This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
.) =
1(0)); if el,···, en is a sequence of edges connecting 0 = (0,0) with
U = (Ul,U2) E Z2, then
II
w(U,>.) =
L(ei'>')w(O,>.).
l:S:i:S:n
One now defines the so-called Calapso transform 9 = T)J : Z2
w-l(u,>.)
G) ~
---+
~N by
(TA;(U)).
Here equivalence of 2 x 1 matrices with entries from ee(~N) is understood modulo simultaneous right multiplication of its entries with one and the same invertible element of the algebra, so that
>.
= 0 one gets
w(-u,O) =
(~) ~
G
(ab1-
1
-f1(U)), so that w-1(u,0)
).
(~)
Clearly, for =
(f~U)).
Therefore, for small >., TAf = f + 0(>'). Prove that the Calapso transform is a discrete isothermic surface with the cross-ratios
001/(1 - >'(01) q(g,gl,g12,g2) = 002/(1- >'(02) . 4.14. Prove that the Calapso transformation 9 = TAf acts on all stars of the net f by Mobius transformations. That is, for every u E Z2 there exists a ~1obius transformation 1I1A(u) : ~N ---+ ~N such that
g(u) = MA(U) . f(u)
and
gi(U) = MA(U) . fi(U)
(i = ±1, ±2).
4.15. Given three pairwise touching spheres, prove that the circle through the touching points is orthogonal to all three spheres. 4.16. Given four pairwise (cyclically) touching spheres, prove that the four touching points are concircular, and that the circle intersects all spheres under equal angles. 4.17. In the previous exercise: if the circle intersects the spheres orthogonally, then the spheres are linearly dependent (hence S-isothermic).
4.7. BibliogTaphicall1otcs
183
4.18. Prove the so-called "Touching coins lemma": whenever four circles in 3-space tOllch cyclically but do not lie on a common sphere, they intersect the sphere which passes through the points of contact orthogonally. 4.19. Complete the details in the proof of Theorem 4.43. 4.20.* Prove Theorem 4.50 and compute the coefficients n, ;1. 4.21. Derive the following discrete Weierstrass representation for circular minimal surfaces:
(1 02~ (1 n1~
611
gIg, -[(1 + 9U}) , g1 + g) , 091 - g g1 - Y g1 - 9
g2 g),
Y29, i(l + !J2Y) , + g2 - 9 g2 - g g2 - 9
621
where Y : ;;32 --> C is a discrete hololl1orphic map, i.e., a solution of the cross-ratio equation °1
q ( g, Y I • Y 12. g2 ) = - ; 0'2
see Chapter 8. Hint: The isothennic Gauss map n : ;;32 stereographic projection of g, (nl
.
+ In2,n:3)
=
(29 I')
1)
lyI 2 'I g-+l 9 12 +1
1
--> §2
is given by the
.
4.22. Derive an explicit formula for the discrete Enneper minimal surface via the discrete Weierstrass representation of Section 4.5.5 by applying it to the standard square grid. The latter is the simplest isothermic net in a plane.
4.7. Bibliographical notes Section 4.1: Discrete Moutard nets in quadrics. General Moutard nets in quadrics were introduced in Bobenko-Suris (2005) (the first online version of this book), along with the most prominent example of the l\loutard representatives of discrete isothermic nets. The latter example was generalized in Bobenko-Suris (2007b), where discrete isothenllic nets in various sphere geometries were investigated. Later the Harne class of nets (Koenigs nets in quadrics) was treated in Doliwa (20071». Section 4.2: Discrete K-nets. The notion of discrete K-nets is due to Sauer (1950) in the case rn = 2 and to Wunderlich (1951) in the case rn = 3. A study of discrete K-surfaces within the framework of the theory of integrable systems was performed ill Bobenko-Pinkall (1996a) geometrically, and in Bobenko-Pinkall (1999) analytically. A presentation in BobenkoMatthes-Suris (2005) is based on the notion of consistency. The study of the GausH map of K-Hurfaces leadH to the notioll of discrete Lorentz-harmonic
184
4. Special Classes of Discrete Surfaces
nets in §2, also introduced in Bobenko-Pinkall (1996a), where the m = 2 case of Theorem 4.8 was first observed. Special classes of discrete K-surfaces were constructed by Hoffmann (1999) (discrete Amsler surfaces; see also Exercise 4.8) and Pinkall (2008) (discrete K-cylinders that touch a plane along a closed curve and those exhibiting a cone point). Color images of discrete K-surfaces are included in the book Bobenko-Seiler (1999). Discrete surfaces in Figures 4.3, 4.5 were produced using a software implementation by Ulrich Pinkall. K-surfaces are reciprocal parallel to geodesic conjugate nets, called Voss surfaces. Discrete Voss surfaces were introduced in Sauer-Graf (1931) as Q-surfaces with the property that the opposite angles at each vertex are equal. Sauer (1950) has shown that the relation between Voss surfaces and K-surfaces is preserved in the discrete setup; see also a modern presentation in Schief-Bobenko-Hoffmann (2008). The study of the angle between the asymptotic lines on discrete Ksurfaces leads to the discretization of the sine-Gordon equation, performed in Bobenko-Pinkall (1996a). The closely related integrable discretization of the sine-Gordon equation was derived by Hirota (1977b) without geometric interpretation. Its symplectic structure was studied in Faddeev-Volkov (1994). Stationary solutions of the discrete sine-Gordon equation describe a discrete pendulum, which was studied in Suris (1989), Bobenko-Kutz-Pinkall (1993). Besides discrete K-surfaces, there exist further remarkable special classes of discrete asymptotic nets: discrete affine spheres studied in Bobenko-Schief (1999a, b), and discrete Bianchi surfaces studied in Doliwa-NieszporskiSantini (2001). Section 4.3: Discrete isothermic nets. Discrete isothermic surfaces were introduced in Bobenko-Pinkall (1996b).
Darboux transformations for discrete isothermic surfaces were introduced in Hertrich-Jeromin-Hoffmann-Pinkall (1999). In particular, Theorem 4.26 on the 3D consistency of the cross-ratio equation was given in this paper for the quaternionic cases N = 3,4 under the name "hexahedron lemma" with a computer algebra proof. An analytic description of the Darboux transformation as a dressing transformation was given in Cieslinski (1999). Three-dimensional discrete isothermic nets were introduced in Bobenko (1999) and Bobenko-Pinkall (1999). A conceptual proof of the 3D consistency in a more general context of an arbitrary associative algebra was given in Bobenko-Suris (2002b).
4.7. Bibliographical notes
185
The Calapso transformation for discrete isothermic surfaces (see Exercises 4.13, 4.14) as well as permutability properties of various transformations are due to Hertrich-Jeromin (2000, 2003). Discrete isothermic surfaces in higher co dimensions were studied by Schief (2001). Besides the discrete isothermic nets, there exists another interesting special class of multidimensional circular nets. These are discrete analogs of Egorov metrics. They are characterized by the property that any elementary quadrilateral (j, Ii, fij, fj) has two right angles at the vertices fi and fj· (Note that for this definition it is essential to fix the directions of all coordinate axes.) The theory of discrete Egorov nets is due to Schief, AkhmetshinVol'vovskij-Krichever (1999) and Doliwa-Santini (2000). Section 4.4: S-isothermic surfaces. The presentation of this section essentially follows Bobenko-Suris (2007b). S-isothermic surfaces, along with their dual surfaces were originally introduced in Bobenko-Pinkall (1999) for the special case of touching spheres. The general class of Definition 4.34, together with Darboux transformations and dual surfaces, is due to Hoffmann. Section 4.5: Discrete surfaces with constant curvature. Circular minimal surfaces were introduced in Bobenko-Pinkall (1996b) as Christoffel duals of their isothermic Gauss maps. The discrete Weierstrass representation was also derived in this paper. Circular surfaces with constant mean curvature appeared for the first time in Bobenko-Pinkall (1999) and Hertrich-Jeromin-Hoffmann-Pinkall (1999) as isothermic nets with a Christoffel dual at constant distance. In the second paper it was shown that equivalently circular surfaces with constant mean curvature can be defined as isothermic surfaces with a Darboux transform at constant distance.
Curvatures of circular surfaces with respect to arbitrary Gauss maps n E §2 based on Steiner's formula were introduced in Schief (2003a, 2006), where it was also shown that the surfaces parallel to a surface with constant Gaussian curvature are linear Weingarten. A curvature theory for general polyhedral surfaces based on the notions of parallel surfaces and mixed area is developed in Pottmann-Liu-Wallner-Bobenko-Wang (2007) and BobenkoPottmann-Wallner (2008). In the circular case this theory yields the same class of surfaces with constant curvatures as originally defined in BobenkoPinkall (1999), Hertrich-Jeromin-Hoffmann-Pinkall (1999); see Corollaries 4.52, 4.53. Discrete surfaces in Figure 4.19 were produced using a software implementation by Peter Schroder. The theory of minimal surfaces of Koebe type was developed in BobenkoHoffmann-Springborn (2006). These surfaces are S-isothermic and their Gauss maps are Koebe polyhedra. Global results in this theory are based
186
4. Special Classes of Discrete Surfaces
on the remarkable fact that a Koebe polyhedron is essentially uniquely determined by its combinatorics. This theory is closely related to the theory of orthogonal circle patterns (see Chapter 8). Section 4.6: Exercises. Ex. 4.6, 4.7: See Bobenko-Pinkall (1996a).
Ex. 4.8: See Bobenko-Pinkall (1996a), Hoffmann (1999). Ex. 4.11: See Wallner-Pottmann (2008); the corresponding theorem for smooth surfaces can be found in Darboux (1914-27, §874). Ex. 4.13, 4.14: See Hertrich-Jeromin (2000, 2003). Ex. 4.18: See Bobenko-Hoffmann-Springborn (2006). Ex. 4.21, 4.22: See Bobenko-Pinkall (1996b).
Chapter 5
Approximation
We have already had several occasions to mention that the notions, constructions and results of discrete differential geometry have not just qualitative similarity with their much more sophisticated counterparts in the smooth theory. Rather, the latter can be obtained from the former through a wellestablished continuous limit. Strictly speaking, such a continuous limit has been established up to now only for those geometries which are described by hyperbolic systems of difference, resp. differential, equations. It is this class of equations for which a rather detailed approximation theory can be developed, which is similar to the corresponding theory for ordinary difference and differential equations. Actually, this hyperbolic theory covers a substantial part of the nets considered in this book.
5.1. Discrete hyperbolic systems To formulate the general scheme that covers the majority of situations encountered so far, we will put our hyperbolic systems into the first order form. It should be stressed that this is necessary only for general theoretical considerations, and will never be done for concrete examples.
Definition 5.1. (Hyperbolic system) A hyperbolic system of first order partial difference equations is a system of the form (5.1 ) for functions Xk : ZM - t X k with values in Banach spaces X k . For each Xk, equations (5.1) are posed for i E ek C {I, ... , M}, the nonempty set of evolution directions of Xk. The complement Sk = {I, ... ,M} \ ek consists of static directions of Xk·
-
187
188
5. Approximation
We think of the variable Xk(U) as attached to the elementary cell ek(U) of dimension #Sk adjacent to the point U E ZM and parallel to 'B Sk : e k = ek(U) = {u
+L
J-liei : J-li E
[0,1]}.
iESk
Here, we recall,
'B S = {u E Z M for an index set S c {I, ... , M}.
: Ui
= 0 if i 1:- S},
Definition 5.2. (Goursat problem) 1) A local Goursat problem for the hyperbolic system (5.1) consists in finding a solution Xk for all k and for all cells e k within the elementary cube of ZM at the origin from the prescribed values Xk(O). The system (5.1) is called consistent if the local Goursat problem for this system is uniquely solvable for arbitrary initial data Xk(O). 2) A global Goursat problem consists in finding a solution of (5.1) on ZM subject to the following initial data:
(5.2) where X k : 'BSk
---+
Xk are given functions.
The following rather obvious but extremely important statement holds: Theorem 5.3. (Well-posed Goursat problem) A Goursat problem for a consistent hyperbolic system (5.1) has a unique solution x on all ofZM. Consistency conditions read: bjbiXk = bibjXk for all i (5.1), one gets the following equations:
(5.3)
bj9k,i(X)
=
bigk,j(X),
i
1= j.
Substituting
1= j,
or 9k,i(X+ gj(x)) - gk,i(X) = gk,j (x+ gi(X)) - gk,j(X), where gi(X) is a vector function whose £-th component is equal to gC,i(X) if i E cp, and is undefined otherwise. Lemma 5.4. For a consistent system of hyperbolic equations (5.1), the function gk,i only depends on those components Xp for which Sp C Sk U {i}. Proof. Equations (5.3) must hold identically in x. This implies that the function gk,i can only depend on those components Xp for which bjxp is defined, i.e., for which j E cp. As (5.3) has to be satisfied for all j E Ck, j 1= i, one obtains that Ck \ {i} c cp for these £. 0 It follows from Lemma 5.4 that for any subset S C {I, ... , M}, equations of (5.1) for k with Sk c S and for i E S form a closed subsystem, in the sense that gk,i depend on Xp with Sp c S only.
5.1. Discrete hyperbolic systems
189
Definition 5.5. (Essential dimension) The number (5.4)
is called the essential dimension of system (5.1). If d = M, system (5.1) has no lower-dimensional hyperbolic subsystems. If d < M, then d-dimensional subsystems corresponding to S with #S = d are hyperbolic. In this case, consistency of system (5.1) is a manifestation of a very special property of its d-dimensional subsystems, which we treat as the discrete integrability (at least if one excludes certain noninteresting situations, such as trivial evolution in some of the directions). Section 6 will be devoted to an extensive treatment of integrability understood as consistency.
Example 1. Consider a difference equation with M = 3 independent variables:
(5.5)
616263x = F(x, 61 x, 62x, 63x, 6162x, 6163x, 6263x).
One can pose a Goursat problem by prescribing the values of x on the coordinate planes 1312,1313,1323. Equation (5.5) can be rewritten as a hyperbolic system of first order equations by introducing auxiliary dependent variables a, b, c, f, g, h: 61 x = a, 62 a = f, (5.6)
62 x = b, 63 a = g, 63 b = h,
61 b = f, 61 c=g, 62 c=h, 63f = 62g = 61 h = F(x, a, b, c, f, g, h).
It is natural to assume that the variable x lives on the points of the cubic lattice U E Z3; the variables a, b, c live on the edges C1(u), C2 (u), C3 (u) of the lattice adjacent to the points u and parallel to the coordinate axes 131,132,133, respectively; and the variables f, g, h are associated to two-cells (elementary squares) CI2 (U), CI3(U), C23(U) adjacent to the points u and parallel to the coordinate planes 1312,1313,1323, respectively. Thus, x has no stationary directions; the stationary directions of a, b, care {1}, {2}, {3}, while the stationary directions of f, g, hare {1, 2}, {1, 3}, {2, 3}, respectively. A Goursat problem for this system would be posed by prescribing the values of x at the point (0,0,0), the values of a, b, c on the axes 13 1,13 2,13 3, respectively, and the values of f, g, h on the planes 1312,1313,1323, respectively. The essential dimension of this discrete system is d = 3 = M. It is instructive to compare this construction with its continuous counterpart: equation (5.5) is a natural discretization of the partial differential equation
(5.7)
[hfhfhx = F(x, 81 x, 82x, 83x, 81 82x, 81 83x, 8283X).
190
5. Approximation
For the latter equation one can introduce auxiliary variables a, b, c, j, g, h via partial derivatives analogously to (5.6). All these variables would be on an equal footing, being defined just at the points U E ]R3.
Example 2. Consider the difference equations which govern M -dimensional Q-nets:
(5.8) Upon introducing auxiliary variables system of first-order equations: 6iX
(5.9)
{
Vi
they can be written as a hyperbolic
= Vi,
6jVi = CijVj
+ CjiVi,
6i Cjk = (TkCij )Cjk
i =I- j, + (TkCji)Cik
-
(TiCjk)Cik,
i =I- j =I- k =I- i.
The last equation is (2.7) from Section 2.1.1, where one can also find details about its origin, as well as about how one can put it in the form with the right-hand side depending only on the unshifted variables Cij. It is natural to assume that the variable x lives at the points of the cubic lattice U E 7l.M; the variables Vi live at the edges ei (u) of the lattice adjacent to the points u and parallel to the coordinate axes 'B i , and the variables Cij, Cji for i < j live oat two-cells (elementary squares) eij(u) adjacent to the points u and parallel to the coordinate planes 'B ij . Thus, x has no stationary directions, the stationary directions of Vi are {i}, and the stationary directions of Cij are {i,j}. A Goursat problem can be posed by prescribing the values of x at the point 0 E 7l. M , the values of Vi on the axes 'B i , and the values of Cij, Cji on the coordinate planes 'B ij . The essential dimension of this discrete system is d = 3, independently of M. In particular, it may well be d < A1. Consistency of this system for any M 2 3 is interpreted as its integrability.
5.2. Approximation in discrete hyperbolic systems To handle approximation results for discrete geometric models, we need to introduce small parameters into hyperbolic systems of partial difference equations. The domain of our functions becomes
'B€ =
El7l.
x ... x
EM7l..
If Ei = 0 for some index i, the respective component in 'B€ is replaced by lR.. For instance, if E = (0, ... ,0), then 'B€ = ]RM. Thus, the domains 'B€ possess continuous and discrete directions, with mesh sizes depending on the parameters Ei. The definitions of translations and difference quotients are modified for functions on 'B€ in an obvious way:
5.2. Approximation in discrete hyperbolic systems
If Ei = 0, then 5i is naturally replaced by the partial derivative multi-index cx = (CXI, ... ,CXM), we set 50: = 5r 1 •••
5't,r.
The definition of elementary cells ified as follows:
ek =
{n
+L
191
fk For a
ek , carrying the variables Xk, is mod-
JLiei: JLi E [0, EiJ}
iESk
(so that the cell size shrinks to zero in the directions with Ei = 0). We see how the discreteness helps to organize the ideas: in the continuous case, when all Ei = 0, all the functions Xk live at points, independently of the dimensions #Sk of their static spaces. In the discrete case, when all Ei > 0, one can clearly distinguish between functions living on vertices (those without static directions), on edges (those with exactly one static direction), on elementary squares (those with exactly two static directions), etc. Having in mind the limit E - t 0, we will only treat the case when the first m ::; M parameters go to zero in a uniform way, EI = ... = Em = E, while the other Joy! - m ones remain constant, Em+l = ... = EM = 1. In this case 13 E = (Ez)m X zM-m, and we set 13 = 13° =]Rm x zM-m. Assuming that the functions gk,i = gk,i on the right-hand sides of (5.1) depend on E smoothly and have limits as E - t 0, we will study the convergence of solutions XE of the difference hyperbolic system (5.1) to the solutions xO of the limiting differential (or differential-difference) hyperbolic system
n {1, ... , m}, Gk n {m + 1, ... , M}.
(5.10)
OiXk
gZ,i(X),
i E Gk
(5.11)
5iXk
gZ,i(X),
i E
Naturally, (5.10), (5.11) describe the respective m-dimensional smooth geometry with ]v! - m permutable transformations. Throughout this chapter, a smooth function 9 : 'D - t X is a function that is infinitely differentiable on its domain, 9 E CXl('D). For a compact set X c 'D, we say that a sequence of smooth functions gE converges to a smooth function gO with order 0;
ii) discrete system (5.1) approximates the differential(-difference) system (5.10), (5.11) with order 0 small enough, solution x k of the Goursat problem exists and is unique on 'BE(1'); moreover, solutions x k converge to smooth functions xZ with order Z is open and dense in Xk . In such a case, one requires in ii) the convergence gk,i ---+ 9£,i in Coo(1)O). Then conclusions of Theorem 5.7 hold for generic initial data.
5.2. Approximation in discrete hyperbolic systems
193
As for condition iii), smooth data X2 : 13sk ~ X k are usually given a priori, and discrete data X k are obtained by restriction to the lattice: X k = X2113 In such a situation, condition iii) is fulfilled automatically. f
"k
•
We will not present a complete proof of Theorem 5.7 here but rather a substantial part of it illustrating all necessary technical ideas. Namely, we will provide the arguments for the simplest situation M = m = 2, and in this particular case we will only demonstrate the convergence k ~ 0 for solutions themselves and omit the proof of 8Q (xk - xV ~ 0 for multi-indices a with lal :::: 1. In other words, we will prove the CO(13)-convergence instead of COO (13)-convergence. The proof of the COO(13)-convergence for general M requires technical care but no essentially new ideas. Thus, we consider the discrete hyperbolic system
x x2
(5.12) with smooth functions
r, gE : X x X
~
X, and with the Goursat data
(5.13) for
Ui E
(5.14)
13i( r). It is supposed that the functions r(a, b) = fO(a, b)
+ (9(E),
r, gE satisfy
gE(a, b) = gO(a, b)
+ (9(E),
uniformly on compact subsets of X x X, and that relation analogous to (5.14) holds for all partial derivatives of the functions gE. Further, it is assumed that the discrete Goursat data (5.13) also have smooth limits:
r,
(5.15) uniformly for Ui E 13i(r). Then the solutions (a E, bE) of the Goursat problems for system (5.12) converge uniformly in 13(1'), with a suitable l' E (0, r], to a pair of Lipschitz functions (a O, bO), (5.16)
a E (Ul' U2) = aO( Ul, U2)
+ (9( E),
bE ( Ul, U2) = bO( Ul, U2)
+ (9( E),
which constitute the unique solution in 13(1') of the Goursat problem for the system (5.17) with the Goursat data
(5.18) for Ui E 13i(r).
Lemma 5.S. (A priori estimate) Let the norms of Goursat data AE, BE be bounded by E-'independent constants. Then there exists some r E (0, r] such that the norms of the solutions (a E , bE) of the Goursat problem (5.12), (5.13) are bounded on 13 E (r) independently of E.
5. Approximation
194
Proof. Let
IA'I, IB'I 'S: Mo, r
and choose Ah > Mo arbitrarily. Define
(M] - AIo)/ sup E
sup {Ir(a, lal,lbl Ml and M > M 2. We now prove the estimate for 81aE. Proceeding from U2 to U2 + E, we find:
181aE(U1, U2
+ E)I
0 such that the canonical Q-nets r : 'Bf(r) ---t ]RN converge, as E ---t 0, to the unique conjugate net f : 'B(r) ---t ]RN with the initial data (Q1,2). Convergence is with order (')(E) in COO('B(r)). Proof. This follows directly from Theorem 5.7, since systems (2.1), (2.7) and (1.1), (1.2) are manifestly hyperbolic (and can be easily rewritten in the first order form). 0 Discretization of an F-transformation. Recall that an F-transform of a given conjugate net is determined by the initial data (F 1,2) (see Section 1.1). We now produce from these the initial data (Ft2) (see Section 2.1) for an E-dependent family of F-transforms of canonical Q-nets corresponding to the initial data (Q1,2). Take the point f+ (0) from (F d· Define the edge functions ail 'Bi' bi! 'Bi by restricting the functions ai I'B i , bi I'Bi to the lattice points, or, better, to the midpoints of the corresponding edges of 'Bi. This gives the data set (F~2); along with the data (Q~2) produced above this yields in a canonical wa; an E-dependent family of Q-nets FE : (Ez)m X {O, I} ---t ]RN, which will be called the canonical Q-nets for the initial data (Q1,2), (F 1,2)'
Theorem 5.12. (Convergence of discrete F-transformations) The canonical Q-nets (r)+ = F E(-, 1) : 'BE(r) ---t ]RN converge to the net f+ : 'B (r) ---t ]RN which is the unique F-transform of f with the initial data (F 1,2), Convergence is with order (')(E) in COO('B(r)). Proof. Again, this follows directly from Theorem 5.7 applied to the hyperbolic systems consisting of (2.10)~(2.12) in the discrete case and of (1.12)~ (1.14) in the smooth case. Note that the discrete equations are implicit, and their solvability for E small enough is guaranteed on the subset of the phase space, {aj -1= 0 : 1 :s; j :s; m}, which is open and dense. 0
5.4. Convergence of discrete Moutard nets Discretization of a Moutard net. Given a Moutard net y : ]R2 ---t ]RN defined by the initial data (M1,2) (see Section 1.2.1), we produce initial data
5. Approximation
198
(Mt2) for an E-dependent family of discrete M-nets yf : (EZ)2 ~ ]RN (see Section 2.3.10). Discrete curves yf r:13 ]R3 (see (fZ)2 ----> Section 2.4.2). Equations (2.69) define the discrete A-surfaces ]R3, called the canonical discrete A -surfaces corresponding to the initial data
r:
(A 1,2). Theorem 5.15. (Convergence of discrete asymptotic nets) Canonical discrete A -surfaces r : 13 E Cr) ----> ]R:3 converge, as f ----> 0, to the unique Asurface f : 13(r) ----> ]R3 w'ith the initial data (Au). Convergence is with order O(E) in C OO (13(r)). Proof. Equations (2.69) are hyperbolic and they approximate equations (1.38). Theorem 5.7 can be applied to prove the convergence of after the E convergence of n has been already proven. 0
r,
Discretization of a Weingarten pair. Initial data (W 1,2) for a Weingarten transformation (see Section 1.3) are nothing but initial data (MT 1,2) for a l\1outard transformation of the Lelieuvre normal field. The construction of Theorem 5.14 delivers the initial data for a family of Lelieuvre normal fields (n E )+ : (fZ)2 ----> ]R3, which are therefore seen as the data (W~2) for transformed A-surfaces (r)+ : (fZ)2 ----> ]R:3, obtained via (2.70) (see Section 2.4.3). Theorem 5.16. (Convergence of discrete Weingarten transformations) Canonical discrete A-nets (r)+: 13 E (r) ----> ]R3 converge to the unique Weingarten transform, f+ : 13(1') ----> ]R3 of f with the initial data (W 1 ,2). Convergence is with order C) ( f) in Coo (13 (1') ). Proof. This is proven by comparing the (identical) formulas (1.40) and (2.70), after the convergence of 11,( and (nE)+ has been established. 0
200
5. Approximation
5.6. Convergence of circular nets In this section, we address the problem of approximating smooth orthogonal nets by discrete circular nets. Recall that the former are governed by the system (1.44)-( 1.47) with constraint (1.48), while the latter are governed by similarly looking equations (3.12)-(3.14), (3.17) with constraint (3.18). We demonstrate that this analogy can be given a qualitative content, so that for a given orthogonal net one can construct an approximating family of circular nets. However, there is a substantial obstruction to accomplishing this, which can be seen by a careful comparison of the constraints (1.48) and (3.18). We think of smooth rotation coefficients as being approximated by discrete ones. But since the discrete rotation coefficients !3kj only have i =1= k, j as evolution directions (that is, they are plaquette variables attached to elementary squares parallel to 'B jk ), there is seemingly no chance to get an approximation of such smooth quantities as Oi!3ij involved in the smooth orthogonality constraint (1.48). In order to be able to achieve such an approximation, we need some discrete analogs of the smooth rotation coefficients which would live on edges. For this aim, we turn to the Mobius-geometric description of circular nets from Section 3.1.4, more precisely, to the frame equations (3.25). Introduce vectors Vi = 1jJV(lp-l; then the frame equations become h1jJ)1jJ-l = -eiVi. Expanding these vectors with respect to the basis vectors ek, we have a formula analogous to (1.53):
(5.26)
Vi
v
= 1jJ i1jJ-l = O"iei -
1
"2
L Pkiek + hieoo . kii
The fact that the eoo-component here is equal to hi, is easily demonstrated. Indeed, from (3.25) it follows that TJ - j = hiVi = h i (Ti1jJ)-lei1jJ. Now equation (3.24) allows us to rewrite this equivalently as [eo, (Ti'tP )1jJ-l] = hiei, which proves the claim above. Observe also the normalization condition
(5.27)
0"; =
1
-l L P~i
.
kii
Coefficients Pki are edge variables analogous to smooth rotation coefficients. Indeed, the vectors Vi are defined on the edges of zm parallel to the coordinate axis 'B i . However, these vectors do not immediately reflect the local geometry near these edges; rather, they are obtained by integration of the frame equations (3.25), and thus are of a nonlocal nature. Thus, in the discrete case we have two different analogs of the rotation coefficients: local plaquette variables !3ij for 1 :S i =1= j :S m, defined on the elementary squares of zm parallel to 'B ij , and nonlocal edge variables Pki for 1 :S i :S m, 1 :S k :S N, k =1= i, defined on the edges of zm parallel to 'B i .
5.6. Convergence of circular nets
Evolution equations for tions (3.25):
Vi
Ti Vj
201
are obtained from (3.13) and the frame equa-
= vJi 1ei(Vj + .Bij Vi)ei.
In the derivation one uses the identity Vi(Vj + .BjiVi)Vi = Vj + .BijVi, which easily follows from (3.18). The resulting evolution equations for the edge variables Pkj read: (5.28)
TiPkj
VJi 1 (Pkj+Pki.Bij),
(5.29)
TiPij
VJi 1(-Pij
Here 1 ::; i # j ::; m, 1 ::; k ::; N, and k (3.18) can be now written as
(5.30)
.Bij
+ .Bji =
+ 2(Ji.Bij). # i, j. The
circularity constraint
1
(JiPij
+ (Jjpji - 2 L
PkiPkj,
k-j.i,j
and gives a relation between local plaquette variables .Bij and nonlocal edge variables Pkj. The system consisting of (5.28), (5.29) and (5.30) can be regarded as the discrete Lame system; d. (1.47), (1.48). Turning to the problem of approximation, we start with approximation of a single orthogonal net f : ]Rm -----t ]RN. For the approximating circular nets, we have M = m and all Ei = E. In all the formulas of Section 3.1 and of the present section, we have to replace the lattice functions hi, .Bij, Pkj by Ehi' E.Bij, EPkj, respectively. Observe that formulas (3.16), (5.27) become Vji
= Vij =
(Ji
(1 - E2.Bij.Bji)1/2
= 1 + (')(E2),
E2 ' " 2) 1/2 = ( 1 - "4 ~ Pki = 1 + (')(E2). k-j.i
Under this rescaling, equations (3.12), (3.13), (3.14) and (3.17) can be put into the standard form (5.1) with functions on the right-hand sides approximating, as E -----t 0, the corresponding functions in equations (1. 7)-(1.10) with order (')(E). Nevertheless, Theorem 5.7 still cannot be applied to orthogonal nets. The reason for this is that the full system of differential equations describing orthogonal nets, consisting of equations (1.7)-(1.10) and constraint (1.54), is nonhyperbolic. Its nonhyperbolicity rests on the fact that the constraint (1.54) is not resolved with respect to the derivatives oi.Bij. Note, however, that constraint (1.54) does not take part in the evolution of solutions starting with the data given in the coordinate planes 'B ij : the constraint is satisfied automatically, provided it is fulfilled for the coordinate surfaces ~ij" Therefore, we will obtain a convergence result for orthogonal nets as soon as it will be established for coordinate surfaces.
n
5. Approximation
202
Discretization of a curvature line parametrized surface. Initial data for a smooth curvature line parametrized surface f : 13 12 ----> JR.N are:
(i) two smooth curves f(O);
f
r~i
(i = 1,2), intersecting orthogonally at
(ii) a smooth function ,12 : 13 12 ,12 = ~(aI/h2 - 2 !hd·
---->
JR., whose designated meaning is
a
Let] r~i be the images of the curves f r~i in the Mobius-geometric model Q~ . Let hi = laJI and Vi = h:;laJ be the metric coefficients and unit tangent vectors of the coordinate curves. Choose an initial frame 1}J(0) E :J-C(X) such that ](0) = 1}J-1(0)eo1}J(0), Vi(O) = 'If,-l(O)ei1}J(O) (i = 1,2). Define the frames 1}J : 'Bi ----> :J-C(X) of the curves ] r~i as the solutions of equations (1.51) for i = 1,2 (considered as ordinary differential equations) with the initial value 1}J(0). Rotation coefficients of the curves Jr~i are the functions (3ki : 'Bi ----> JR. defined by the formula (1.53) for i = 1,2. Define the discrete coordinate curves
]£ r~1
by restricting the functions
Jr ~i to the lattice points.
Let hi = I JR. in the expansions £ £ + E1£lie(X). (Jiei - "2E '\""' L Pkiek k#i Finally, let the plaquette function '12 : 1312 ----> JR. be obtained by restricting ,12 to the lattice points (or to the midpoints of the corresponding plaquettes of 1312).
A£(nl.t)-l Vi£ = nl.£ 'f/ Vi 'f/
=
Thus, we get valid Goursat data for a hyperbolic system of first order difference equations for the variables vi, hi, Phi' consisting of (3.12), (3.13), (3.14), (5.28) and (5.29) with distinct i. j E {1,2} and 1 :::; k :::; N, where the following expressions should be inserted:
if,
t E E E (312 = (J1P12-"2
(1"2 'LPk1Pk2-'12 \""' E
£
£
)
,
k>2 The nets ]E : 1312 ----> Q~ defined as solutions of the Goursat problem just described are circular surfaces, since they fulfill the circularity constraint (5.30). They will be called canonical circular surfaces constructed from the above initial data. Theorem 5.17. (Convergence of circular surfaces) There exists r > 0 such that the canonical circular surfaces ]E : 1312 (r) ----> Q~ converge, with
5.6. Convergence of circular nets
203
order ()(E) in C OO (13 12(r)), to the unique curvature line parametrized surface j : 13 12 (r) -> Qij with the initial data n:Bi (i = 1, 2) and ~ (oIlh2 - 02/321) = ')'12· Edge rotation coefficients pL and plaquette rotation coefficients /312, /3~1 of the circular surfaces jt converge to the corresponding rotation coefficients /3ki of the curvature line parametrized surface j. Proof. We begin with showing the convergence of the frames, 'lj;: -> 'lj;, and of the rotation coefficients, Pki -> /3ki , along the discrete curves r:Bi' This follows from two observations. First, Vf(O) = Vi(O) + ~(OiVi)(O) + ()(E2), so that (Ti - 1)'1//(0) = -~ ei'lj;(O) (OiVi) (0) + ()(E2).
r
Second, combining frame equations on two neighboring edges of finds that
(Ti - Ti-1)'lj;t = -ei'lj;t(l- Ti- 1 )v'f = -Eei'lj;t(OiVi) everywhere on theory.
13i.
13i,
one
+ ()(E2)
The claim follows by standard methods of the ODE
Now an application of Theorem 5.7 shows that the functions jt : 1312 -> Qij converge to the functions j : 13 12 -> Qij which solve the Goursat problem for the hyperbolic system of first order differential equations, consisting of (1.44)-(1.47) with distinct i,j E {I, 2} and 1 ::; k ::; N, and
Solutions /3ki satisfy the relation ~ (01/312 - 02/321) = ')'12 and the orthogonality constraint (1.48). D Discretization of an m-dimensional orthogonal net. Given the initial data (0 1,2) for an m-dimensional orthogonal net (see Section 1.4), we can apply the procedure described in the previous paragraph, with an initial frame 'lj; (0) E :J{oo such that
j(O) = 'lj;-l(O)eo'lj;(O),
Vi(O) = 'lj;-l(O)ei'lj;(O)
(1::; i ::; m),
to produce, in a canonical way, the circular surfaces jt
r:B'. t)
and their pla-
quette rotation coefficients /3ij . Thus, we get the data (Ot2) (see Section 3.1) for an E-dependent family of circular nets jt : (Ez)m -> Qij. These nets will be called the canonical circular nets corresponding to the initial data (0 1,2), Theorem 5.18. (Convergence of circular nets) The canonical circular nets jt : 13€(r) -> ]RN converge, as E -> 0, to the unique orthogonal net
204
5. Approximation
j : 'B(r) ~ jRN with the initial data (0 1,2), Convergence is with order O(E) in COO('B(r)). Proof. The data (0~2) yield a well-posed Goursat problem for the hyperbolic system of first order difference equations for the variables jE, vi, hi, f3ij , consisting of (3.12), (3.13), (3.14), (3.17). The convergence of these Goursat data is assured by Theorem 5.17. Now the claim of the theorem follows directly from Theorem 5.7. 0 Discretization of a Ribaucour transformation. Given the initial data (R1,2) for a Ribaucour transform of an orthogonal net (see Section 1.4), define the plaquette rotation coefficients f3 Mi on the "vertical" plaquettes along the edges of the coordinate axes 'Bi by restricting the corresponding functions d)i to lattice points or, alternatively, to midpoints of the corresponding edges of 'Bi: U
E
'Bi,
1
~
i
~
m.
Thus, we get the data (R~2) (see Section 3.1), which, together with (0~2)' allow us to construct in a ~anonical way circular nets FE : (Ez)m X {O, 1}' ~ jRN. They will be called the canonical circular nets corresponding to the initial data (0 1,2), (R1,2).
Theorem 5.19. (Convergence of discrete Ribaucour transformations) The canonical circular nets (r)+ = P(-, 1) : 'BE(r) ~ jRN converge to the unique Ribaucour transform f+ : 'B(r) ~ jRN of f with the initial data (R1,2). Convergence is with order O(E) in COO('B(r)). Proof. Define vlf(O) as the unit vector parallel to 8f(0) = f+(O) - f(O), and set hM(O) = 18f(0)1. These data along with f3 Mi on the coordinate axes, added to the previously found r(O), vi, hi, f3ij for 1 ~ i,j ~ m, form valid Goursat data for the system (3.12), (3.13), (3.14), (3.17). The circularity constraint (3.18) implies that f3iM = -2(vi, vlf ) - EOi on all edges of 'Bi. Perform the substitution
vM = y+O(E),
hM = £+O(E),
f3Mi = EOi+0(E2), (jiM =
-2(Vi, y)+O(E)
in equations (3.13), (3.14), (3.17) with one of the indices equal to M. Taking into account that in this limit one has
ViA} = vM~ = 1 - E(Vi' y)Oi + 0(E 2 ), one sees that the limiting equations coincide with (1.57), (1.58), (1.59). A 0 reference to Theorem 5.7 finishes the proof.
5.7. Convergence of discrete K-surfaces
205
5.7. Convergence of discrete K-surfaces Discretization of a K-surface. Given the initial data (K) for a K-surface (see Section 1.6), we define the initial data (K.6.) (see Section 4.2) for an Edependent family of discrete K-surfaces with EI = E2 = E by restricting nf 'B; to the lattice points, as for general A-surfaces. Thus, we get two intersecting discrete curves in §2. Define discrete M-nets nf : (EZ)2 -> §2 as solutions of the difference equations (4.31) with the initial data (K.6.). Finally, define the discrete K-surfaces (EZ)2 -> ffi.3 with the help of the discrete Lelievre representation (2.69). These will be called the canonical discrete K-surfaces corresponding to the initial data (K).
r :
Theorem 5.20. (Convergence of discrete K-surfaces) Canonical discrete K-surfaces r: :Bf(r) -> ffi.3 converge, as E -> 0, to the unique K-surface f : :B(r) -> ffi.3 with the initial data (K). Convergence is with order CJ(E) in COO(:B( r)). Proof. We have for n = nf: € (n, Tin + T2n) 2 + E(n, (hn aI2 = 1 + (Tin, T2n) = 2 + E(n, 6In + 62n)
+ 62n) + E2(6In, 62n) .
Since (n,6in) = CJ(E), we find that ab
=
1 - ~E2(6In, 62n)
+ CJ(E4).
Comparing this with (1.68), we see that Theorem 5.7 can be applied and yields convergence of the net n€: (EZ)2 -> §2 to the smooth net n: ffi.2 -> §2. Finally, convergence of to f follows exactly as for general A-surfaces. 0
r
Discretization of a Backlund pair. Let the initial data (B) for a Backlund transformation of a given K-surface f, i.e., the point n+(O), be given. Take it as the initial data (B.6.) for the discrete Backlund transformations (r)+ : (EZ)2 -> ffi.3 of the family of discrete K-surfaces constructed in Theorem 5.20.
r
Theorem 5.21. (Convergence of discrete Backlund transformations) Canonical discrete K-surfaces (r)+ : :B€(r) -> ffi.3 converge to the unique Backlund transform f+ : :B(r) -> ffi.3 of the K-surface f with initial data (B). Convergence is with order CJ(E) in COO(:B(r)). Proof. For the Backlund transformation, equations (4.43), (4.44) hold. In the smooth limit we find: (6In, n+ + n) PI = 1 - (Tin, n+) P2 = -
(62n, n+ - n) 1 + (T2n, n+)
206
5. Approximation
Comparing this with (1.41)-(1.42) and applying Theorem 5.7, we prove convergence of the Gauss maps. 0
5.8. Exercises 5.1. Check that each of the following four difference equations approximates the ordinary differential equation ax = f(x) for x : ~ ~ X (where a stands for the ordinary derivative d/dt):
8x = f(x),
8x = f(TX),
~ TX),
8x = f(X
for x : EZ ~ X, where 8x(t) = (x(t theorem where appropriate.
+ E) -
8x = f(x)
~ f(TX)
X(t))/E. Use the implicit function
5.2. Put the Hirota equation for ¢ : (62:)2 ~ ~, 1 sin 4(TIT2¢ - TI¢ - T2¢
E2
1
+ ¢) = "4 sin 4(TIT2¢ + TI¢ + T2¢ + ¢),
and the sine-Gordon equation for ¢ : ~2 ~ ~,
al a2¢ = sin ¢, into the form of hyperbolic first order systems, and check that the former approximates the latter as E ~ O. 5.3. Put the difference equation for discrete Lorentz-harmonic functions n: (EZ)2 ~ §2, TIT2n+n=
(n, TIn + T2n) ( . ) (TIn+T2 n ), 1 + TIn, T2n
and the differential equation for Lorentz-harmonic functions n : ~2 ~ §2, aIa2n= -(aIn,a2n )n,
into the form of hyperbolic first order systems, and check that the former approximates the latter as E ~ O. Why would the approximation claim fail for similar equations in the case of functions with values in IL,N+I,I? 5.4. Prove the following form of the discrete Gronwall lemma: Let Z+ ~ ~+ be three nonnegative sequences satisfying
~,a,
b:
n-I
~(n) :::; an
+ 2: bkb.(k). k=1
Then
n-I
~(n) :::; an
+ 2: akbk
n-I
II
(1 + bj)' k=O j=k+l An interesting (and important) particular case is that of constant coefficients an = /'l, and bn = K.
5.9. Bibliographical notes
207
5.9. Bibliographical notes Geometric convergence theorems are available in the literature for problems described by elliptic partial differential equations, such as the Plateau problem in the theory of minimal surfaces; see, for example, Dziuk-Hutchinson (1999). Convergence of metric and geometric properties of general polyhedral surfaces was shown in Hildebrandt-Polthier-Wardetzky (2006) based on the analysis of the "cot an" Laplace operator. For surfaces described by hyperbolic partial differential equations, first approximation results were obtained in Bobenko-Matthes-Suris (2003, 2005). The presentation of this section follows these papers. The complete proof of the main approximation Theorem 5.7 can be found in Matthes (2004). A related purely geometric construction of circular nets approximating general curvature line parametrized surfaces is given in Bobenko-Tsarev (2007).
Chapter 6
Consistency as Integrability
Up to now we have encountered many instances of multidimensional nets which serve as discretizations of smooth geometries traditionally associated with, and described by integrable systems. The idea of consistency (or compatibility) is in the core of the integrable systems theory. One is faced with it already at the very definition of the complete integrability of a Hamiltonian flow in the Liouville-Arnold sense, which means exactly that the flow may be included into a complete family of commuting (compatible) Hamiltonian flows. It is impossible to list all applications or reincarnations of this idea. We mention only some of them relevant for our present account.
• In the theory of solitons nonlinear integrable equations are represented as a compatibility condition of a linear system called the zero curvature representation (also known as Lax, or ZakharovShabat representations). Various analytic methods of investigation of soliton equations (such as the inverse scattering method, algebrogeometric integration, asymptotic analysis, etc.) are based on this representation. • It is a characteristic feature of soliton (integrable) partial differential equations that they appear not separately but are always organized in hierarchies of commuting (compatible) flows.
• Another indispensable feature of integrable systems is that they possess Biicklund-Darboux transformations. These special transformations are often used to generate new solutions from the known ones.
-
209
6. Consistency as Integrability
210
In fact all these properties are interrelated and it is customary to understand the integrability as the presence of OIl(' (or some combination) of these features.
In this chapter we show how the development of discrete differential geometry leads to a new l.Ulderstanding of the very notion of integrability and its properties.
6.1. Continuous integrable systems Consider one of the most celebrated integrable systems having numerous applications in differential geometry as well as in mathematical physics, the sine-Gordon equation (6.1) for a function ¢ : ]R2 ---t]R. Recall the geometric interpretation of the sineGordon equation. Let f : ]R2 ---t ]R3 be a surface parametrized along its asymptotic lines. Surfaces of constant negative Gaussian curvature K = -1 (K-surfaces, for short) in the asymptotic lines parametrization are characterized by the additional requirement that 10tIi does not depend on U2, and lodl does not depend on lJ1. Reparametrizing the asymptotic lines of a K-surface if necessary, one can assume that 10tIi = 102fl = 1. Then the angle ¢ = ¢( u) between the vectors otI and 02f satisfies the sine-Gordon equation (6.1). Integrability of the sine-Gordon equation has many manifestations, two of which will be of special importance for us: the zero curvature representation and the existence of Backlund transformations. To formulate the zero curvature representation of the sine-Gordon equation, consider the matrices
~ ( -A
(6.2)
U
=
(6.3)
V
= :2i
ad)
(
0
-A -010
A-le-iq,
),
A-lei )
o
.
They depend on 'u E ]R2 through the function ¢ and its partial derivatives, and also depend on a (real) parameter A, known in the theory of integrable systems as the spectral parameter. It is usual to think about U, V as functions of U E ]R2 which take values ill the twisted loop algebra
g[A]
= {~ : ]R*
--t
su(2): ~(-A) = (}3~(A)(}:3}.
6.1. Continuous integrable systems
211
Then it is a matter of a straightforward computation to check that 1; is a solution of equation (6.1) if and only if the zero curvature condition EhU - 8 1 V
(6.4)
+ [U, V] = 0
is satisfied identically in A. The name "zero curvature" comes from the fact that (6.4) expresses the flatness of the connection (or, better, the oneparameter family of connections) on ]R2 given by the differential one-form U dU1 + V dU2. This condition assures the solvability of the following system of linear differential equations: (6.5) for a function '11 : ]R2 ~ G[A] with values in the twisted loop group
G[A]
= {S:]R*
~ SU(2) : S(-A)
= 0"3S(A)0"3}.
The existence of the zero curvature representation is considered as one of the main integrability features of the sine-Gordon equation (and the likes). On a general note, it relates a nonlinear equation (6.1) to the system of linear equations (6.5), which are amenable to analysis. In particular, the spectral theory of the first equation in (6.5) lies in the basis of the inverse spectral transformation approach to the solution of certain boundary value problems for the sine-Gordon equation. Also conserved quantities (integrals) of the sine-Gordon equation can be derived directly from its zero curvature representation. Furthermore, the zero curvature representation allows one to reconstruct a K-surface corresponding to a solution 1; of the sine-Gordon equation. Given a solution 1; : ]R2 ~ JR., introduce the matrices (6.2), (6.3) satisfying (6.4). Define the function '11 : JR.2 ~ G[A] as the solution of equations (6.5) with the initial condition '11(0,0; A) = 1. Then the immersion f : ]R2 ~ ]R3 obtained by the Sym formula, (6.6)
f(u) = 2AW-1(u; A)
8W~~; A) 1,\=1 '
under the canonical identification (4.9) of su(2) with ]R3, is an asymptotic lines parametrized K-surface, with the angle 1; between the asymptotic directions. The function '11 is known as the extended frame of f. Moreover, the right-hand side of (6.6) with various A not necessarily equal to 1 delivers a whole family of immersions fA : ]R2 ~ ]R3, all of which turn out to be asymptotic lines parametrized K-surfaces. These surfaces fA constitute the so-called associated family of f. The classical Backlund transformation is the next common feature of all known integrable systems. In the case of the sine-Gordon equation, it is given by the following construction. For a given solution 1; of (6.1), a
212
new solution equations:
6. Consistency as Integrability
(p+ can be found by solving the following system of differential
(6.7) This system is compatible, Eh(81 ¢+) = 8 1 (82 ¢+), provided ¢ is a solution of the sine-Gordon equation, and then ¢+ is also a solution. It is determined by the parameter n and the value ¢+(O,O) at one point. Geometrically, ¢+ is the angle between asymptotic directions of the Backlund transform f+ : IR2 ----) IR3 of a given K-surface f : IR2 ----) IR3 characterized, according to Definition 1.26, as follows: the straight line segments [J (u), f+ (u) 1 are tangent to both surfaces f and f+, and their length is independent of u. It can be checked by a direct computation that equations (6.7) are equivalent to the following matrix differential equations, which are satisfied identically with respect to the spectral parameter A:
(6.8)
81 W = U+W -
wu,
82 W = V+W -
wv,
where the matrix W is given by the formula
(6.9)
W=
(
ei(¢+-¢)/2
-inA
-inA
e- i (¢+-¢)/2
)
On the other hand, (6.8) constitute a solvability condition for the system consisting of (6.5) and similar equations for the matrix function I}i+ = WI}i.
(6.10) One can show that formed surface f+.
I}i+
serves as the extended frame of the Backlund trans-
A remarkable property of Backlund transformations is given by Bianchi's permutability theorem: if ¢(1) is a Backlund transformation of ¢ with parameter nand ¢(2) is a Backlund transformation of ¢ with parameter (3, then there exists a unique solution ¢(12) of the sine-Gordon equation which is simultaneously a Backlund transformation of ¢(1) with parameter (3 and a Backlund transformation of ¢(2) with parameter n; this solution is given by the formula
(6.11) So, integrable systems, for which the sine-Gordon equation is a prototypical example, are characterized by such features as zero curvature representation and Backlund transformations with permutability properties. The origin and the very existence of these features is considered in the classical theory of integrable systems as something mysterious and transcendental.
6.2. Discrete integrable systems
213
6.2. Discrete integrable systems The theory of discrete integrable systems has been developed for some time as part of the general theory of integrable systems. Its aims at the early stages were not very ambitious: just to find difference analogs of integrable differential systems, enjoying the same integrability features. In this introductory section we give an illustration by the example of the integrable discretization of the sine-Gordon equation, known as the Hirota equation: (6.12)
1 sin 4"(T1 T2¢ - T1¢ - T2¢ + ¢)
E2
1
= "4 sin 4"(T1 T2¢ + T1¢ + T2¢ + ¢).
Here ¢ is a real-valued function on (EZ)2, and the shift symbols stand for
Tk¢(U) = ¢(u + Eek). The Hirota equation (6.12) turns out to describe discrete K-surfaces, i.e., discrete A-surfaces f : (EZ)2 ----* IR3 with all edges of the same length E£, so that 18dl = 182 fl = £. Here, of course, 8k f(u) = (f(u + Eek) - f(u))/E. The discrete zero curvature representation of equation (6.12) is formulated in terms of the matrices U, V: (EZ)2 ----* G[A], defined by the formulas ei(Tl'P-¢)/2
(6.13)
U
iEA)
(
£, ;1.)
i(A
e_ il ".2".;)/2
'
2
1
(6.14)
~ e (T2H¢)/2) 2
(
V =
£2;1.)
"- e- ih q,+1>1/2 21.
1
'
2A where the normalizing factors £1 (A) = (1 + E2 A2 /4)1/2 and £2(A) = (1 + E2A- 2/4)1/2 are introduced in order to assure that U, V E G[A]. The matrix equation (6.15) is satisfied identically in A if and only if the function ¢ solves (6.12). Equation (6.15) is called a discrete zero curvature representation of the Hirota equation (6.12). It expresses the flatness of a discrete G[A]-valued connection, given by the matrices U assigned to the directed edges (u, u + Eel) and the matrices V assigned to the directed edges (u, u + Ee2) of the lattice (EZ)2; see Figure 6.1. In its turn, this condition assures the solvability of the following system of linear difference equations:
(6.16) for a function W : (EZ)2
T1 W = Uw, ----*
G[A].
T2W = VW
6. Consistency as Integrability
214
v u Figure 6.1. Discrete flat connection.
As in the continuous case, the discrete zero curvature representation can be used as a starting point for application of the analytical machinery of the inverse spectral methods. It also yields the conserved quantities (integrals) of the Hirota equation. Moreover, it can be used to reconstruct the underlying discrete K-surface, corresponding to a given solution ¢ : (fZ)2 ----+ lR of the Hirota equation, in literally the same fashion as in the smooth case. Given a solution ¢ of equation (6.12), introduce matrices (6.13), (6.14) satisfying (6.15). Define the function 1]/ : (fZ)2 ----+ G[A] as the solution of (6.16) with the initial condition 1]/(0,0; A) = 1. Then the Sym formula (6.6) determines a net j : (fZ)2 ---) lR3 , which is a discrete K-surface with the characteristic angle function ¢ and with the edge length ff, where £ = (1 +1'2/4)-1. Again, the right-hand side of (6.6) for various A not necessarily equal to 1 delivers an associated family 1>.. of discrete K-surfaces. The Backlund transformation for equation (6.12) is given by the following difference analogs of formulas (6.7): (6.17)
. 1 sm4:(T1¢+-¢++T1¢-¢)
I'
20:
sin 4:1(T1 ¢ + + ¢ + - T1 ¢ - ¢ ) ,
(6.18)
Statements analogous to those for the sine-Gordon equation hold. Difference equations (6.17), (6.18) are compatible; that is, T1(T2¢+) = T2(T1¢+), provided ¢ is a solution of (6.12), and then ¢+ is also a solution (determined by the parameter 0: and the value ¢+(O, 0) at one point). Also the geometric meaning of the Backlund transformation is similar to the continuous case: the straight line segments connecting the corresponding points of a discrete K-surface j and its Backlund transform j+ lie in the tangent planes of both surfaces, and their length is independent of U E (EZ)2. A direct computation shows that equations (6.17), (6.18) are equivalent to the matrix equations (6.19)
6.3. Discrete 2D integrable systems on graphs
215
which are satisfied identically in A, with the same matrix W as in (6.9). These equations assure the solvability of the system consisting of (6.16) and similar equations for the matrix function \[1+ defined by (6.10). This latter matrix \[1+ is nothing but the extended frame of the transformed surface. Bianchi's perIllutability theorem is formulated exactly as in the continuous case, and is expressed by the saIlle formula (6.11).
6.3. Discrete 2D integrable systems on graphs Before we turn to the explanation of the crucial idea that the 3D consistency property of 2D equations should be taken as the definition of their integrability, we provide a bit more details on the notion of integrability, corresponding to the traditional view of integrable systems, which is based on discrete zero curvature representations. This latter notion works in a more general context than systems on a regular square lattice 71}, namely it is naturally formulated for systems on graphs. A gmph 9 will mean for us not just a combinatorial object, but will be provided with an additional structure of a strongly regular polytopal cell decomposition of an oriented surface. The set of its vertices will be denoted by V(9), the set of its directed edges, by £7(9), and the set of its faces, by F(9). To any such 9 there canonically corresponds a dual cell decomposition 9*; it is only defined up to isotopy, but can be fixed uniquely with the help of the Voronoi-Delaunay construction. The vertices of 9* are in a one-to-one correspondence with the faces of 9 (actually, they can be chosen as some points inside the corresponding faces; cf. Figure 6.2).
X2
I I
"-
I
""-
.1'3
"-
I
Xl
/Yf, , /
/ /
X"
Figure 6.2. A facE' of 9 and the corrE'sponding vertE'X of 9*.
The variables of a discrete system (fields in the terminology of mathematical physics) will be understood as elements f of some set X (the phase
6. Consistency as Integrability
216
space of a system), assigned either to the vertices or to the edges of 9. (One can imagine also a mixed situation, where part of fields are assigned to the vertices and the others to the edges.) The system itself will be of the following nature. Consider a closed path of directed edges which constitute the boundary of a face of 9:
... , Then, in the case of fields assigned to the vertices, it is supposed that the fields f(xd, ... , f(xn) satisfy a certain condition, of a geometric or an analytic nature, called the equation associated to the face: Q(J(xd,···, f(xn)) = O.
(6.20)
If the fields are assigned to the edges, f(el), ... , f(e n ), then the equation should read correspondingly: (6.21 )
A discrete system is a collection of such equations associated with all faces of 9. One says that such a system admits a discrete zero curvature representation if there is a collection of matrices L(e;.\) E G[.\] from some loop group G[.\], associated with every directed edge e E £(9) (so called transition matrices), with the following properties. For a system with fields on vertices, L(e;.\) depends on the fields f(XI), f(X2) if e = (Xl, X2); for a system with fields on edges, L(e;.\) depends just on the field f(e). The argument.\ of the loops from G[.\] is known in the theory of integrable systems as the spectral parameter. It is required that: • for any directed edge e = (Xl, X2), if -e
L( -e,.\) = (L(e,.\)
(6.22)
= (X2, xd, then
rl;
• for any closed path of directed edges
... , we have (6.23) In the case when the path bounds a face of 9, the discrete zero curvature condition (6.23) must be equivalent to (or at least a consequence of) the equation for the corresponding face. Under conditions (6.22), (6.23) one can define a wave function W : V(9) - t G[.\] on the vertices of 9, by the following requirement: for any
217
6.4. Discrete Laplace type equations
directed edge e = (Xl, X2) E £(9), the values of the wave functions at its ends must be connected via (6.24)
For an arbitrary graph, the analytical consequences of the zero curvature representation for a given collection of equations are not clear. However, in the case of regular graphs, such as those generated by the square lattice Z + iZ c C, or by the regular triangular lattice Z + e27ri / 3 Z C C, such a representation may be used to determine conserved quantities for suitably defined Cauchy problems, as well as to apply powerful analytical methods for finding concrete solutions.
6.4. Discrete Laplace type equations There exist discrete equations on graphs which are not covered by the constructions of Section 6.3.
Definition 6.1. (Discrete Laplace type equations) Let 9 be a graph, with the set of vertices V(9) and the set of edges E(9). Discrete Laplace type equations on the graph 9 for a function f : V(9) - t C read: (6.25)
L
¢(j(xo), f(x); v(xo, x)) =
o.
xE star(xo)
There is one equation for every vertex Xo E V(9); the summation is extended over star(xo), the set of vertices of 9 connected to Xo by an edge (see Figure 6.3); the function ¢ depends on some parameters v : E(9) - t C, assigned to the edges of 9.
The classical (linear) discrete Laplace equations on 9 are a particular case of this definition: (6.26)
L
v(xo, x) (j(x) - f(xo)) = 0,
xE star(xo)
with some weights v : E(9)
-t
lR+ assigned to the (undirected) edges of 9.
The notion of integrability of discrete Laplace type equations is not well established yet. We discuss here a definition which is based on the notion of the discrete zero curvature representation and works under an additional assumption about the graph 9. Namely, like in the previous section, it has to come from a strongly regular polytopal cell decomposition of an oriented surface. We consider, in somewhat more detail, the dual graph (cell decomposition) 9*. Each e E E(9) separates two faces of 9, which in turn correspond to two vertices of 9*. A path between these two vertices is then declared the edge e* E E(9*) dual to e. If one assigns a direction to an edge
6. Consistency as Integrability
218
Xo
Xa
_--~-4':""
IXO
I I
6--
Y4 X4
Figure 6.3. Star of a vertex Xo in the graph
9·
X4
Figure 6.4. Face of 9* dual to a vertex Xo of 9.
e E E(9), then it will be assumed that the dual edge e* E E(9*) is also directed, in a way consistent with the orientation of the underlying surface, namely so that the pair (e, e*) is positively oriented at its crossing point. This orientation convention implies that e** = -e. Finally, the faces of 9* are in a one-to-one correspondence with the vertices of 9: if Xo E V(9), and Xl, ... , Xn E V(9) are its neighbors connected with Xo by the edges el = (Xo, Xl)"'" en = (Xo, Xn) E E(9), then the face of 9* dual to Xo is bounded by the dual edges ei = (YI, Y2), . .. , e~, = (Yn, yI); see Figure 6.4. We will say that a system of discrete Laplace type equations on 9 possesses a discrete zero curvature representation if there is a collection of matrices L(e*; A) E G[A] from some loop group G[A], associated to directed edges e* E E(9*) of the dual graph 9*, such that: • the matrix L(e*; A) depends on the fields f(xo), f(x) at the vertices of the edge e = (xo, x) E E(9), dual to the edge e* E E(9*), and • the flatness conditions (6.22), (6.23) on the dual graph are satisfied. The matrix L(e*; A) is interpreted as a transition matrix along the edge e* E E(9*), that is, a transition across the edge e E E(9). The wave function 'IJI in this situation is defined on the set V(9*) of vertices of the dual graph.
6.5. Quad-graphs Although one can consider 2D integrable systems on very different kinds of graphs on surfaces, there is one kind - quad-graphs -- supporting the most fundamental integrable systems.
219
6.5. Quad-graphs
Definition 6.2. (Quad-graph) A quad-graph'D is a strongly regular polytopaZ cell decomposition of a surface with all quadrilateral faces. Since we are interested mainly in the local theory of integrable systems of quad-graphs, and in order to avoid the discussion of some subtle boundary effects, we shall always suppose that the surface carrying the quad-graph has no boundary. Quad-graphs are privileged because from an arbitrary strongly regular polytopal cell decomposition 9 one can produce a certain quad-graph 'D, called the double of 9. The double 'D is a quad-graph, constructed from 9 and its dual 9* as follows. The set of vertices of the double 'D is V('D) = V(9) u V(9*). Each pair of dual edges, say e = (xo, Xl) E E(9) and e* = (Yl, Y2) E E(9*), defines a quadrilateral (xo, Yl, Xl, Y2). These quadrilaterals constitute the faces of a cell decomposition (quad-graph) 'D. Thus, a star of a vertex Xo E V(9) generates a flower of adjacent quadrilaterals from F('D) around Xo; see Figure 6.5. Let us stress that edges of 'D belong neither to E(9) nor to E(9*).
v--_XI
Figure 6.5. Faces of
1)
around the vertex Xo.
Quad-graphs 'D coming as doubles are bipartite: the set V('D) may be decomposed into two complementary halves, V(TJ) = V(9) UV(9*) ("black" and "white" vertices), such that the ends of each edge from E('D) are of different colors. Equivalently, any closed loop consisting of edges of 'D has an even length. The construction of the double can be reversed. Start with a bipartite quad-graph 'D. For instance, any quad-graph embedded in a plane or in an open disc is automatically bipartite. Any bipartite quad-graph produces two dual polytopal (in general, no more quadrilateral) cell decompositions 9 and
6. Consistency as Integrability
220
9*, with V (9) containing all the "black" vertices of 'D and V (9 *) containing all the "white" ones, and edges of 9 (resp. of 9*) connecting "black" (resp. "white") vertices along the diagonals of each face of 'D. The decomposition of V('D) into V(9) and V(9*) is unique, up to interchanging the roles of 9 and 9*. Notice that if a quad-graph 'D is not bipartite (i.e., if it admits loops consisting of an odd number of edges), then one can easily produce from 'D a new even quad-graph 'D', simply by refining each of the quadrilaterals from F('D) into four smaller ones. Since we are interested mainly in the local theory, we always assume (without mentioning it explicitly) that our quad-graphs are cellular decompositions of an open topological disc. In particular, our quad-graphs 'D are always bipartite, so that 9 and 9* are well defined.
6.6. Three-dimensional consistency An attentive examination of examples in Sections 6.1, 6.2 leads to remarkable observations which relate to the main philosophy of this book. For the continuous sine-Gordon equation the theory seems to consist of several components of a rather different nature: the main object is a partial differential equation, its Backlund transformations are described by a compatible system of two ordinary differential equations, while the superposition formula of Backlund transformations is expressed in purely algebraic terms. In the discrete context situation changes dramatically. All components of the discrete theory have essentially one and the same structure: equation (6.12) which describes discrete K-surfaces, equations (6.17), (6.18) for Backlund transformations of discrete K-surfaces, and equation (6.11) for the superposition principle ofthe latter. Their common structure is captured in the following formula for a function rP : zm ----) lR on an m-dimensionallattice:
(6.27)
. -1 ( rPjk sm
4
+ rPk - rPj - rP ) = -Ok.sm -1 (rPjk - rPk + rPj - rP ) . OJ
4
Here the subscript j stands for the shift in the j-th lattice direction, and parameters OJ are assigned to all edges parallel to the j-th lattice direction. Actually, in the geometric context, we are dealing with the case m = 4. The subscripts 1,2 label the coordinate directions of the discrete surfaces, while the subscripts 3,4 are used as replacements of the Backlund superscripts (1), (2). The relevant values of the parameters are: 01 = f/2, 02 = 2/f, 03 = 0, and 04 = /3. Equations (6.17), (6.11) are exactly of the form (6.27), and equations (6.12), (6.18) are brought into this form upon changing
6.6. Three-dimensional consistency
221
the sign of ¢ on every second hyperplane complementary to the second coordinate direction, i.e., upon the change of variables ¢( u) ----t (-1 )U 2¢( u). This reflects the fact that the underlying geometric properties of discrete K-surfaces and their Backlund transformations are identical and are captured in the definition of multidimensional K-nets, i.e., A-nets (nets in IR3 with planar vertex stars) satisfying the additional requirement that in every elementary quadrilateral the opposite sides have equal length. A discrete K-surface is a K-net with m = 2, iterated Backlund transformations of a discrete K-surface form a K-net with m = 3, while Bianchi's permutability theorem for two Backlund transformations of a discrete K-surface deals with K-nets with m = 4. The variable transformation Ijk
(6.28)
I
I = exp(i¢/2)
puts (6.27) into the form
CYjIj - CYkik CYjik - CYkIj ,
which is also known as the Hirota equation.
Ijk
Iijk
CYi
CYj Ijk
ik
CYi
ik
CYj
CYj Iik
CYk
I
CYk CYj I
CYk
CYkl
CYk
CYj
Ij
/'
~a
CYk Iij 2
CYj
/' I
Figure 6.6. 2D equation.
CYi
Ii
Figure 6.7. 3D consistency.
The Hirota equation (6.28) is a two-dimensional discrete equation, since it relates the variables I at the vertices of any elementary two-dimensional cell (square) of the m-dimensional lattice in such a way that any three variables determine the fourth one uniquely. The possibility to impose this equation everywhere on the m-dimensionallattice hinges on the case m = 3. The corresponding property of three-dimensional consistency should be understood as follows: suppose that four values I, Ii, Ij, ik are given (consult Figure 6.7 for notation). Then equation (6.28) defines Iij, Ijk and Iik' and a further application of this equation gives three a priori different values of Iijk. These three values turn out to automatically coincide for arbitrary
6. Consistency as Integrability
222
initial data. Indeed, a direct computation shows:
(6.29)
r.
k tJ -
+ ai(a~ - aJ)fj/k + aj(ar - a~)fkfi aT)fk + ai(a~ - aJ)fi + aj(ar - a~)fj
ak(o:; - ar)fdj
--=---;c:------;:-:----...."----=--,,.----,----~____,,_,--
ak(aJ -
This coincidence is the meaning of the 3D consistency of the Hirota equation. As a consequence, the Hirota equation can be consistently imposed on all elementary squares of a multidimensional lattice.
6.7. From 3D consistency to zero curvature representations and Backlund transformations Now we are in a position to expose the main idea concerning the understanding of discrete integrable systems, namely that the property of 3D consistency observed in Section 6.6 for the Hirota equation is actually of a fundamental importance and leads directly to the core of the whole theory. We show that other features of integrable systems, such as zero curvature representations and Backlund transformations, are consequences of 3D consistency. The present section is devoted to a realization of this idea for systems on quad-graphs with fields on vertices and with labelled edges. A typical representative of this class of equations is the Hirota system, which we write here once more in the form (6.30)
h2 f
a2!2 ad2 - a2h'
adl -
In the geometric context of K-surfaces we had f = exp(i¢/2) E §1. In the present analytic study we will assume the fields f to be any complex numbers assigned to the vertices of 'I}, while ai are (complex) parameters naturally assigned to the edges of Z2 parallel to 'Bi and constant along the strips in the complementary direction. In a different fashion, one can view ai as fields satisfying the labelling property (6.31)
(6.32) One more example of such a system with vertex variables and edge parameters having the labelling property is given by the cross-ratio equation: (6.33) We already studied this system in Section 4.3 in the context of discrete isothermic surfaces, where the fields f are points of]RN and parameters ai
6.7. From 3D consistency to zero curvature representations
223
are real numbers, the cross-ratio being defined according to the Clifford multiplication in e£(Il~N). Here we will consider a simpler version with fields and parameters being complex numbers. The commutativity of complex multiplication makes the check of the 3D consistency of the cross-ratio equation the matter of a straightforward computation, leading to (6.34)
h23 = (0:1 - 0:2)hh + (0:2 - 0:3)1213 + (0:3 - O:l)hh . (0:2 - 0:1)13 + (0:3 - 0:2)h + (0:1 - 0:3)12
A general system of this class consists of equations (6.35) Here j : Z2 ----) C are complex fields, and O:i are complex parameters on the edges of Z2 parallel to 'B i , satisfying the labelling condition (6.31); see Figure 6.6. Actually, just from the outset we would like to generalize this setup by considering systems on arbitrary quad-graphs instead of Z2. In this case (6.35) should be read as a relation for fields j : V('D) ----) C, with 0: : E('D) ----) C being a labelling of edges of 'D, i.e., a function taking equal values on any pair of opposite edges of any quadrilateral from F('D). In the context of equations on general quad-graphs, there are no distinguished coordinate directions; nevertheless it will be convenient to continue to use notation of (6.35), with the understanding that indices are used locally (within one quadrilateral) and do not stand for shifts into globally defined coordinate directions. So, j, h, h2, 12 can be any cyclic enumeration of the vertices of an elementary quadrilateral. Sometimes we will stress the absence of global coordinate directions by writing (6.35) in a different system of notation, using just a cyclic enumeration of vertices: (6.36) see Figure 6.8.
(3
Figure 6.8. A face of a labelled quad-graph; fields on vertices.
6. Consistency as Integrability
224
For the very possibility to pose equation (6.36) on general quad-graphs, this equation should be uniquely solvable for anyone of its arguments fi E C; therefore the following assumption is natural by considerations in that generality: Linearity. The function Q is a polynomial of degree 1 in each argument fi (multiaffine), with coefficients depending on the parameters a, (3:
(6.37)
Q(h, 12, 13, f4; a, (3) =
al(a, (3)h1213f4
+ ... + a16(a, (3).
For the Hirota equation (6.30) one can take
while for the cross-ratio equation (6.33) with complex-valued arguments one can take
Assume now that equation (6.35) possesses the property of 3D consistency; cf. Figure 6.7. We will demonstrate that this remarkable property automatically leads to two basic structures associated with integrability in the soliton theory: Backlund transformations and zero curvature representation.
Theorem 6.3. (3D consistency yields Backlund transformations) Let equation (6.35) be 3D consistent. Then for any solution f : V('D) --t C of the corresponding system (6.36) on a quad-graph 'D, there is a two-parameter family of solutions f+ : V('D) ----; C of the same system, satisfying (6.38) for all edges (J,Ji) E E('D). Such a solution f+ is called a Backlund transform of f, and is determined by 'its value at one vertex of 'D and by the parameter .A.
Proof. We formally extend the planar quad-graph 'D into the third dimension. For this aim, consider the second copy 'D+ of 'D and add edges connecting each vertex f E V('D) with its copy f+ E V('D+). (We slightly abuse the notation here, by using the same letter f for vertices of the quadgraph and for the fields assigned to these vertices.) On this way we obtain a "3D quad-graph" D, with the set of vertices V(D) = V('D)
u V('D+),
with the set of edges E(D) = E('D) U E('D+) U {(J, f+) : f E V('D)},
6.7. From 3D consistency to zero curvature representations
225
and with the set of faces
F(D) = F(1J) UF(1J+) U {(J,fr,ft,f+): f,fr E V(1J)}. Elementary building blocks of D are combinatorial cubes (j, fro fr2' 12, f+, ft, f0., fi)' as shown in Figure 6.9. The labelling on E(D) is defined in the natural way: each edge (J+, f i+) E E(1J+) carries the same label O'i as its counterpart (J, Ii) E E(1J), while all "vertical" edges (J, f+) carry one and the same label A. Clearly, the content of Figure 6.9 is the same as of Figure 6.7, up to notation. Now, a solution f+ : V(1J+) -----t C on the first floor of D is well defined due to the 3D consistency, and is determined by its value at one vertex of 1J+ and by A. We can assume that f+ is defined 011 V(1J) rather than 011 V(1)+), since these two sets are in a one-to-one 0 correspondence.
fi
0'1
0'2
f0. 0'2
f+
ft
0'1
A
A: I I I
A
A
12 __ --- ~1_
fr2
/
/ / 0'2
f
0'2
0'1
fr
Figure 6.9. Elementary cube of D.
Theorem 6.4. (3D consistency yields zero curvature representation) Let equation (6.35) be 3D consistent. Then the corresponding system (6.36) on a quad-graph 1) admits a zero curvature representation with spectral parameter dependent 2 x 2 matrices: there exist matrices associated to directed edges of 1),
L(e, O'(e); A) : £(1))
(6.39)
-----t
GL(2, A-I, so that equations on the vertical faces of Figure 6.9 read:
This gives the Mobius transformation (6.41) with (6.46) The determinant of this matrix is constant (equal to 1 - AOOi); therefore no further normalization is required. A more usual form of the transition matrices of the zero curvature representation for the complex cross-ratio equation is obtained by the gauge transformation
AU) =
G{),
which leads to the matrices (6.47) These matrices (6.47) are interpreted as matrices of the Mobius transformations acting on the shifted quantities:
To summarize: 3D consistency of 2D quad-equations with complex fields on vertices and with labelled edges implies the existence of Backlund transformations and of the zero curvature representation. This is not a pure existence statement but rather a construction: both attributes can be derived in a systematic way starting with no more information than the equation itself, they are in a sense encoded in the equation provided it is 3D consistent.
6.8. Geometry of boundary value problems for integrable 2D equations There are several important aspects of the problem of embedding of a quadgraph into a regular multidimensional square lattice, related to integrable equations.
6. Consistency as Integrability
228
6.8.1. Initial value problem. We start with the question of correct initial value (Cauchy) problems for discrete 2D equations on quad-graphs. Let P be a path in the quad-graph 'D, i.e., a sequence of edges Cj = (Xj, xj+d E E('D). We denote by E( P) = {Cj} and V (P) = {x j} the set of edges and the set of vertices of the path P, respectively. One says that the Cauchy problem for a path P is well posed if for any set of data f p : V (P) ----) C there exists a unique solution f : V('D) ----) C such that V(P)= fp. It is not difficult to find examples of paths on the square lattice for which the Cauchy problem is well posed. The task we are interested in is to characterize, for a given quad-graph 'D, all paths with this property.
n
-
-
f--- -
-~
-
f-- - --
-
- - f--~
c_ _ _
Figure 6.10. One-corner initial path.
~~
__
u
Figure 6.11. Staircase initial path.
A solution of this problem can be given with the help of the notion of a strip in 'D. Definition 6.5. (Strip) A strip in 'D is a sequence of quadrilateral faces qj E F('D) such that any pair qj-I, qj is adjacent along the edge Cj = qj-I n qj, and Cj, CHl are opposite edges of qj. In other words, a strip is a path in 'D* consisting of edges cj = (qj-l, qj) E E('D*) such that any consecutive pair cj, Cj+l enter and leave the quadrilateral qj along a pair of opposite edges Cj, CHI. The edges Cj are called traverse edges of the strip. So, in a labelled quad-graph 'D any strip may be associated to a label a sitting on all its traverse edges Cj. The strips come to replace coordinate directions in a regular square lattice, and can be considered as a discrete analog of characteristics for hyperbolic systems of partial differential equations with two independent variables. Theorem 6.6. (Well posed Cauchy problems on quad-graphs) Let'D be a finite simply connected quad-graph without self-crossing strips, and let P be a path without self-crossings in 'D. Consider a 3D consistent equation of the type (6.35) on the quad-graph 'D. Then:
6.S. Boundary value problems for integrable 2D equations
229
i) If each strip in 'D intersects P exactly once, then the Cauchy problem for P is well posed. ii) If some strip in 'D intersects P more than once, then the Cauchy problem for P is overdetermined (has in general no solutions).
iii) If some strip in 'D does not intersect P, then the Cauchy problem for P is underdetermined (has in general more than one solution). Proof. We shall only sketch the proof of the claim i). It is based on an embedding T of V ('D) into the unit cube of '!In, where n is the number of edges in P (the number of distinct strips in 'D). Choose any vertex :£0 E V('D), and set T(xo) = E '!In. The image of any other vertex x E V('D) is defined recurrently along a path connecting Xo to x with the help of the following rule:
°
For any two neighbors x, y E V('D), if the edge (x, y) E E('D) belongs to the strip number i E {1, 2, ... , n}, then T(y) = T(x) +ei (mod 2), where ei is the i-th coordinate vector of '!In. The result does not depend on the path connecting x to xo, since any closed path has an even intersection index with any strip; therefore contribution of any strip to T along a closed path vanishes. Edges and faces of 'D correspond to edges and two-faces of the unit cube in '!In. The T-image of the path P is the path (0,0,0, ... ,0), (1,0,0, ... ,0), (1,1,0, ... ,0), ... , (1,1,1, ... ,1). It is clear that for a 3D consistent equation the data along this path define a well-posed Cauchy problem for the unit cube in '!In. In particular, these data uniquely determine the values of the solution on T(V('D)). 0 It should be mentioned that this theorem is not valid for equations without the 3D consistency property. The next theorem is based on the zero curvature representation with a spectral parameter; therefore it is also specific for 3D consistent equations on labelled quad-graphs. We will formulate this theorem for a concrete equation (cross-ratio equation), but actually it applies under much more general circumstances. See, however, Exercise 6.5, illustrating an instance where this theorem is not valid. Theorem 6.7. (Relating data on two Cauchy paths) Consider a generic solution of the cross-ratio equation (6.33) on a simply connected quad-graph 'D. Let each of the two paths P = (xo, Xl, ... , x n ) and P = (xo, Xl, ... ,xn ) in 'D with a common starting point Xo = Xo and a common end point Xn = xn intersect each strip in 'D exactly once. Then the the fields (fo, iI,···, fn) along P determine the fields (fo, 11, ... , 1n) along P uniquely, as soon as the sequences of labels ai = a(Xi-l, Xi) along P and ai = a(Xi-l, Xi) along P are known, that is, without knowing any additional information on the combinatorics of 'D.
6. Consistency as Integrability
230
Proof. The proof is based on the zero curvature representation of the crossratio equation with the transition matrices L given in (6.47). By the hypothesis of the theorem, the sequence (al, ... , an) is a permutation of the sequence (al, ... , an). From the zero curvature condition (6.23) it follows that (6.48) Generally speaking, such an equality does not hold automatically for nonnormalized transition matrices, but in our case det L(fi, fi-1, ai; A) = 1 Aai, which yields the equality of determinants of the both sides of (6.48). Denote the left-hand side of (6.48) by
IIL(fi, ii-I, ai; A).
T(A) =
All entries of this matrix are polynomials in A. We want to show that this matrix can be uniquely refactorized as
T(A) =
II L(fi, h-1' ai; A),
with io = fo, in fn, and with a prescribed permutation (ai) of the parameters (ai) along the path P. We show that there is a unique matrix of the form
L(f1 ,fo, a1; A)
= (
Aa 1(fo
~ id- 1
fo
~ fr)
such that all entries of the matrix T(A)L- l (f1, io, a1; A) are polynomials in A. Since det T(A) = I1(1 - Aad, the points A = ail are exactly those where T(A) is degenerate. For a generic solution, rank T(ail) = 1, so that dim ker T (ai 1) = 1. Define by
11
kerT(all) (recall that io
= fo).
= lR (i1
~ io)
Then the elements of the vector
T(A) (i1
~ io)
are polynomials divisible by 1 - Aal. Now observe that -
L -l(fl ,fo, al; A) = 1 -\a1 (fl
-
~ fO)
-
((fl
-
T
-/0)-1) -
((fl _ °io)-1
which immediately implies that T(A)L -1 (f1,fo, al; A) is a polynomial in A. An inductive application of this procedure yields the desired refactorization of the matrix T(A). It remains to show that for the so found sequence
6.8. Boundary value problems for integrable 2D equations
231
eli), we have ln = fn. But this follows immediately from the fact that the free term of the (12) entry of T().) is equal to fo - fn = In. This finishes the proof. D
10 -
This theorem has rather surprising consequences. Consider a quad-graph obtained from the regular square lattice by replacing some Tn x n rectangle by a finite simply connected quad-graph with the same boundary vertices. The resulting quad-graph is called a regular "quare lattice with a localized defect. We say that a defect is weak if all strips entering the defect leave it in the same direction, possibly in a different order. Figure 6.12 illustrates a weak 3 x 2 defect.
~
Figure 6.12. A weak localized defect in the regular square lattice.
Consider a Cauchy problem for the cross-ratio equation on a regular lattice with a weak defect, with the initial data outside of the defect. Suppose that all horizontal edges outside of the defect carry the same label 0' and all vertical edges outside of the defect carry the same label {3 (so that in Figure 6.12 there should be 0'1 = 0'2 = 0'3). Compare the solution of this problem with the solution of the same Cauchy problem but on the regular square lattice without defects. Surprisingly, as a consequence of Theorem 6.7, the solutions will coincide outside of the defect. One can say that for the cross-ratio equation (and the likes) with a homogeneous labelling the weak defects are transparent. 6.8.2. Extension to a multidimensional lattice. The problem of embedding of a quad-graph 'D into a regular multidimensional cubic lattice has also aspects of a different flavor.
6. Consistency as Integrability
232
Theorem 6.8. (Rhombic embedding) A quad-graph 'D admits an embedding in C with all rhombic faces if and only if the following two conditions are satisfied:
i) No strip crosses itself or is periodic. ii) Two distinct strips cross each other at most once. For a proof of this theorem we refer the reader to Kenyon-Schlenker (2004). We will show that rhombic embeddings are closely related to 3D consistency of equations on 'D. Given a rhombic embedding p : V('D) ---4 C with edges of unit length (which can always be achieved by scaling and will be assumed from now on), one defines the following function on the directed edges of 'D with values in §l = {O E C : 101 = I}: (6.49)
O(x, y) = p(y) - p(x),
\7'(x, y)
E
E('D).
This function can be called a labelling of directed edges, since it satisfies O(-e) = -O(e) for any e E E('D), and the values of 0 on two opposite and equally directed edges of any quadrilateral from F('D) are equal. See Figure 6.13. For any labelling 0 : E('D) ---4 §l of directed edges, the function 0: = 0 2 : E('D) ---4 §l is a labelling of (undirected) edges of'D in our usual sense.
f Figure 6.13. Labelling of directed edges.
Definition 6.9. (Quasicrystallic rhombic embedding) A rhombic embedding p : V('D) ---4 C of a quad-graph 'D is called quasicrystallic if the set of values of the function 0 : E('D) ---4 §l defined by (6.49) is finite, say
e=
{±Ol, ... , ±Od}.
An example of a quasicrystallic (actually periodic) rhombic quad-graph with d = 3 is the so-called dual kagome lattice shown in Figure 6.14. A prototypic example of a nonperiodic quasicrystallic rhombic quadgraph with d = 5 is the famous Penrose tiling shown in Figure 6.15.
6.8. Boundary value problems for integrable 2D equations
233
Figure 6.14. Dual kagome lattice.
Figure 6.15. Penrose rhombic tiling.
It is of a central importance that any quasicrystallic rhombic embedding p can be seen as a sort of projection of a certain two-dimensional sub complex
(combinatorial surface) 01) of a multidimensional regular square lattice 7l,d. The vertices of 01) are given by a map P : V(1)) --t 7l,d constructed as follows. Fix some Xo E V(1)), and set P(xo) = o. The images in 7l,d of all other vertices of 1) are defined recurrently by the property: For any two neighbors x, y E V(1)), if p(y) - p(x) = ±Oi E e, then P(y) - P(x) = ±ei, where ei is the i-th coordinate vector of 7l,d. The edges and faces of 01) correspond to the edges and faces of 1), so that the combinatorics of 01) is that of 1).
6. Consistency as Integrability
234
To exploit possibilities provided by the 3D consistency, we extend the labelling () : £('D) -+ §1 to all edges of'lL,d, assuming that all the edges parallel to (and directed as) ek carry the label ()k. This gives, of course, also the labelling a = ()2 of undirected edges of 'lL,d. Now, any 3D consistent equation can be imposed not only on 01), but on the whole of'lL,d: (6.50)
1
:s; j
-1= k
:s; d.
Here indices stand for the shifts into the coordinate directions. Obviously, for any solution f : 'lL,d -+ C of (6.50), its restriction to V(01)) rv V('D) gives a solution of the corresponding equation on the quad-graph 'D. As for the reverse procedure, i.e., for the extension of an arbitrary solution of (6.36) from 'D to 'lL,d, more thorough considerations are necessary. An elementary step of such an extension consists in finding f at the fourth vertex of an elementary square from the known values at three vertices according to (6.50). Due to 3D consistency this extension is well defined. In particular, one can find f at the eighth vertex of an elementary 3D cube from the known values at seven vertices; see Figure 6.16. This can be alternatively viewed as a flip (elementary transformation) on the set ofrhombically embedded quadgraphs 'D, or on the set of the corresponding surfaces 01) in 'lL,d. Any quadgraph'D (or any corresponding surface 01)) obtainable from the original one by such flips, carries a unique solution of (6.50) which is an extension of the original one.
Figure 6.16. Elementary flip.
Definition 6.10. (Hull) For a given set V C 'lL,d, its hull Je(V) is the minimal set Je C 'lL,d containing V and satisfying the condition: if three vertices of an elementary square belong to Je, then so does the fourth vertex.
'lL,d
One shows by induction that for an arbitrary connected sub complex of with the set of vertices V, its hull is a brick
(6.51)
6.9. 3D consistent equations with noncommutative fields
235
where (6.52) and in the case that nk are unbounded from below or from above on V, we set ak(V) = -00, resp. bk(V) = 00. Combinatorially, all points of the hull JC(V(n1»)) can be reached from by the extension procedure described above. However, there might be obstructions for extending solutions of (6.36) from a combinatorial surface (two-dimensional subcomplex of Zd) to its hull, having nothing to do with 3D consistency. For instance, the surface n shown in Figure 6.17 supports the solutions of (6.36) which cannot be extended to the solutions of (6.50) on the whole of JC(V(n)): the recursive extension will lead to contradictions. The reason for this is nonmonotonicity of n: it contains pairs of points which cannot be connected by a monotone path in n, i.e., by a path in n with all directed edges lying in one octant of Zd. However, such surfaces n do not come from rhombic embeddings, and in the case of n1) there will be no contradictions.
n1)
Figure 6.11. A nonmonotone surface in Z3.
Theorem 6.11. (Extension of solutions from quad-surfaces to Zd) Let the combinatorial surface n1) in Zd come from a quasicrystallic rhombic embedding of a quad-graph 'D, and let its hull be JC(V(n1»)) = IIa,b. An arbitrary solution of a 3D consistent equation (6.36) on n1) can be uniquely extended to a solution of equation (6.50) on IIa,b' The proof of this theorem can be found in Bobenko-Mercat-Suris (2005).
6.9. 3D consistent equations with noncommutative fields The validity of the message formulated in the last paragraph of Section 6.7, saying that 3D consistency of a quad-equation yields a construction of Backlund transformations and of the zero curvature representation, is by no
6. Consistency as Integrability
236
means restricted to the situation for which it was demonstrated (complex fields on vertices). In the present section, we show that it can be extended to equations with fields on vertices taking values in some associative but noncommutative algebra A with unit over a field X, and with edge labels with values in X. The transition matrices of the zero curvature representation are in this case 2 x 2 matrices with entries from A. They act on A according to (6.42), where now the order of the factors is essential. Actually, the proof of Theorem 4.26 given in Section 4.3.7 is an example of a derivation of a zero curvature representation for the cross-ratio equation (6.33) with fields in A = e£(JR N ) and with parameters ai from X = JR, which governs discrete isothermic surfaces in JRN (one has to interpret the arbitrary parameter a3 in equation (4.96) as the spectral parameter >.). The literal generalization of this proof for an arbitrary associative algebra A leads to the following statement.
Theorem 6.12. (Cross-ratio equation in an associative algebra) The cross-ratio equation in an associative algebra A is 3D consistent. It possesses a zero curvature representation with transition matrices (6.47), where the inversion is treated in A. We provide here two more examples of similar results for equations with noncommutative fields. The first will be dealing with the noncommutative Hirota equation. It turns out that the correct way to write such a noncommutative generalization is the following: (6.53)
Theorem 6.13. (Hirota equation in an associative algebra) The noncommutative Hirota equation (6.53) is 3D consistent. It admits a zero curvature representation with the transition matrices (6.54)
Proof. The noncommutative Hirota equation on the face (f, fi, lij, fj) can be written as a formula which gives fij as a linear-fractional transformation of fj: (6.55)
where (6.56)
L(fi,f,ai,aj) =
(
aj I -ad-
-adi ajf-Ifi
).
Here we use the same notation as in the proof of Theorem 4.26 given in Section 4.3.7 for the action of of the group GL(2,A) on A. Thus, equation
6.9. 3D consistent equations with noncommutative fields
(6.53) on the faces as
e 13 , e 23
of the elementary 3D cube
e l23
237
can be written
L(h, j, aI, (3)[h] , L(12, j, a2, (3)[h]·
(6.57) (6.58)
By the shift in the second, resp. the first, coordinate direction we derive the expressions for h23 obtained from equation (6.53) on the faces 72e13, 71 e 23, respectively:
(6.59) (6.60)
h23 h23
L(J12, 12, aI, (3)[123], L(J12, h, a2, (3)[jd·
Substituting (6.57), (6.58) on the right-hand sides of equations (6.60), (6.59), respectively, we represent the equality between the two values of h23 (which we want to demonstrate) in the following matrix form:
In fact, a stronger claim holds, namely (6.61) L(J12, h, a2, (3)L(h, j, aI, (3) = L(J12, 12, aI, (3)L(12, j, a2, (3)' Indeed, the (11) entries on both sides of this matrix identity are equal to a~ + ala2!12j-l. Equating (12) entries on both sides is equivalent to the Hirota equation of the face (J, h,J12, h), and the same holds for the (21) entries. Finally, equating the (22) entries is equivalent to the condition that h2!-1 commutes with 12[1 1 , and this is, of course, true by virtue of (6.53). This proves the 3D consistency of the noncommutative Hirota equation. The claim about the zero curvature representation is nothing but relation (6.61) just proven with a3 replaced by A. 0 We consider here one more equation of this kind: (6.62) with the vertex variables j taking values in A and with the edge labels a from X. In the case of real-valued fields f, this equation is known under the name of the discrete KdV equation; among other things, it expresses the Bianchi-type superposition formula for the Backlund transformations of the Korteweg-de Vries equation. In the case when the fields j are considered to belong to ]RN C A = e£(]RN), the solutions of this equation are special T-nets in ]RN: (6.63)
In the vector form (6.63) this equation is known as the discrete Calapso equation.
6. Consistency as Integrability
238
Theorem 6.14. (Discrete KdV equation in an associative algebra) Equation (6.62) in an associative algebra A with unit is 3D consistent. It possesses a zero curvature representation with the transition matrices (6.64)
Proof. Equations (6.62) on the vertical faces of Figure 6.9 read:
This gives the transition matrices, which can then be used to prove the 3D consistency, in the same manner as in the proof of Theorem 4.26 given in Section 4.3.7 and in the proof of Theorem 6.13. 0 Our last example in this section is of a geometric origin and of a slightly different nature than the previous ones. In Section 4.1, in our study of T-nets in quadrics, we encountered the equation with vertex variables f : 71} ---7 Q = {f E jRN : (j, f) = ",2}: (6.65)
h2 - f
=
a(12 - h),
a
(j,h - h) = ---'-:-----'---''",2 - (jl, h)
2(j,h - h)
Ih-121 2
.
A priori it does not contain any parameters. However, the quantities
(Xi
=
2(j, fi), being functions of the vertex variables f rather than parameters of the equation, possess the labelling property (6.31). Comparing (6.65) with (6.63), we see that the former can be regarded as a particular instance of the latter.
Theorem 6.15. (T-nets in quadrics) Equation (6.65), describing T-nets in quadrics, is 3D consistent. It possesses a zero curvature representation with transition matrices with entries from eC(jRN): (6.66)
Proof. 3D consistency has been proven geometrically in Theorem 4.3. As for the transition matrices, we can take those from (6.64) with (Xi
= 2(j, fi) = - f fi - fif·
Note the geometrical meaning of the spectral parameter: A = 2(j, f+) for the Backlund transformation f+ from which the transition matrices are constructed. 0
239
6.10. Classification of discrete integrable 2D systems. I
6.10. Classification of discrete integrable 2D systems with fields on vertices. I The notion of 3D consistency, being fundamental for definition and study of 2D integrability, proves extremely useful also in various classification problems of the integrable systems theory. Because of its constructive nature, it can be put into the basis of classification within certain Ansiitze. We will present here the solution of a very general problem concerning the 3D consistent systems of (possibly different) quad-equations with complex fields on vertices. Quad-equations will be of the form (6.67) where the field variables Xi E ClP'l are assigned to the vertices of a quadrilateral (ordered cyclically), and Q satisfies only one assumption, namely that of linearity, formulated already in Section 6.7: the function Q is supposed to be an irreducible polynomial of degree 1 in each variable. This implies that equation (6.67) can be solved for any variable, and the solution is a rational function of the other three variables. The problem we would like to solve is that of the 3D consistency of six a priori different quad-equations put on the faces of a coordinate cube: the system of six quad-equations, (6.68)
A(x, Xl, X12, X2) = 0, B(x, X2, X23, X3) = 0, C(x, X3, X23, Xl) = 0,
0, 0, = 0,
A(X3, X13, X123, X23) =
B(Xl' X12, X123, X13) = C(X2' X23, X123, X12)
should admit a unique solution X123 for arbitrary initial data X, Xl, X2, X3; see Figure 6.18. The functions A, . .. ,C are affine linear and a priori are not supposed to be related to each other in any way. In solving a classification problem, one should factor out a possibly large group of transformations that leave the class of objects being classified invariant. In our problem each quad-equation preserves its form under the group (Mob)4 which acts by Mobius transformations on all the vertex fields independently. It will be convenient to denote by p~ the set of polynomials in n variables which are of degree m in each variable, with the following action of Mobius transformations on a polynomial P E p~:
where (6.69)
6. Consistency as Integrability
240
X123
A X3 e:_-------'------'-----. X13
C B - -- ~ - -----,;/ '- - -- -- - f3 , ,
:c .---~---------
/
X2
A
x Figure 6.18. A 3D consistent system of six different quad-equations: A and A are associated to the bottom and to the top faces of the cube, Band E, to the left and to the right faces, and C and C, to the front and to the back ones.
Pi.
Thus, quad-equations (6.67) are characterized by polynomials Q E An important step in the solution of our problem will be classifying such polynomials and finding their normal form modulo the action of (Mob)4. The full problem we are aiming at is classifying and finding normal forms for 3D consistent systems (6.68) modulo the action of (Mob)8 (independent Mobius transformations of all eight vertices of the cube). We will solve these problems under certain nondegeneracy conditions. In formulation of these conditions, as well as in the whole theory, the following operations play an important role: (6.70) (Variables placed as subscripts stand for partial differentiations.) The operation bXi,xj applied to an affine linear polynomial Q(Xl, X2, X3, X4) eliminates the variables Xi, X j, the result being a biquadratic polynomial of the remaining variables Xk,Xl (so that {i,j,k,l} = {1,2,3,4}), which we will denote by hkl(Xk, Xl) = hlk(Xk, xt). Thus, from any Q E the operations bXi,xj produce six biquadratic polynomials hkl E P~, four of them corresponding to the edges of the underlying quadrilateral, and the remaining two corresponding to the diagonals. Note that the operations bXi,xj are covariant with respect to Mobius transformations:
Pi
(6.71) with
~i
given in (6.69).
6.10. Classification of discrete integrable 2D systems. I
241
Definition 6.16. (Nondegenerate biquadratic) A biquadratic polynomial h(x, y) E P~ is called nondegenerate if no polynomial in its equivalence class with respect to Mobius transformations is divisible by a factor x - c or y - c (with c = canst). Thus, a polynomial h(x, y) E P~ is nondegenerate if it is either irreducible or of the form (alxy + f31x + '"'fly + 6d(a2xy + f32x + "(2Y + 62) with ai6i - f3i'Yi i= O. In both cases the equation h = 0 defines Y as a two-valued function on x and vice versa. An example of a degenerate biquadratic is given by h(x, y) = x - y2 (considered as an element of P~), since under the inversion x f---> l/x it turns into x(l - xy2).
Definition 6.17. (Quad-equation of type Q) A multiaffine function Q E Pi, or the corresponding quad-equation Q = 0, is said to be of type Q if its four accompanying edge biquadratics h jk E P~ are nondegenerate, and it is said to be of type H otherwise. It turns out that multiaffine equations of type Q admit an exhaustive classification modulo (Mob)4, with only four normal forms.
Theorem 6.18. (Classification of equations of type Q) Any multiaffine equation Q(Xl' X2, X3, X4) = 0 of type Q is equivalent, modulo Mobius transformations from (Mob)4 acting on each of the variables independently, to one of the equations in the following list, called the list Q: (Q4)
sn( a)(xlx2 +X3X4) - sn(f3) (XIX4 + X2X3) - sn( a - (3)(XIX3 +X2X4) + sn(a - (3) sn(a) sn(f3)(l + k2xIX2X3X4) = 0,
(Q3)
sin( a )(XIX2+X3X4) -sin(f3) (XIX4 +X2X3) -sin( a- (3)(XIX3 +X2X4) +6 sin(a - (3) sin(a) sin(f3) = 0,
(Q2)
a(xlx2 + X3X4) - f3(XIX4 + X2X3) - (a - (3)(XIX3 + X2X4) +af3(a-f3)(xl+X2+X3+x4)-af3(a-f3)(a2-af3+f32) = 0,
(Q1)
a(xlx2 + X3X4) - f3(XIX4 +6a;3( a - (3) = o.
+ X2X3)
- (a - (3)(XIX3
+ X2X4)
In equation (Q4) the notation sn(a) = sn(a; k) is used for the Jacobi elliptic function with modulus k. The parameter 6 in equations (Q3), (Ql) can be scaled away, so that one can assume without loss of generality that 6 = 0 or 6 = 1. It is important to observe that there were no a priori built-in parameters a, 13 in the polynomial Q E Pi; they appear in the course of classification. They turn out to be naturally assigned to the edges of the quadrilateral (Xl, X2, X3, X4). Equation (Q4) is the most general one of the list; it is parametrized by two points of an elliptic curve. Equations (Q1)-(Q3) are obtained from (Q4) upon degenerations of an elliptic curve into rational
6. Consistency as Integrability
242
curves. One could be tempted to reduce the list Q to one item (Q4). However, the limit procedures necessary for that are outside of our group of admissible (Mobius) transformations, and, on the other hand, in many situations the "degenerate" equations (Q1)-(Q3) are of interest by themselves (for instance, the simplest equation of the list, (Q1) with 6 = 0, is nothing but the complex cross-ratio equation). This resembles the situation with the list of the six Painleve equations and the coalescences between them. It remains to find out how the equations of Theorem 6.18 can be assembled into 3D consistent systems on a cube.
Theorem 6.19. (3D consistent systems of type Q) Each equation of the list Q is 3D consistent. Conversely, any 3D consistent system (6.68) with all six equations of type Q is equivalent, modulo Mobius transformations from (M ijb)8 acting on each variable independently, to the system
(6.72) where (ij k) stands for any of the three cyclic permutations of (123), and Q(Xl, X2, X3, X4; ex, (3) is one of the polynomials (Q1)-(Q4)· In the next section, we sketch the main ideas behind the proof of this remarkable result.
6.11. Proof of the classification theorem 6.11.1. 3D consistent systems, biquadratics and tetrahedron property. Biquadratic polynomials h ij for a given Q E :P~ are closely related to the so-called singular solutions of the basic equation (6.67). Generically, the equation Q(XI, X2, X3, X4) = 0 can be solved with respect to any variable: if Q = p(Xj, Xb Xl)Xi + q(Xj, Xk, Xl) then Xi = -q/p for generic values of Xj, xk, Xl. However, Xi is not determined if the point (Xj, Xk, Xl) lies on the curve Si in (CJP'1)3 defined by (6.73)
Si:
p(Xj, Xk, Xl) = q(Xj, Xk, Xl) = O.
Since p = QXi and q = Q - XiP = Q - XiQx;, equations (6.73) are equivalent to Q(XI,X2,X3,X4) = Qx;(XI,X2,X3,X4) = o. Definition 6.20. (Singular solution) A solution (Xl,X2,X3,X4) of equation (6.67) is called singular with respect to Xi if it satisfies also the equation Qx; (Xl, X2, X3, X4) = O. The set of solutions singular with respect to Xi zs parametrized by the curve (6.73) called the singular curve for Xi. The projection of the curve Si onto the coordinate plane (k, l) is exactly the biquadratic hkl = pqXj - PXjq = Qx;QXj - QQX.i,Xj = O.
6.11. Proof of the classification theorem
243
Lemma 6.21. (Singular solutions and biquadratics) If a solution (Xl, X2, X3, X4) of equation (6.67) is singular with respect to Xi, then h jk = h jl = hkl = 0 on this solution. Conversely, if hkl = 0 for some solution, then this solution is singular with respect to either Xi or Xj. Proof. Since hkl = QXiQXj - QQxi,Xj' the equation hkl of the equation Q = 0 is equivalent to QXiQXj = O.
= 0 on the solutions 0
In the following theorem we will deal with biquadratic polynomials corresponding to various multiaffine ones; we will denote the biquadratics by the same letters as their parent quad-equations, with the superscripts for the remaining variables, so that, e.g., AO,1 = 6X2,X12A is the result of eliminating X2,X12 from A(X,XI,XI2,X2). Theorem 6.22. (Tetrahedron property and biquadratics for 3D consistent systems) Consider a 3D consistent system (6.68) with all six functions A, ... ,C being of type Q. Then:
• The system (6.68) possesses the tetrahedron property: the value of Xl23 as a function of the initial data x, Xl, X2, X3 does not depend on X; see Figure 6.19 . • For any edge of the cube, the two biquadratic polynomials corresponding to this edge, coming from the two faces sharing this edge, coincide up to a constant factor. Proof. The values of Xl23 obtained from the equations A = 0, B = 0 and C = 0, respectively, result from elimination of X12, Xl3 and X23, which can be expressed by the equations 2 I
I
3
I
F(x, Xl, X2, X3, XI23)
= AX13,X23BC 2 3
I
I
AX23BCX13 - AX13Bx23C + ABx23CX13
= 0,
I
G(x, Xl, X2, X3, XI23)
= BX12,X13CA - BX13CAx12 - BX12CX13A + BCX13Ax12 = 0, 2 I
3
I
H(x, Xl, X2, X3,
I X123)
= CX12,X23AB - CX12ABx23 - CX23Ax12B + CAX12 B X23 = O. Here the numbers over the arguments of the polynomials F, G, H indicate their degrees in the corresponding variables. These degrees are in the projective sense, that is in agreement with the action of Mobius transformations, and can be read off the right-hand sides. Due to 3D consistency, the expressions for Xl23 as functions of X,XI,X2,X3 found from these three equations,
6. Consistency as Integrability
244
coincide. Therefore the polynomials F, G, H must factorize as: F
2
= J(x,x3)K,
G
2
= g(x,xdK, 1
1
1
H
2
= h(X,X2)K,
1
K = K(X,X1,X2,X3,X123), where the polynomial K yields the common value of X123 as a function of x, Xl, X2, X3. Here J, g, h are some polynomials of degree 2 in the second argument. The degrees of J, g, hand K in X remain to be determined. We do this by analyzing singular solutions. Let the initial data x, Xl, X2 be free variables, and let X3 be chosen to satisfy the equation J(x, X3) = O. Then F == 0, and thus the system B = C = A = 0 does not determine the value of Xl23. Moreover, the equation B = 0 can be solved with respect to X23 since otherwise the initial data must be constrained by the equation BO,2(x, X2) = O. Analogously, the equation C = 0 is solvable with respect to Xl3. Therefore, the uncertainty appears from the singularity of equation A = 0 with respect to Xl23. Hence, the relation A3,23(X3, X23) = 0 is valid. In view of the assumption of the theorem, X23 is a (two-valued) function of X3 and does not depend on X2. This means that the equation B = 0 is singular with respect to X2 and therefore BO,3(x, X3) = O. Analogously, CO,3(x, X3) = O. Thus, we have proven that if X3 = 'P(x) is a zero of the polynomial J then it is also a zero of the polynomials BO,3, CO,3. If one of these three polynomials is irreducible, then this already implies that they coincide up to a constant factor. If the polynomials are reducible, then we could have J = a 2, BO,3 = ab, CO,3 = ac, where a, b, care multiaffine in x, X3. In any case, deg x J = 2, and this is sufficient to complete the proof. Indeed, this implies deg x K = 0, so the tetrahedron property is valid, and the first statement of the theorem is proven. In turn, the tetrahedron property can be used to prove the relation (6.74)
The variables in this relation separate:
B O,3 CO,3
AO,l BO,2 - CO,l . AO,2 '
so that BO,3/CO,3 may only depend on x. Due to nondegeneracy of the biquadratics BO,3, CO,3, this ratio is constant, which proves the second statement of the theorem.
245
6.11. Proof of the classification theorem
So, it remains to prove equation (6.74). For this goal, rewrite the system
(6.68) in the form
= a(x,xI,X2), X23 = b(X,X2,X3), Xl3 = C(X,XI,X3), Xl23 = a(x3,xI3,X23) = b(XI,XI2,XI3) = C(X2,XI2,X23). Assuming the tetrahedron property, i.e., Xl23 = d(XI, X2, X3), we find by Xl2
differentiation: dXl dX2 dX3
= aX13 CXl , = bXl2 a X2 , = CX23 b X3 ,
dX2 =
aX23 b X2 ,
dX3 = dXl =
bXl3 CX3 '
aXl3 c X + aX23 b x , 0= bx12 a x + bXl3 C X ' 0= CX23 b x + cXl2 a X '
0=
cX12 a Xl ,
These equations readily imply the relation aX2bx3cXl
+ a Xl b X2 c X3
= O.
It is equivalent to (6.74) in view of the identity
a X2 /a Xl
X23,. _ _ _ _ _~8
=
AD,I/AD,2.
0
Xl23
I
_- - - r
X3
(q.."-'----.-----"CI3 : "
I "
)'.-
".l.'''' X 2).'9:", -,- -
x
Xl2
. _ - - - - - - 0 · Xl
Figure 6.19. Tetrahedron property.
The astonishing tetrahedron property, possessed by all 3D consistent systems of type Q, is illustrated in Figure 6,19. It means that the fields Xl, X2, X3, Xl23 sitting at the vertices of the white tetrahedron are connected by a certain multiaffine relation K(XI, X2, X3, X123) = O. Of course, for symmetry reasons, a multiaffine relation L(x, X12, X23, X13) = 0 also holds for the fields at the vertices of the black tetrahedron. 6.11.2. Analysis: descending from multiaffine Q to quartic r. In the further analysis, one more operation similar to (6.70) will be useful, namely
(6,75)
8Xk : P~ ~
pi,
8Xk (h) =
h;k -
2hh xkXk '
The operation 8Xk applied to a biquadratic polynomial h(XI, X2) actually computes its discriminant with respect to the variable Xk which gets eliminated, the result being a quartic polynomial which we will denote by rl (xz)
6. Consistency as Integrability
246
(where {k, l} = {1, 2}). Thus, from any h E P~, the operations 6Xk produce two quartic polynomials rl E The operation Xk is covariant with respect to Mobius transformations:
Pi.
6
(6.76)
The following statement is proved by a straightforward computation.
Lemma 6.23. (Commutativity of discriminants) For any multi affine polynomial Q(XI, X2, X3, X4) E
Pi,
6Xk (6 x;,xj(Q)) = 6xj (6 x;,Xk(Q)),
(6.77)
so that the following diagram is commutative:
r4(x4)
6x1 (6.78)
6X3
1
hI4(XI' X4)
r3(x3)
------>
16x1 ,x2 6X2 ,X3 f-----
1
6x2
6X1 ,X4
Q(XI, X2, X3, X4)
----+
h 23 (X2, X3)
1 6x3 ,x4
6x41 rl (xI)
6X4
h 34 (X3, X4)
+--
6X2
1 6x3 6X1
hI2(XI' X2)
+--
r2(x2)
------>
In fact, this diagram can be completed by the polynomials h 13 , h24 corresponding to the diagonals (so that the graph of the tetrahedron appears), but we will not need them. Further on we will make an extensive use of relative invariants of polynothese mials under Mobius transformations. For quartic polynomials r E relative invariants are well known and can be defined as the coefficients of the Weierstrass normal form r = 4 x3 - g2X - g3. For a given polynomial r(x) = r4x4 + r3x3 + r2x2 + rlX + ro they are given by
Pi
g2(r) = 418 (2rrIV - 2r'r'" + (r")2) =
1~ (12ror4 -
g3(r) = _1_(12rr"r IV - 9(r')2r IV - 6r(r"'f
=
3456 1 2 432 (72r or2r4 - 27rl r4
Under the Mobius change of x = the constant factors:
+ 9r lr2 r 3 -
Xl
3rlr3
+ r~),
+ 6r'r"r'" 2
2(r")3)
3
27ror3 - 2r2)·
these quantities are just multiplied by
247
6.11. Proof of the c1assiEcation theorem
For a biquadratic polynomial h E P~ , (6.79) h(x, y) = h22x2y2+h2IX2y+h2ox2+hI2xy2+hllxy+hlOX+ho2y2+hOly+hoo, the relative invariants are defined as
Notice that i3 can be defined also by the formula
Under the Mobius change of x = Xl and y = X2,
ik(M[h]) = ~~ ~~ik(h),
k = 2,3.
The following properties of the operations 6x ,y, 6x are proven straightforwardly. Lemma 6.24. (Opposite biquadratics and all four quartics have equal invariants) For any multiaffine polynomial Q(XI, X2, X3, X4) E Pl set: h I2 (XI,X2) =6X3 ,X4(Q) andh 34 (x3,x4) =6X1 ,X2(Q). For any biquadratic polynomial h(XI' X2) E P~ set: rl(xI) = 6X2 (h) and r2(x2) = 6X1 (h). Then (6.80)
ik(hI2) = idh34),
(6.81)
gk(r2) = gk(rl),
k = 2,3, k = 2,3.
In other words, in the diagram (6.78), the pairs of biquadratic polynomials on the opposite edges have the same invariants i2, i3, and all four quartic polynomials ri have the same invariants g2, g3. These results suggest the following approach to the classification of multiaffine equations Q = 0 modulo Mobius transformations. Suppose that, for a given Q E Pl, the four quartic polynomials ri(xi) associated to the vertices of the quadrilateral in the diagram (6.78) are known. Then one can use Mobius transformations to bring these polynomials into a canonical form. After that, one can reconstruct the edge biquadratics h ij from the pairs of vertex polynomials ri, rj. Finally, one can reconstruct the multiaffine Q from the edge biquadratics. 6.11.3. Synthesis: ascending from quartic r to biquadratic h. According to formulas (6.71), (6.76),
6X1 (6Xj ,Xk (M[Q])) = ~]~~~fM[6xl(6xj,xk(Q))l = C~i2Mh]'
6. Consistency as Integrability
248
where C = ~I ~~~~~~. Since the polynomial Q is defined up to an arbitrary constant factor, we may assume that Mobius changes of variables in the equation Q = induce transformations
°
ri
f-+
~i2 Mhl
of the polynomials r i. This allows us to bring each r i into one of the following six canonical forms: (6.82) r=(x 2 -1)(k 2x 2 -1),
r=x2-1,
r=x2,
r=x,
r=1,
r=O,
according to the six possibilities for the root distribution of r: four simple roots, two simple roots and one double, two pairs of double roots, one simple root and one triple, one quadruple root, or, finally, r vanishes identically. Note that in the first canonical form it is always assumed that k 2 1= 0,1, so that the second and third forms are not considered as particular cases of the first form. Not every pair of such polynomials is admissible as a pair of polynomials at two adjacent vertices, since the relative invariants of the polynomials of such a pair must coincide according to (6.81). We identify all admissible pairs, and then solve the problem of reconstruction of the biquadratic polynomial (6.79) by the pair of its discriminants (6.83)
which is equivalent to a system of 10 (bilinear) equations for 9 unknown coefficients of the polynomial h.
Lemma 6.25. (Reconstructing biquadratic from two discriminants) Nondegenerate biquadratic polynomials with a given pair of discriminants (rl(x),r2(y)) in the canonical form (6.82) exist if and only ifrl(x) = r(x) and r2 (y) = r(y) with one and the same canonical form r. These polynomials h can be brought into the following normal forms, possibly after Mobius transformations of x, y not affecting r:
(q4)
r(x)
=
(x 2 -1)(k 2x 2 -1):
h
=
1
2a(x2+y2-2Axy-a2-k2a2x2y2),
where A2 = r(a);
(q3)
r(x) = 8 - x 2 : where 8
h =
1
. ( ) (x 2 + y2 - 2 cos (a) xy) -
2 sm a
= 0, 1;
(q2)
r (x) = x :
(q1)
r(x)
=8:
1
h = 4a (x - y) 2 h
=
a
-
a3
2" (x + y) + 4 ;
(x - y)2 8a 2a - "2'
where
8 = 0,1.
8 sin(a) 2
,
6.11. Proof of the classification theorem
249
In the cases (q4), (q3)8=1 and (q2) any biquadratic h with a given pair of discriminants (r(x), r(y)) is automatically of the form given in the lemma; in the cases (q3)8=O and (q1) an additional Mobius transformation might be necessary to bring h to this form (for instance, in the case (q1)8=O, that is, r(x) = r(y) = 0, any biquadratic h = (KXY + AX + flY + v)2 has this pair of discriminants, and any Mobius transformation of x, y preserves this form of h). One clearly sees the origin of the elliptic curve in the case (q4): the solution of the problem of finding a biquadratic h(x, y) with the pair of discriminants (r(x), r(y)) in the case r(x) = (x 2 -1)(k 2x 2 -1) is parametrized by a point (a, A) of the corresponding elliptic curve. Introducing the uniformizing variable 0: by a = sn( 0:), so that A = sn' (0:) = cn( 0: )dn( 0:), we can write the corresponding biquadratic (q4) in the form (6.84) h(x, y; 0:) = 2 s~( 0:) (x 2 +y2 - 2 cn( o:)dn( o:)xy - sn2 ( 0:)(1 + k 2x 2y2)).
One can recognize this polynomial as the addition theorem for the elliptic function sn(x; k); more precisely, h(x, y; 0:) = 0 if and only if X = sn(~; k) and y = sn(7]; k) with ~ - 7] = ±o:. 6.11.4. Synthesis: ascending from biquadratics h ij to multiaffine Q. The next step is the reconstruction of the multiaffine polynomials from the biquadratic ones. In doing this, the following facts are useful (they are proven by a direct computation). Lemma 6.26. (Reconstructing multiaffine equation from edge biquadratics) For any multiaffine polynomial Q E Pl, with the notation h ij = bXk,Xl (Q) E P~, the following identities hold:
4i3(h12)h14
(6.85)
where
(6.86)
f
=
= det
23 h 34 h 23 h 34 - h X3 X3X3 X3
12 h;;
12 h Xl 12 h X1X2
12 h X2X2
12 X 2 h X1X2
C
f fX2
),
f X2X2
34 . + h 23 h X3X3 '
12 h 34 _ h 14 h 23 + h 23 h 34 _ h 23 h 34 h'Xl Xl X3 X3 h 12 h 34 - h 14 h 23
Identity (6.85) shows that h14 can be expressed through the other three biquadratic polynomials (provided i3(h12) i= 0). Differentiating (6.86) with respect to X2 or X4 leads to a relation of the form Q2 = F[h 12 , h 23 , h 34 , h14], where F is a rational expression in terms of hij and their derivatives. Therefore, if the biquadratic polynomials on three edges (out of four) are known, then Q can be found explicitly. Of course, it is seen from Lemma 6.26
6. Consistency as Integrability
250
that not every set of three biquadratic polynomials comes as h ij from some
QE
Pl.
Proof of Theorem 6.18. We demonstrate the reconstruction procedure in the most interesting case (Q4). Let the polynomials h 12 , h2:3, h 34 and h14 be of the form (q4), with parameters denoted by (a, A), (b, B), (ii, A) and (b,.8), respectively, all of them lying on the elliptic curve A 2 = r( a). The relative invariants i2, i3 of hP and h 34 must coincide because of (6.80), and it is easy to check that this condition allows only the following possible values for (ii, A):
(a, A),
(-a, -A),
1 ka 2 (a, -A),
1
ka 2 (-a, A).
According to (6.71), a Mobius change of variables in the equation Q = 0 yields JX/r,xll\;I[Q])
=
~k~IM[JXhl;I(Q)]
=
C~il~jlM[hij],
where C = ~1~2~3~4. Since Q is only defined up to a multiplicative constant, we may assume that a Mobius change of variables induces transformations h ij ~ ~il ~jl M[h ij ] of the biquadratic polynomials h ij . In particular, if
h 34 = h(X3' X4; -a, -A)
or
11 31 = h( X3, Xl; kIa' -
then the corresponding Mobius transformation, X3 will change h 34 to
-h( -X3, X4; -a, -A),
resp.
~
k~2)'
-X3 or X3
~
1/(kx3),
2 ( 1 1 A )' - kX3h kX3' X4; ka' - ka 2
both of which coincide with h(X3, X4; a, A) due to the symmetries of the polynomial (q4). Thus, performing a suitable Mobius transformation of the variable X3 (which does not affect the polynomial r(x3)), we may assume without loss of generality that (ii, A) = (a, A). After that, the polynomial h14 is uniquely found according to formula (6.85), and it turns out that the equality (b,.8) = (b, B) is fulfilled automatically. Thus, the change of one variable allows us to achieve the equality of the parameters corresponding to the opposite edges of the square. A direct computation using formula (6.86) yields the equation
a(XIX2 + X3X4) + b(XIX4 + X2X3) - C(XIX3 + X2X4) - abc(l + k 2 XIX2.r3:r'1) = 0, where c = (aB +bA)/(l- k 2a 2b2). The uniformizing substitution a = sn(a), b = - sn(;3) , so that A = sn'(a), B = sn'(;3), and therefore c = sn(a - (3), brings it to the form (Q4).
6.11. Proof of the classification theorem
251
Also in the other cases (Q1)-(Q3), suitable Mobius changes of the variables X2, X3, X4 allow us to bring the polynomials into the form hl2 = h(XI,X2ia), h 23 = h(X2,x3i(3), h 34 = h(x3,x4ia). A direct computation with formula (6.85) proves that this yields hl4 = h(XI' X4i (3). Then the multiaffine Q is found by using (6.86). 0 6.11.5. Putting equations Q = 0 on the cube. Proof of Theorem 6.19. Given a 3D consistent system (6.68) with all equations of type Q, one can use the Mobius transformations from (Mob)8 to bring all six equations into the canonical form from the list Q. Since by Theorem 6.22 biquadratics coming to an edge from two adjacent faces must coincide up to a constant factor, all six equations have to be of one and the same type (Q1)-(Q4). Moreover, the parameters k 2 in the case (Q4) and § in the cases (Q3), (Q1) have to be the same on each face of the cube. Therefore, the equations on all faces may differ only by the values of a and (3. Consider the equations corresponding to three faces adjacent to one vertex, say to x:
A(X,Xl,X12,X2) = Q(X,XI,XI2,X2ia,P) = 0, B(x, X2, X23, X3) = Q(x, X2, X23, X3i (3,;Y) = 0, C(x, X3, X13, xt) = Q(x, X3, x13, Xli 'Y, a) = O. We will show that one can write these three equations as (6.87) For the polynomials (Q1)-(Q4) from the list Q we have:
(6.88) (6.89)
h I2 (Xl, X2) h 14 (XI,X4)
= §x3,x4Q(Xl, X2, X3, X4i a, (3) = ",(a, (3)h(XI, X2i a), =
§x2,x3Q(XI,X2,X3,x4ia,(3) = ",((3,a)h(xI,x4i(3),
with the biquadratics h(x, Yi a) listed in Theorem 6.25 as (q1)-(q4). Thus, we find:
AO,l(X, Xl) = ",(a, P)h(x, Xli a),
02
-
-
BO,2(X, X2) = ",((3, ;Y)h(x, x2i (3),
A ' (x, X2) = ",((3, a)h(x, x2i (3), BO,3(x, X3) = ",(;y, (3)h(x, x3; ;Y),
CO,3(x, X3) = ",(,,(, a)h(x, x2i 'Y),
CO,I(x, Xl) = ",(a, 'Y)h(x, Xl; a).
According to the second statement of Theorem 6.22 and to formula (6.74), the following relations must hold: (6.90) h(X,XI;a) _ ( _) h( X,XI;a ) - rn a, a , (6.91)
6. Consistency as Integrability
252
For the most complicated case (Q4), to which we will restrict ourselves in this proof, the biquadratic (q4) is given in (6.84), and a direct computation gives K,(a,{3) = 2sn(a) sn(f3) sn(a - {3). Equations (6.90) yield that & may only take the values ±a, which correspond to m(a, &) = ±1, and analogously for {3,'Y. Equation (6.91) with the above-mentioned values of K,(a,{3) yields m(a, &)m({3, (J)m("(, 1') = 1. Thus, up to a change of enumeration, two cases are possible: or
&
= a,
{3 = -{3,
'Y
= -"(.
In the first case the equations A = 0, B = 0, C = 0 have the desired form (6.87) with al = a, a2 = {3, a3 = 'Y. In the second case it is enough to observe that the equation B = 0 is not affected by the replacement of parameters ({3,1') with (-{3, -i') = (-{3, 'Y), which again leads to the desired form (6.87) with al = a, a2 = -{3, a3 = 'Y. Continuing to argue in a similar manner for faces adjacent to other vertices, one shows that the signs of edge parameters can always be adjusted on the whole cube as in system 0 (6.72).
6.12. Classification of discrete integrable 2D systems with fields on vertices. II In the previous two sections, we classified quad-equations Q = 0 of type Q, that is, those with all nondegenerate edge biquadratics, and showed their 3D consistency. However, quad-equations of type H, i.e., those with (some of) the edge biquadratics being degenerate, are by no means less interesting or less important. It is enough to mention that the very prominent Hirota equation is of type H (which is the reason for the choice of the latter notation). A classification of multiaffine equations of type H seems to be a rather complicated and tiresome task. Nevertheless, postulating some additional properties, a classification can be achieved. Our assumptions for the quad-equation Q = 0 will be as follows: ~ Linearity. The left-hand side of the equation (6.92) is a polynomial of degree 1 in each variable, depending on two parameters assigned to the edges. ~
Symmetry. The function Q has the symmetry properties
(6.93)
EQ(XI,X4,X3,x2;{3,a),
E
(6.94)
aQ(x2,x3,x4,xI;{3,a),
a = ±1.
~
= ±1
Tetrahedron property. The value X123, existing due to 3D consistency, depends on Xl, X2 and X3, but not on x.
6.12. Classification of discrete integrable 2D systems. II
253
The symmetry properties are natural to require, because to enable us to pose our equations on arbitrary quad-graphs, the equations should not depend on the enumeration of vertices. Note that the normal forms of the list Q possess these symmetries. The tetrahedron property is admittedly a less natural classification assumption, but it holds for the vast majority of known interesting examples, including all the equations of the list Q and the Hirota equation itself; see formula (6.32). We consider here the problem of 3D consistency for equation (6.92) in the sense of the system (6.72), with one and the same polynomial Q. Due to the symmetry assumption, the natural transformation group, which can be used to put the equation in the normal form, is essentially smaller than in Section 6.10; namely, all vertex fields should be acted on by one and the same Mobius transformation. Theorem 6.27. (Classification of symmetric equations with tetrahedron property) Any 3D consistent quad-graph equation (6.92) possessing the linearity, symmetry, and tetrahedron properties is equivalent, modulo Mobius transformations acting simultaneously on all variables Xi and modulo point transformations of the parameters a, (3, to one of the equations of the following lists. List Q from Theorem 6.18: (Q4)
sn( a) (XIX2 +X3X4) - sn({3)(xIX4 +X2X3) - sn( a - (3)(XIX3 + X2X4) + sn( a - (3) sn( a) sn({3) (1 + k2xIX2X3X4) = 0,
(Q3)
sin( a) (XIX2+X3X4) -sin({3) (XIX4 +X2X3) -sin( a- (3)(XIX3+X2X4) +0 sin(a - (3) sin(a) sin({3) = 0,
(Q2)
a(xlx2 + X3X4) - (3(XIX4 + X2X3) - (a - (3)(XIX3 + X2X4) +a{3(a-{3)(xI +X2+X3+x4)-a{3(a-{3)(a2-a{3+{32) = 0,
(Q1)
a(xlx2 + X3X4) - (3(XIX4 +oa{3( a - (3) = 0;
+ X2X3)
- (a - (3)(XIX3
+ X2X4)
list H:
+ X3X4)
+ X2X3) + 0(a 2 - (32) = 0, X4) + ({3 - a)(XI + X2 + X3 + X4) + {32 -
(H3)
a(xlx2
(H2)
(Xl - X3)(X2 -
(H1)
(Xl - X3)(X2 - X4)
- (3(XIX4
+ (3 -
a
a 2 = 0,
= 0;
and list A: (A2)
sin(a)(xlx4 + X2X3) - sin({3) (XIX2 + X3X4) - sin(a - (3)(1 + XIX2X3X4) = 0,
(A1)
a(xlx2 + X3X4) - (3(XIX4 -oa{3(a - (3) = 0.
+ X2X3) + (a -
(3)(XIX3
+ X2X4)
Remarks. 1) The parameter 0 in equations (Q3), (Q1), (H3), (A1) can be scaled away, so one can assume without loss of generality that 0 = or 0 = 1.
°
6. Consistency as Integrability
254
2) If one extends the transformation group of equations by allowing Mobius transformations to act on the variables Xl, X3 differently than on x2, X4 (white and black subgraphs of a bipartite quad-graph), then equation (A2) turns into (Q3)8=O by the change (X2, X4) I---> (1/X2, 1/X4), and equation (A1) turns into (Q1) by the change (X2, X4) I---> (-X2, -X4). SO, really independent equations are given by the lists Q and H. 3) Equation (H3) is the most general in the list H, since (H1) and (H2) can be considered as its limiting cases. Note that (H1) is the discrete KdV equation and (H3)8=O is a version of the Hirota equation with the symmetry properties (6.93), (6.94). The general scheme of the proof of Theorem 6.27 is the same as in Section 6.11. We start with the "analysis" part. Due to the symmetry assumption, all edge biquadratics for the polynomial Q(Xl, X2, X3, X4; a, (3) are given by one and the same biquadratic polynomial g(x, y; a, (3), so that
h 12 (Xl, X2) = c5 X3 ,X4 (Q) = g(Xl, X2; a, (3), h 14 (Xl,X4) = c5 X2 ,X3(Q) = g(Xl,x4;(3,a). Moreover, the polynomial 9 is symmetric: g(x, y; a, (3) = g(y, x; a, (3). Lemma 6.28. (Descending from multiaffine Q to quartic r) The biquadratic g(x, y; a, (3) admits a representation (6.95)
g(X, y; a, (3) = k(a, (3)h(x, y; a),
where the factor k is antisymmetric, k((3, a) = -k( a, (3), and the coefficients of the polynomial h(x, y; a) depend on a single parameter a in such a way that the discriminant r(x) = c5 y (h) does not depend on a at all. Proof. In the proof of Theorem 6.22 we used the previously demonstrated tetrahedron property to derive formula (6.74). In the present setup the tetrahedron property has been postulated, thus we can still use formula (6.74), which, due to the symmetry assumptions, takes the following form:
g(x, Xl; aI, a2)g(x, X2; a2, a3)g(x, x3; a3, ad
= -g(x, Xl; aI, a3)g(x, X2; a2, adg(x, X3; a3, a2). This relation implies that the fraction g(x,xl;al,a2)/g(x,xl;al,a3) does not depend on Xl, and due to the symmetry it does not depend on X either. We see that the symmetry assumptions has been used in this argument to replace the nondegeneracy of biquadratics which has been required in Theorem 6.22 to come to the same conclusion. We find:
g(x, Xl; aI, a2) g(x, xl; aI, a3)
K:(al, a2) K:(al,a3)'
6.12. Classification of discrete integrable 2D systems. II
255
where the function /(x, y; a, (3) = lJI(x, y; a - (3), (6.105) (Ql)8=O: Additive three-leg form with ¢(x, y; a, (3) = 'lj;(x, y; a - (3), 'lI'{Xo, Xl; a)
(6.106)
a
= --Xo - Xl
(H3): Multiplicative three-leg form with (6.107)
iJ>(x, y; a, (3) =
(3x - ay ax-
(3' y
lJI(XO, Xl; a) = XOXI
+ ba.
(H2): Multiplicative three-leg form with (6.108)
iJ>(x,y;a,(3)
=
x-y+a-(3 x-y-a+
(3'
lJI(xo,xI;a)=xo+xI+a.
(HI): Additive three-leg form with (6.109)
¢(x, y; a, (3)
a-(3
= -- , x-y
'l/}(xo, Xl; a)
=
Xo
+ Xl.
Proof. The proof of this theorem is obtained by a direct computation; see Exercise 6.15. However, this does not give any insight in how these three-leg forms could be found. A general way to derive three-leg forms is the subject 0 of Exercise 6.16. Remark. It should be mentioned that the existence of a three-leg form allows us to derive (and, in some sense, to explain) the tetrahedron property of Section 6.11. Indeed, consider three faces adjacent to the vertex Xl23 in Figure 6.19, namely the quadrilaterals (Xl, X12, X123, X1.'3) , (X2' X23, X123, XI2), and (X3, X13, X123, X23). A summation (resp. multiplication) of the three-leg forms (centered at X123) of equations corresponding to these three faces leads to the additive equation (6.110)
¢(XI23, xl; a2, a3)
+ ¢(X123, X2; a3, al) + ¢(XI23, X3; aI, a2) =
0,
respectively to the multiplicative equation (6.111)
This is the equation which relates the fields at the vertices of the "white" tetrahedron in Figure 6.19. Note that it can be interpreted as a discrete Laplace type equation coming from a spatial flower with three petals. The functions ¢(x,y;a,(3) (resp. iJ>(x,y;a,(3)) corresponding to the "long" legs, yield additive (resp. multiplicative) Laplace type equations on arbitrary planar graphs. Studying the list of Theorem 6.32, one sees that there are only six "long" legs functions. Three of them are rational in x, y;
6. Consistency as Integrability
260
each of the corresponding Laplace type systems admits two different extensions to a quad-graph system: one from the list Q, where the form of the "short" legs coincides with the form of the "long" ones, (Q3)8=O, (Q1)8=1, and (Q1)8=O, and the other from the list H, with different "short" legs, (H3), (H2), and (H1), respectively. The other three functions are rational in Y only, and require a uniformizing change of the variable x. The corresponding Laplace type systems admit only one extension to a quad-graph system, (Q4), (Q3)8=1, and (Q2). Theorem 6.33. (Integrability of Laplace type equations) Additive Laplace type equations (6.98) with the "long legs" functions ¢(x, Yi Cl:, (3), resp. multiplicative Laplace type equations (6.100) with the "long legs" functions (x, Yi Cl:, (3) from Theorem 6.32, are integrable in the sense of Section 6.4. Proof. Observe that the functions ¢, resp. , always contain the parameters Cl:, (3 in the combination Cl: - (3. This means that the edge parameters of the extension of the Laplace type system to a quad-graph system are only defined up to an additive constant A. Let the equations of the extended quad-graph system read
Q(xo, Yb Xk, Yk+li Cl:k
+ A, Cl:k+l + A)
=
o.
Rewrite this equation as a Mobius transformation, (6.112) Then the (normalized, if necessary) matrix L(xo, Xb Cl:k, Cl:k+li A) is the transition matrix across the edge (xo, Xk) E E(9), or along the edge (Yk, Yk+d E E(9*), in the zero curvature representation of the Laplace type system on 9. 0 Example. Consider the additive Laplace type system on 9, corresponding to the equation (Q1)8=O (cross-ratio equation): Cl:k - Cl:k+l =
'"""'
(6.113)
L....; xkE
star(xo)
O.
Xo - Xk
The extension to the quad-equation reads:
(xo - Yk)(Xk - Yk+l) Cl:k + A (Yk - Xk)(Yk+l - xo) Cl:k+l + A' which can be written as the Mobius transformation (6.112) with the matrix L(XO,Xk,Cl:k,Cl:k+li A) (6.114)
( \ + n,k)l + Cl:k A
LX.
Cl:k+l
Xo - Xk
(Xk 1
-XOXk) -Xo
261
6.14. Fields on edges: Yang-Baxter maps
These matrices give a zero curvature representation of the Laplace type system (6.113) on an arbitrary surface graph 9.
6.14. Fields on edges: Yang-Baxter maps We now turn to the study of another large class of 2D systems on quadgraphs with fields assigned to the edges. In this situation it is natural to assume that each elementary quadrilateral carries a map F : X x X -> X x X, with X being the set where the fields x, Y take values, so that F(x, y) = (X2' yd; see Figure 6.22. The concept of 3D consistency of such maps may be interpreted in several ways, depending on the initial value problem one would like to pose on the elementary 3D cube.
x
Figure 6.22. Map encoded by an elementary quadrilateral with fields on edges.
One way is to choose the initial data x, y, z on three edges of an elementary cube adjacent to one vertex. One computes first
F(x, y) = (X2' Yl),
F(y, z) = (Y3, Z2),
F(z, x) = (Zl' X3),
and then
F(X3, Y3) = (X23, Y13),
F(Yl, Zl) = (Y13, Z12),
F(Z2' X2) = (Z12' X23),
so that there are two a priori different answers for any of the fields X23, Y13, Z12 with two indices; see Figure 6.23. Definition 6.34. (3D consistent map) A map F : X x X -> X x X is called 3D consistent if the two answers for each of the fields (X23, Y13, Z12) in Figure 6.23 coincide for any initial data (x, y, z).
An important example of a 3D consistent map, coming from discrete differential geometry, is discussed in Exercises 6.19, 6.20. Definition 6.34 has one notational inconvenience: since each initial edge is used on the first step by two different maps, it is not possible to express the property of 3D consistency in terms of compositions of maps. This can be overcome by a different choice of initial data, namely by choosing them on a path consisting of three edges of three different coordinate directions.
6. Consistency as Integrability
262
Y3
Y13 I
X3
Z2 1 I I
Z
I X }- ____2_
Y/ /
/
X
Figure 6.23. 3D consistency of 2D systems with fields on edges.
This leads to the notion of Yang-Baxter maps (traditionally denoted by R rather than by F). Definition 6.35. (Yang-Baxter map) A map R : X x X ---+ X x X is called a Yang-Baxter map if it satisfies the Yang-Baxter relation (6.115) where each ~j X 3 1----+ X 3 acts as the map R on the factors i, j of the Cartesian product X 3 , and acts identically on the third factor.
Equation (6.115) is understood as follows. The fields x, yare supposed to be assigned to the edges parallel to the I-st and the 2-nd coordinate axes, respectively. Additionally, consider the fields z assigned to the edges parallel to the 3-rd coordinate axis. Initial data are the fields x, y, z on a path consisting of three edges of different coordinate directions; see Figure 6.24. The left-hand side of this figure corresponds to the composition of maps on the left-hand side of equation (6.115), which are visualized as maps along the three front faces of the cube:
(Here and below we slightly abuse the notation by omitting the arguments on which our maps act identically.) Similarly, the right-hand side of the figure corresponds to the chain of maps on the right-hand side of (6.115), which are visualized as maps along the three back faces of the cube:
So, equation (6.115) assures that the two ways of obtaining from the initial data (x, y, z) lead to the same results.
(X23, Y13, Z12)
6.14. Fields on edges: Yang-Baxter maps
263
X23
;1:2:3
Yl3
Y13
Rl~
Y3
X3
~
Z2
Z12 R13
I
\'i'l I
Z Z12
Z
YI~_ X2 _ _
//Rl~
Y
R13
X
Y
X Figure 6.24. Yang-Baxter relation.
The notion of the zero curvature representation makes perfect sense for Yang- Baxter maps, and can be expressed as
L(x, >")L(y, >..) = L(Yl, >")L(X2' >..).
(6.116)
There is a construction of zero curvature representations for Yang-Baxter maps with no more input information than the maps themselves, close in spirit to Theorem 6.4. Consider a parameter-dependent Yang-Baxter map R(o, (3), with parameters CI', (3 E C assigned to the same edges of the quadrilateral in Figure 6.22 as the fields x, y, opposite edges carrying the same parameters. Although this can be considered as a particular case of the general notion, by introducing :X = X x C and R(:E, 0; y, (3) = R(o, (3)(x, y), it is convenient for us to keep the parameter separately. Thus, in Figure 6.24 all edges parallel to the :1: (resp. y, z) axis carry the parameter 0: (resp. (3, 1'), and the corresponding version of the Yang-Baxter relation reads: (6.117) Theorem 6.36. (Zero curvature representation for Yang-Baxter maps) Suppose that there is an effective action of thc linear group G = G L( N, q on the set X (i. e.. A E G acts identically on X only if A = I), and that the Yang-Baxter map R(o:, (3) has the following special form:
(6.118)
X2
=
B(y, (3, o:)[x] ,
Yl
=
A(x, o,fJ)[y].
Here A, B : X x ex C ---t G are some matrix-valued functions on X depending on parameters ° and (3, and A[x] denotes the action of A EGan x E X. Then, whenever (X2' yd = R(o, (3)(.T, y), we have (6.119)
A(x,o:,>")A(y,(3,>")
A(Yl, (3, >")A(X2' 0:, >..),
(6.120)
B(y,(3,>")B(x, 0:, >..)
B(X2' 0, >")B(YI' (3, >..).
6. Consistency as Integrability
264
In other words, both A(x, a,).) and B-1(x, a,).) (or BT(x, a, ).)) form zero curvature representations for R. Proof. Look at the values of Z12 produced by the two sides of the YangBaxter relation (6.117): the left-hand side gives Z12 = A(x, a, ,)A(y,,8, ,)[z], while the right-hand side gives Z12 = A(Yl,,8,,)A(x2,a,,)[z]. Now since we assume that the action of G is effective, we immediately arrive at the relation
A(x,a,,)A(y,,8,,) = A(Yl,,8,,)A(x2,a,,), which holds whenever (X2, yd = R(a, ,8)(x, y). This coincides with (6.119), an arbitrary parameter, playing the role of the spectral parameter ).. Similarly, one could look at the values of X23 produced by the two sides of (6.117): the left-hand side gives X23 = B(Y3,,8, a)B(z2", a)[x], while the right-hand side gives X23 = B(z", a)B(y,,8, a)[x]. Effectiveness of the action of G again implies:
B(Y3,,8, a)B(z2", a) = B(z", a)B(y,,8, a), whenever (Y3, Z2) = R(,8, ,)(y, z). This coincides with (6.120); here the role of the spectral parameter). is played by an arbitrary parameter a. 0 In order to cover all known examples, the scheme of Theorem 6.36 must be extended in the following way. We say that A(x, a,).) gives a projective zero curvature representation for the Yang-Baxter map R if the relation (6.116) holds up to multiplication by a scalar matrix cI, where c may depend on all the variables in the relation. Assume that the action of G = GL(N, q on X is projective, i.e., scalar matrices and only they act trivially. Then the previous considerations show that the matrices A(x, a,).) and B-1(x, a,).) give projective zero curvature representations for the corresponding YangBaxter maps (6.118). In practice, the natural choices of matrices A, B in (6.118) actually lead to proper zero curvature representations, as the following examples show. Example 1: Adler map. Here X =
_ a-,8 x = y - --, x+y
(6.121)
C]!Dl
and the map has the form
_
,8-a x+y
y = x - --.
Then fj
= x _,8 - a = x 2 + xy - (,8 - a) = A(x,a,,8)[y], x+y
x+y
where
A(x,a,).)=
(
X
1
x2
+ xa
-). )
'
and the group G = GL(2, q acts projectively on C]!Dl by Mobius transformations. In this example B(x, a,).) = A(x, a, ).), so the matrices BT = AT provide us with an alternative zero curvature representation.
6.14. Fields on edges: Yang-Baxter maps
265
Example 2: Interaction of matrix solitons. Our next example comes from mathematical physics. The matrix Korteweg-de Vries equation Ut + 3UUx + 3Ux U + Uxxx = 0 admits one-soliton solutions of the form U(x, t) = 2a 2P sech2 (ax - 4a 3 t), where a is the parameter measuring the soliton velocity, and the matrix amplitude P must be a projector: p2 = P. Projectors of rank 1 have the form P = ~r? / (~, 17). It turns out that the change of the matrix amplitudes P of two solitons with velocities a1 and a2 after their interaction is described by the following Yang-Baxter map:
R(a1' a2) : (6,171; 6, 172)
->
(~l' ih; ~2' T72),
(6.122) (6.123) In this example X is the set of projectors P of rank 1 which is the variety CpN-1 x CpN-1, and a projective action of the group G = GL(N, q on X is induced by A[(~, 17)] = (A~, A'I]). It is easy to see that formulas (6.123) can be written as
with the matrices ~17T
2a A-a
A(~,17,a,'\)
= 1+ - \ - . -(-)' ~,17
Thus, the matrices A(~, 17, a,'\) give a projective zero curvature representation for the interaction map, but it is not difficult to see that this is actually a genuine zero curvature representation. As in Example 1, B(~, 17, a,'\) = A(~, 17, a, '\).
Example 3: Yang-Baxter maps ansmg from geometric crystals. Let X = C n , and define R : X x X -> X x X by the formulas (6.124)
xJ --
X
p
_J_ J
-
p'
Yj
j-1
Pj-1 = Yj -po '
j = 1, ...
,n,
J
where (6.125)
Pj =
t (IT a=l
k=}
Xj+k
IT
YJ+k)
k=a+}
(in this formula subscripts j + k are taken (mod n) ). Clearly, the map (6.124) keeps the subsets Xn x Xf3 C X x X, where n
Xn =
{(X}, ...
,xn ) EX:
II k=}
Xk
= a},
6. Consistency as Integrability
266
invariant. It can be shown that the restriction of R to Xn x Xt3 may be written in the form (6.118). For this, the following trick is used. Embed Xn x Xt3 into ClP'n-l x ClP'n-l via J(:c,y) = (z(x),w(y)), where
z(x) = (1 : ZI
: ... :
zn-d,
w(y) =
j
Zj
=
II
(WI: ... : W n-l :
1),
n
Xk ,
Wj
=
k=l
II
Yk·
k=j+1
Then it is easy to see that in the coordinates (z, 111) the map R is written as
z = B(y, (3, ex)[z] ,
'Ii! = A(x, ex, (3)[w] ,
with certain matrices B, A from G = GL(n, q, where the standard projective action of GL(n, q on ClP'n-1 is used. Moreover, a simple calculation shows that the inverse matrices are cyclic two-diagonal:
(6.126)
(6.127)
B- 1 (y, (3, ex)
A-I(x, ex, (3)
Yl
-1
0 0
Y2
0 -1
0
Y3
0 -ex
0 0
0 0
Xl
0
-1 0
X2
0 0
-1
X3
0 0
0 0
0 0
0 0 0
0 0 0
Yn-l
-1
0
Yn
0 0 0
-(3 0 0
.Tn -1
0
-1
Xn
To be more precise, the matrices A. B are defined only up to multiplication by scalar matrices. These scalar matrices are chosen in (6.126), (6.127) in such a way that the dependence of the matrices B- 1 , A-Ion their "own" parameters ((3, resp. ex) drops out, so that the only parameter remaining in the zero curvature representation is the spectral one. In other words, the zero curvature representation does not depend on the subset Xn x Xt3 to which we restricted the map. It can be checked that this is actually a genuine (not only projective) zero curvature representation. Note also that in this example the matrices BT coincide with A (so they cannot be used to produce an alternative zero curvature representation for R).
6.15. Classification of Yang-Baxter maps Consider Yang-Baxter maps R : X x X ----+ X x X, (.T,y) f---+ ('U,v) in the following special framework. Suppose that X is an irreducible algebraic variety, and R is a birational automorphism of X x X. Thus, the birational
6.15. Classification of Yang-Baxter maps
267
map R- I : X x X --> X x X, (u,v) ~ (x,y) is defined. This is depicted in the left square in Figure 6.25. Furthermore, a nondegeneracy condition is imposed on R: rational maps u(',y) : X - t X and v(x,') : X --> X must be well defined for generic x, resp. y. In other words, birational maps If : X x X --> X x X, (x, v) ~ (u, y) and If-I: X x X - t X X X, (u, y) ~ (x, v), called companion maps to R, must be defined. This requirement is visualized in the right square in Figure 6.25. Birational maps R satisfying this condition are called quadrirational. A formal definition of a slightly more general notion (where different spaces are allowed for the arguments x and y) looks as follows.
Definition 6.37. (Quadrirational map) Let Xl, X2 be two irreducible algebraic varieties over C. A rational map F : Xl X X2 --> Xl X X 2 , identified with its graph, an algebraic variety r F C Xl X X2 X Xl X X 2 , is called quadrirational if for any fixed pair (x, y) E Xl X X 2 , except possibly some closed subvarieties of codimension ~ 1, the variety r F intersects each of the sets {x} x {y} X Xl X X 2 , Xl X X2 X {x} x {y}, Xl x {y} X {x} X X2 , and {x} X X2 X XI X {y} exactly once, i. e., if r F is a graph of four rational maps F,F-I,P,P-I: Xl X X2 --> Xl X X 2 .
u
u
Y;(,V x
x
Figure 6.25. A map F on X x X, its inverse and its companions.
It is possible to classify all quadrirational maps in the case Xl = X2 = ClP'I; we give a short presentation of the corresponding results. Birational isomorphisms of ClP'1 x ClP'l are necessarily of the form
(6.128)
F. u = a(y)x + b(y) . c(y)x + d(y) ,
A(x)y + B(x) v - ----'-'-----'- C(x)y + D(x) ,
where a(y), .. . , d(y) are polynomials in y, and A(x), .. . , D(x) are polynomials in x. For quadrirational maps, the degrees of all these polynomials are :S 2. Depending on the highest degree of the coefficients of each fraction in (6.128), we say that the map is [1:1], [1:2], [2:1], or [2:2]. The richest and most interesting subclass is [2:2]. For the maps of this subclass the polynomials .6.(x) = A(x)D(x) - B(x)C(x) and 8(y) = a(y)d(y) - b(y)c(y) are of
6. Consistency as Integrability
268
degree four. A quartic polynomial belongs to one of the following five types, depending on the distribution of its roots: I: four simple roots,
II: two simple and one double root, III: two double roots, IV: one simple and one triple root, V: one quadruple root. It turns out that a necessary condition for a map of the subclass [2:2] to be quadrirational is that the polynomials ~(x) and 8(y) are simultaneously of one of the types I-V. Sufficient conditions are more complicated and can be expressed in terms of singularities of the map F, i.e., those points (~, 'T1) E «][»1 x CIP'1 where both the numerator and the denominator of at least one of the fractions in (6.128) vanish:
(6.129) or A(~)'T1
(6.130)
+ B(O = 0,
C(O'T1 + D(~) =
o.
For instance, if both polynomials ~(x) and 8(y) are of type I, then the necessary and sufficient condition for the quadrirationality of the map (6.128) is that the roots Xi, Yi (i = 1, ... ,4) of ~(x), 8(y) can be ordered so that both equations (6.129), (6.130) be satisfied for (~, 'T1) = (Xi, Yi), i = 1, ... ,4; in other words, the four singularities of both fractions in (6.128) be at the points (Xi, Yi). One can find normal forms for all quadrirational [2:2] maps with respect to the action of the natural transformation group, which in this case is the group (Mob)4 of Mobius transformations acting independently on each of the fields X, y, u, v.
Theorem 6.38. (Classification of quadrirational maps on CIP'1 x C1P'1) Any quadrirational (2:2) map on CIP'1 xCIP'1 is equivalent, under some change of variables acting by Mobius transformations on each field X, y, u, v independently, to exactly one of the following five normal forms: Fr:
u = oyP,
v = (3xP,
Frr:
y u=-P, 0
v=
Fm:
y u= -P, 0
X
73 P, X
v=
73 P,
P =
P= P=
(1 - (3)x + (3 - 0 + (0 - l)y (3(1 - o)x + (0 - (3)yx + 0((3 - l)y'
ox - (3y + (3 - 0 , x-y ox - (3y , x-y
6.15. Classification of Yang-Baxter maps
F rv : u Fv:
=
yP,
u = y + P,
v
xP,
P
v=x
+ P,
=
with suitable constants
0:,
=
1 + (3 -
269
0: ,
x-v
p=o:-(3 ,
x-v
(3.
Each one of the maps F r , ... ,Fv is an involution and coincides with its companion maps, so that all four arrows in Figure 6.25 are described by the same formulas. Note also that these maps come with the intrinsically built-in parameters 0:, (3. Neither their existence nor a concrete dependence on parameters is presupposed in Theorem 6.38. A geometric interpretation of these parameters can be given in terms of singularities of the map; it turns out that the parameter 0: is naturally assigned to the edges x, u, while (3 is naturally assigned to the edges y, v. For instance, for the map Fr the parameter 0: is nothing but the cross-ratio of the four roots Xi of the polynomial ~(x), and similarly (3 is the cross-ratio of the four roots Yi of the polynomial 8(y). The most remarkable fact about the maps F r , ... ,Fv is their 3D consistency. For 'J = I, II, III, IV or V, denote the corresponding map Fy of Theorem 6.38 by Fy(o:, (3), indicating the parameters explicitly. Moreover, for any 0:1, 0:2, 0:3 E C, denote by Fij = F J-( O:i, O:j) the corresponding maps acting nontrivially on the i- th and the j- th factors of (ClP l )3. Theorem 6.39. (Normal forms of quadrirational maps on ClP l x ClP l are 3D consistent) For any 'J = I, II, III, IV or V, the system of maps Fij is 3D consistent, and also satisfies the Yang-Baxter relation with parameters (6.117). Proof. The proof can be obtained by a direct computation (Exercise 6.22). It will also follow from Theorem 6.40 below, after we provide a geometric interpretation of the maps F'J'. D
Actually, 3D consistency of quadrirational maps on ClP l x ClP l holds not only for the normal forms F'J' but also under much more general circumstances. The only condition for quadrirational [2:2] maps consists in matching singularities along all edges of the cube. Similar statements hold also for quadrirational [1:1] and [1:2] maps, so that in the case Xl = X2 = ClP l the properties of being quadrirational and of being 3D consistent are related very closely. The maps F'J' of Theorem 6.38 admit a very nice geometric interpretation. Consider a pair of nondegenerate conics Ql, Q2 on the plane ClP2 , so that both Qi are irreducible algebraic varieties isomorphic to ClPl. Take
270
6. Consistency as Integrability
X E Ql, Y E Q2, and let £ = (XY) be the line through X, Y (well-defined if X =1= Y). Generically, the line £ intersects Ql at one further point U =1= X, and intersects Q2 at one further point V =1= Y. This defines the map
(6.131)
:t(X, Y)
= (U, V);
see Figure 6.26 for the ]R2 picture. The map :t is quadrirational, it is an involution and moreover coincides with both its companions. This follows immediately from the fact that, knowing one root of a quadratic equation, the second is a rational function of the input data. Intersection points X E Ql n Q2 correspond to the singular points (X, X) of the map :t.
Figure 6.26. A quadrirational map on a pair of conics.
Generically, two conics intersect at four points; however, degeneracies can happen. There are five possible types I - V of intersection of two conics: I: II: III: IV: V:
four simple intersection points; two simple intersection points and one point of tangency; two points of tangency; one simple intersection point and one second order tangency point; one point of third order tangency.
All conics sharing a quadruple of points build a linear family, or a pencil of conics. There are five types I-V of pencils of conics. Using rational parametrizations of the conics:
Cpl :3 x
I--f
X(x) E Ql
C
Cp2,
resp.
Cpl:3 y
I--f
Y(y) E Q2
C
Cp2,
it is easy to see that :t pulls back to the map F : (x, y) I--f ('U, v) which is quadrirational on Cpl x Cpl. One shows by a direct computation that the maps F for the above five situations are exactly the five maps listed in Theorem 6.38. Now, we obtain the following geometric interpretation of the statement of Theorem 6.39.
6.15. Classification of Yang-Baxter maps
271
Figure 6.27. 3D consistency on a linear pencil of conics.
Theorem 6.40. (3D consistent maps on a pencil of conics) Let Qi, i = 1,2,3, be three nondegenerate members of a linear pencil of conics. Let X E Ql, Y E Q2 and Z E Q3 be arbitrary points on these conics. Define the maps 'Jij as in (6.131), corresponding to the pair of conics (Qi,Qj). Set (X2' YI) = 'J12(X, Y), (X3, Zl) = 'J13(X, Z), and (Y3, Z2) = 'J23(Y, Z). Then
X 23 = (X 3Y3) n (X 2Z 2) E Ql, (6.132)
Y13
= (X3Y3) n (Y1ZI)
E
Q2,
Z12 = (Y1Z1) n (X 2Z 2) E Q3.
In other words, the maps 'Jij are 3D consistent.
Proof. We will work with equations of lines and conics on Cp2 in homogeneous coordinates, and use the same notation for geometric objects and homogeneous polynomials vanishing on these objects. Construct the lines a = (YZ), b = (XZ) and c = (XY), respectively. Let
X 2 = (c n Qd \ X,
Y1 = (c n Q2) \ Y,
X3 = (b n QI) \ X,
Y3
= (a n Q2) \ y.
Next, construct the line C = (X 3 Y3 ), and let
X 23 = (CnQd \X3,
Y13 = (CnQ2) \ Y3·
Finally, construct the lines A = (Y1 Y13 ) and B = (X 2X 23 ). We have four points X, X 2, X3 and X 23 on the conic Ql, and two pairs of lines (C, c) and (B, b) through two pairs of these points each. Therefore, there exists J-ll E Cpl such that the conic Ql has the equation Ql = 0 with
Ql = J-l1Bb + Cc.
6. Consistency as Integrability
272
Similarly, the conic Q2 has the equation Q2 = 0 with Q2
= /-l2Aa + Cc.
Consider the conic Q1 - Q2 = /-llBb - /-l2Aa = O.
It belongs to the linear pencil of conics spanned by Q1 and Q2. Furthermore, the point Z = a n b lies on this conic. Therefore, the conic Q1 - Q2 must coincide with Q3, which has therefore the equation Q3 = 0 with
Q3 = /-ll Bb - /-l2 Aa.
Moreover, the two points Z2 = an Band Zl = b n A also lie on Q3. Since Z2 E B, we have B = (X2Z2). Similarly, since Zl E A, we have A = (Y1Zd. Finally, we find that the point Z12 = An B = (Y1Zd n (X 2Z 2) also lies on Q3, which is equivalent to (6.132). 0
6.16. Discrete integrable 3D systems The major part of this chapter has been devoted to the very rich theory of integrability of 2D equations, the root of which has been identified in their 3D consistency. In this last section we turn our attention to integrability of 3D systems, now understood as 4D consistency. The most striking feature is that the number of integrable systems drops dramatically with the growth of dimension: we know of only half a dozen of discrete 3D systems with the property of 4D consistency. All of them are of a geometric origin and in fact appeared already in Chapters 2, 3, and 4. We are going to briefly discuss their general algebraic features. In the 3D context, there are a priori many kinds of systems, according to where the fields are assigned: to the vertices, to the edges, or to the elementary squares of the cubic lattice. 6.16.1. Fields on 2-faces. Consider first the situation when the fields (assumed to take values in some space X) are assigned to the the elementary squares. Denote by a, b, c the fields attached to the 2-faces parallel to the coordinate planes 12, 13, 23, respectively. The system under consideration is a map F : X3 f---+ X3, which we write as F(a,b,c) = (T3a,T2b,T1C) = (a3, b2, cd. One can think of the fields a, b, c as sitting on the bottom, front, and left faces of a cube, and a3, b2 , q, on the top, back, and right faces. This is visualized in Figure 6.28. The concept of 4D consistency of such a map assumes that one can extend it to a four-dimensional square lattice. Thus, in addition to the fields a, b, c, there are fields d, e, f, attached to the 2-faces parallel to the coordinate planes 14, 24, 34, respectively. Initial data a, b, c, d, e, f are the
6.16. Discrete integrable 3D systems
273
c
b I
b2
I
/
/
}-----
a
/
Figure 6.28. 3D system on an elementary cube: a map with fields on 2-faces.
fields on six 2-faces of a 4D cube adjacent to one vertex. They allow one to apply the map F on four 3-faces of a 4D cube (the inner, bottom, front, and left ones in Figure 6.29): H23 : (a, b, c)
f---->
(a3, b2, cd,
H24 : (a, d, e)
f---->
(a4, d2, ed,
F 134 : (b, d, j)
f---->
(b 4, d3, h),
F234: (c,e,j)
f---->
(c4,e3,12).
Here Fijk denotes the map F acting on a 3-face of the coordinate directions 'B ijk . Now one can apply the map F on the other four 3-faces (the outer, top, back, and right ones):
F123 : (a4, b4, C4)
f---->
(a34, b24 , C14),
H24: (a3,d3,e3)
f---->
(a34,d23,e13),
F134 : (b2, d2, h)
f---->
(b24, d23, 112)'
F234 :
f---->
(C14, e13,
(Cl, el,
h)
h2).
Thus, one obtains two answers for each of the six fields a34, b24, C14, d23 , e13, h2, and the map F is 4D consistent if these pairs of answers identically coincide for all six fields and for all initial data. We mention here two examples of systems of the kind just discussed, both of geometric origin. The first is the discrete Darboux system which describes Q-nets in the affine setting; see Section 2.1.3. For this system, each 2-face of the coordinate direction 'B ij , i < j, carries a field consisting of a pair of real numbers hij"ji). The map is given by the formulas (6.133)
Tklij =
lij + liklkj , 1 -'jk,kj
k=l=i,j.
Theorem 6.41. (4D consistency of the discrete Darboux system) The discrete Darboux system (6.133) is 4D consistent.
The second example is the star-triangle map which describes T-nets; see Section 2.3.8. For this system, each 2-face of the coordinate direction 'Bij, i < j, carries just one real number aij, and the convention aij = -aji holds.
6. Cow;istency as Integrability
274
f
c
:b a "-----t . . . . . . . . . . . . . . .
c Cl..J
d
Figure 6.29. Initial data and results of two-stage application of a 4D consistent map with fields on 2-faces.
The map is given by the formulas (6.134)
Tkaij
aij
=aijajk
+ (ljkaki + (lkiaij
k
i- i. j.
A symmetric appearance of this formula is due to the above convention. If one would like to consider aij with i < j only, there would appear some minus signs in the denominator. Thus. in the index-free notation for the fields a = a12, b = a13, e = a23, used at the beginning of this section, the star-triangle map is written as (6.135)
F(a,b.e) = (a3,b 2,ed = (
a . b , e ). ab + be - ca ab + be - ea ab + be - ca
Theorem 6.42. (4D consistency of the star-triangle map) The startriangle map (6.135) is 4D consistent. As in the case of 2D systems (see Section 6.14). our definition of consistency cannot be written in terms of composition of maps. since each piece of the initial data is used simultaneously in two different maps. It turns out to be possible to change the initial value problem on a 4D cube in such a way that the resulting consistency condition can be formulated in terms of compositions. It is not difficult to realize that for this the initial data should be prescribed on six 2-faces (of all six two-dimensional coordinate directions) which form a surface topologically equivalent to a disk. Such a surface is depicted on the left in Figure 6.:30. One can apply to this initial surface two different sequences of flips of the kind depicted in Figure 6.28, both leading to the surface on the right in Figure 6.30. One sequence starts with flipping the inner 3-face. and then
6.16. Discrete integrable 3D systems
275
a34
d e
q4 b
b24
C
f
a
h2
- - - - - - ...
e13 d 23
Figure 6.30. Initial data surface for a map satisfying the fUllctional tetrahedron equation, and the result of its four-fold flipping.
the top, front, and right ones. Denote the maps corresponding to these flips by 3 ij k; they are "companion maps" for the original F, i.e., they arise from F by regarding it along various diagonals of the basic cube. There appears a composition of maps:
3 123 : (a, b, c) e-t (a3, b2, cd, 3 134 : (b 2, d 2, f) e-t (hl, d 23 , h),
3 12 4 : (a3, d, e)
e-t
3 234 : (cl,el,h)
(a3.J,
e-t
d2 , ed,
(cl,t,e13,f12).
Another sequence starts with flipping the left 3-face, and proceeds with the back, bottom, and outer ones:
3 234 : (c,e,f) 3 124 : (a,d3,e3)
e-t
e-t
(c.l,e3,h),
(a4,d 23 ,e13),
3 134 : (b, d, h)
e-t
3 123 : (a4, b4, C4)
(b 4, d3, !I2),
e-t
(a34' b24 , C14).
The requirement that the two chains of maps lead to identical results can be thus encoded in the formula (6.136)
Definition 6.43. (Functional tetrahedron equation) A map 3 : X3 --> X3 'is said to satisfy the functional tetrahedron equation if (6.136) holds, where each 3 ijk is a map on X6(a, b, c, d, e, f) acting as 3 on the factors of the Cartesian product X6 corresponding to the variables sitting on the faces parallel to the planes ij, ik, j k, and acting tr-ivially on the other three factor-so Thus, we see that the concept of functional tetrahedron equation essentially coincides with the concept of 4D consistency of 3D systems with fields on 2-faces, the main difference lying in how the initial value problem is posed for the system at hand. It can be demonstrated (see Exercise 6.26)
6. Consistency as Integrability
276
that the 4D consistency of the star-triangle map (6.135) is translated into the following result.
Theorem 6.44. (Star-triangle solution of the functional tetrahedron equation) The map (6.137)
S(a, b, c) = (a3, b2, cd = (
ab b ,a + c - abc, be b) a+c-ac a+c-ac
satisfies the functional tetrahedron equation (6.136). The map (6.137) is related to the map (6.135) via conjugation by b 1---* 1/ b. One of the integrability features of 4D consistent maps (or, equivalently, of maps satisfying the functional tetrahedron equation) is a 3D analog of the zero curvature representation. For the map (6.137) it is discussed in Exercise 6.27. 6.16.2. Fields on vertices. Another version of 3D systems deals with fields assigned to vertices. In this case each elementary cube carries just one equation (6.138) relating the fields x E X in its eight vertices. Such an equation should be solvable for any of its eight arguments in terms of the other seven. This is shown in Figure 6.31.
X23 __- - - - - - {
Xl23
X3 e----i-----{.X13
I
X2 •
- - - - - - -
Xl2
/ / /
X - - - - - - - - { . Xl
Figure 6.31. 3D system on an elementary cube: an equation with fields on vertices.
The 4D consistency of such a system is defined as follows. Initial data on a 4D cube are 11 fields x, Xi (1 ::::: i ::::: 4), Xij (1 ::::: i < j ::::: 4). This data allow one to uniquely determine, by virtue of (6.138), all the fields Xijk (1 ::::: i < j < k ::::: 4). Then one has four different possibilities to find X1234, corresponding to the four 3-faces adjacent to the vertex Xl234 of the 4D cube; see Figure 6.32. If all four values coincide for any initial data, then
6.16. Discrete integrable 3D systems
277
equation (6.138) is 4D consistent. For such systems, one can consistently impose equations (6.138) on all three-dimensional cubes of the lattice Z4.
X1234 1
X3
"
X134
1
"
1
~13
X2 ' 1
X
•
X12
--
," ',' X
.. ' • __ ~, __________ - ~ .. ' , , ' X24 "
X124
X14',
Figure 6.32. 4D consistency of a 3D system with fields on vertices.
The only examples of 4D consistent equations with one scalar field attached to each vertex we know are related to the star-triangle relation and appear through different factorizations of the face fields aij' Given a (complex-valued) solution aij of equation (6.134), the relations (6.139)
'Tkaij
'Tiajk
'Tjaki
aij
ajk
aki
yield the existence of a function (6.140)
Z :
zm -+ .) = L(Y1, >')L(x, >.) with the matrices L(x,>')=>'l+x. Can you derive this zero curvature representation, following the ideas of Theorems 6.4, 6.36? 6.21. Specializing the map (6.149), or, equivalently, (6.150) to the case when all fields belong to the algebra lHI of quaternions, show that the real parts of the quaternions x, X2 are equal, as well as the real parts of the quaternions y, Y1, and that the imaginary parts of the four quaternions build a (nonplanar) quadrilateral in ]R3 with opposite sides of equal length (Chebyshev quadrilateral). In other words, there exist (t, (3 E ]R and f, h, 12, f12 E ~lHI = su(2) c:::'. ]R3 such that
+ (h - J), = a1 + (112 - h),
x = (t1 X2
+ (12 - J), = (31 + (112 - h),
y = (31
Yl
with (6.151)
Ih2 -
121 =
Ih
-.n
Ih2 - hi
=
112 - fl·
Moreover, we have the relation (6.152)
h2 - h - 12
+f
=
1 (112 - J) x 2(a - (3)
(12 - h),
which fixes the proportionality coefficient between the vectors h2- ft and (112 - J) x (12 - h); these vectors are parallel due to
(112 - h - 12
12+ f
+ f, h2 - 1) = (112 - h - 12 + f, 12 - h) = 0,
which is equivalent to relations (6.151). The 3D consistency of map (6.149) yields the 3D consistency of equation (6.152) with ]R3- valued fields on vertices, if a, (3 is considered as a real-valued edge labelling.
6.17. Exercises
283
6.22. Check (by hand or with the help of a computer algebra system) that the maps from Theorem 6.38 satisfy the Yang-Baxter relation. 6.23. Construct zero curvature representations for the Yang-Baxter maps from Theorem 6.38, based on Theorem 6.36. 6.24. Consider a pencil of conics having a triple tangency point at the point (WI: W2 : W3) = (0 : 1 : 0) (in homogeneous coordinates on ClP'2). Conics of this pencil and their rational parametrization are given (in nonhomogeneous coordinates) by the formulas
Q(a): W2 -
wf - a
= 0,
X(x)
= (WI (x), W2(x)) =
Check that if QI = Q(a), Q2 = Q((3), then the map given in coordinates by Fv of Theorem 6.38.
(x, x 2 + a).
:.r defined
in (6.131) is
6.25. Consider a pencil of conics through four points 0= (0,0), (0,1), (1,0), (1, 1) E ClP'2, where nonhomogeneous coordinates (WI, W2) on the affine part C 2 of ClP'2 are used (any four points on ClP'2, no three of which lie on a straight line, can be brought into these four by a projective transformation). Conics of this pencil are described by the equation
Q(a): W2(W2 - 1) = aW1(W1 - 1). A rational parametrization of such a conic is given, e.g., by
x-a x(x-a)) X(x) = (WI(X), W2(x)) = ( 2 ' 2 . X -a x-a Here the parameter x has the interpretation of the slope of the line (OX). The values of x for the four points of the base locus of the pencil on Q(a) are x = a, 00, 0 and 1. Show that if Q1 = Q(a), Q2 = Q((3), then the map :.r defined in (6.131) is given in coordinates by H of Theorem 6.38. 6.26. The geometric content of the discrete Moutard equation Xij - x = aij (x j - Xi) is the parallelism of the lines (XXij) and (XiX j ). Therefore, there are in principle four ways to introduce the field aij as the proportionality coefficient between the two vectors under consideration: ±aij and ±1/aij. Prove that one can introduce the fields aij for six two-dimensional coordinate directions in £:;4 so that all four maps Sijk in (6.136) be given by the formulas (6.137). 6.27.* Check the following 3D analog of the zero curvature representation for the map (6.137):
L23(C)L13(b)LI2(a) = L12(a3)L13(b2)L23(cd, where
L12(a) =
(l~aa
I-a a
o
L13(b)
= (
-b 0
0
1 1+b 0
1-o b) , b
6. Consistency as Integrability
284
and L 23(C)
~G
o
-c l+c
Can you derive this representation? 6.28. Consider the system of linear equations
X2 - X = a(xI - x),
(6.153)
X3 - X = b(XI - x)
for a scalar-valued function x on Z3. Equations in (6.153), as well as the coefficients a, b, are naturally assigned to triangles; see Figure 6.33. Show
a3 . _---
x
X2
- -- -- --- - - - - -- -Xl2
Figure 6.33. Equations on triangles.
that the compatibility of these equations is assured as soon as the coefficients a, b satisfy the following equations: (6.154)
(a3 - l)(b - 1) = (b 2 - l)(a - 1),
which should be understood as a map (a, aI, b, bd f----+ (a3, b2)' Valid initial data for such a map can be prescribed on a surface shown in Figure 6.34.
Figure 6.34. Initial conditions for the system (6.154).
6.17. Exercises
285
Show that, due to the second equation in (6.154), there exists a scalar function 1 on Z3 such that a = hi h, b = hi h, and that this function solves the equation
h - 12 + 12 - h + h - h = o.
(6.155)
112
123
113
The function x on Z3 solves the multiratio equation, also known as the Schwarzian discrete KP equation: (Xl -
(6.156)
XI2)(X2 - X23)(X3 - X13) _
-1
(X12 - X2)(X23 - X3)(XI3 - xd -
.
Give a geometric interpretation of equations (6.153), (6.156) (hint: these equations encode a Menelaus configuration). Can you find a linear system similar to (6.153) which would generate the so-called bilinear octahedron equation (or bilinear Hirota equation, or discrete KP equation): (6.157) It is natural to call equations of the type (6.155), (6.156), (6.157), which do not involve the fields at the vertices X and X123 of an elementary cube, octahedron equations, as opposed to the general cube equations (6.138).
6.29. Octahedron equations (6.155), (6.156), (6.157) have a sort of 4D consistency property. One imposes such an equation for three 3D coordinate directions (ij4): X12
(6.158)
=
1(Xl,X2,X4,XI4,X24),
X13 = g(Xl, X3, X4, X14, X34), X23
=
h(x2, X3, X4, X24, X34).
Compared to the usua14D consistency of cube equations, the vertices x and Xij4 do not appear in this system, and only three equations are considered. Check that the following holds: equations (6.159)
Xl23
=
l(g, h, X34,
fI, h) = g(J, h, X24,j, h) =
h(J, g, X14,
j, fI),
are satisfied identically with respect to 11 independent variables (chosen as initial data):
In equations (6.159) the "hat" denotes the shift in the 4-th direction:
j = 74(J) =
1(xI4, X24,
X44,
X144, X244),
etc.
Verify also that for each of the above systems, an equation of the form (6.160)
6. Consistency as Integrability
286
holds. For instance, for the multiratio equation (6.156), (X14 -
X12)(X24 -
X23)(X34 -
X13) =
(X12 -
X24)(X23 -
X34)(X13 -
X14)
-1 .
6.18. Bibliographical notes Sections 6.1, 6.2: Continuous and discrete integrable systems. The theory of integrable systems (called also the theory of solitons) is a vast field in mathematical physics with huge literature. The focus of different publications in this area varies from algebraic geometry, enumerative topology, statistical physics, quantum groups and knot theory to applications in nonlinear optics, hydrodynamics and cosmology. We mention here a selection of mathematical monographs (in chronological order): Toda (1978), Novikov-Manakov-Pitaevskii-Zakharov (1980), Ablowitz-Segur (1981), Calogero-Degasperis (1982), Newell (1985), Faddeev-Takhtajan (1986), Ablowitz-Clarkson (1991), Dubrovin (1991), Matveev-Salle (1991), Hirota (1992), Korepin-Bogoliubov-Izergin (1992), Belokolos-Bobenko-Enol'skii-Its-Matveev (1994), Hitchin-Segal-Ward (1999), Kupershmidt (2000), Rogers-Schief (2002), Babelon-Bernard-Talon (2003), Reyman-SemenovTian-Shansky (2003), Suris (2003), Dubrovin-Krichever-Novikov (2004), Fokas-Its-Kapaev-Novokshenov (2006). Concerning the basic concrete example of these sections, the sine-Gordon equation: the Backlund transformation was found by Backlund (1884); the permutability theorem is due to Bianchi (1892). The zero curvature representation is due to Ablowitz-Kaup-Newell-Segur (1973) and Takhtajan (1974). The immersion formula for surfaces with constant negative Gaussian curvature in terms of the frame is in Sym (1985). The discretization (6.12) of the sine-Gordon equation along with its Backlund transformation is due to Hirota (1977b). The geometric meaning was uncovered in BobenkoPinkall (1996a). Sections 6.3, 6.4, 6.5: Integrable systems on graphs. Our presentation of the general theory of integrable systems on graphs follows BobenkoSuris (2002a). Examples of integrable systems on the regular triangular lattices were considered in Adler (2000), Bobenko-Hoffmann-Suris (2002) and Bobenko-Hoffmann (2003). The fundamental role of quad-graphs for discrete integrability was understood in Bobenko-Suris (2002a). A different framework for integrable systems on graphs was developed by Novikov with collaborators. In particular, the Laplace transformations on graphs were studied in Dynnikov-Novikov (1997), the theory of discrete Schrodinger operators on graphs was developed in Novikov (1999a, b ), and the scattering theory on trees is due to Krichever- Novikov (1999).
6.18. Bibliographical notes
287
Sections 6.6, 6.7: From 3D consistency to zero curvature representations and Backlund transformations. The idea of consistency (or compatibility) is at the core of the theory of integrable systems. It appears already in the very definition of complete integrability of a Hamiltonian flow in the Liouville-Arnold sense, which says that the flow may be included into a complete family of commuting (compatible) Hamiltonian flows; see Arnold (1989). In the discrete context the (d + 1)-dimensional consistency of d-dimensional equations was observed many times. In the case d = 1 it was used as a possible definition of integrability of maps in Veselov (1991). A clear formulation in the case d = 2 was given in Nijhoff-Walker (2001). A decisive step was made in Bobenko-Suris (2002a) and independently in Nijhoff (2002): it was shown that the existence of a zero curvature representation follows for two-dimensional systems from the three-dimensional consistency. Section 6.8: Geometry of boundary value problems for integrable 2D equations. The discussion of the Cauchy problem on quad-graphs in Subsection 6.8.1 follows Adler-Veselov (2004). Embedding of quad-graphs into cubic lattices as a purely combinatorial problem was studied in a more general setting of arbitrary cubic complexes in Dolbilin-Stan'ko-Shtogrin (1986, 1994) and Shtan'ko-Shtogrin (1992). Theorem 6.8 is due to KenyonSchlenker (2004). The notion of the quasicrystallic rhombic embeddings and the extension to multi-dimensional lattices in Subsection 6.8.2 is due to Bobenko-:l\1ercat-Suris (2005). Note that intersections of 01) with bricks correspond to combinatorially convex subsets of 'D, as defined in Mercat (2004). Section 6.9: 3D consistent equations with noncommutative fields. The notion of 3D consistency in the noncommutative setup was introduced in Bobenko-Suris (2002b), where also the derivation of the zero curvature representation was given. Further examples due to Adler and Sokolov can be found in Adler-Bobenko-Suris (2007). The discrete Calapso equation (6.63) together with its zero curvature representation appeared in Schief (2001). There is a big literature on noncommutative integrable systems. One of the fundamental results in the theory of quantum integrable systems with discrete space-time is the quantization of the Hirota system by FaddeevVolkov (1994). A systematic exposition of noncommutative integrable systems is given in Kupershmidt (2000). References on discrete noncommutative systems include Matveev (2000), Nimmo (2006), Schief (2007). Sections 6.10, 6.11, 6.12: Classification of discrete integrable 2D systems with fields on vertices. The classification of discrete integrable 2D systems based on the notion of 3D consistency was given in
288
6. Consistency as Integrability
Adler-Bobenko-Suris (2003, 2007). The first of this papers deals with equations possessing the cubical symmetry and the tetrahedron property (Theorem 6.27). In Sections 6.10, 6.11 we present the classification of the second paper made under much weaker assumptions (Theorems 6.18, 6.19). Equations (H3)o=o and (HI) are perhaps the oldest in the lists; they can be found in the work of Hirota (1977a,b). Equations (Ql) and (Q3)o=o go back to Quispel-Nijhoff-Capel-Van der Linden (1984). Equation (Q4) was found in Adler (1998) (in the Weierstrass normalization of an elliptic curve). This equation in the Jacobi normalization is due to Hietarinta (2005). Equations (Q2), (Q3)o=1, (H2) and (H3)o=1 appeared explicitly for the first time in Adler-Bobenko-Suris (2003). The master equation (Q4) was investigated in Adler-Suris (2004), where its relation to various 2D integrable systems was revealed. Special solutions to this equation were found in Atkinson-Hietarinta-Nijhoff (2007). Symmetries of quad-equations from our lists were studied in Papageorgiou-TongasVeselov (2006) and Rasin-Hydon (2007). A 3D consistent equation without the tetrahedron property was found in Hietarinta (2004). This equation was shown to be linearizable by RamaniJoshi-Grammaticos-Tamizhmani (2006). Its geometric interpretation is given in Adler (2006). Section 6.13: Integrable discrete Laplace type equations. The relation of discrete (hyperbolic) systems on quad-graphs to Laplace type (elliptic) equations was discovered in Bobenko-Suris (2002). Examples of Laplace type equations on graphs previously appeared in Adler (2001). The threeleg forms of integrable quad-equations were found in Adler-Bobenko-Suris (2003) (with a formula for (Q4) in the Weierstrass normalization). In AdlerSuris (2004) the three-leg form of (Q4) was used to derive elliptic Toda systems on graphs. Section 6.14: Yang-Baxter maps were introduced in Drinfeld (1992) under the name of set-theoretical solutions of the Yang-Baxter equation. In Veselov (2003) the term "Yang-Baxter maps" was proposed instead of "set-theoretical solutions", and various notions of integrability were studied. In particular, commuting monodromy maps were constructed and zero curvature representations were discussed. A general construction of zero curvature representations (Theorem 6.36) was given subsequently in SurisVeselov (2003). A good survey on the topic is by Veselov (2007).
The map of Example 1 first appeared in Adler (1993). Example 2 is treated in Goncharenko-Veselov (2004) along with more general Yang-Baxter maps on Grassmannians. Example 3 is investigated in Noumi-Yamada (1998) and in Etingof (2003).
6.18. Bibliographical notes
289
Section 6.15: Classification of Yang-Baxter maps. Quadrirational Yang-Baxter maps were introduced and classified in Adler-Bobenko-Suris (2004). On pencils of conics used in this classification one can read, for example, in Berger (1987). Section 6.16: Discrete integrable 3D systems. Various algebraic structures relevant for integrability of higher-dimensional discrete systems appeared in the literature. The role played in 2D by the zero curvature representation goes in 3D to the so-called local Yang-Baxter equation introduced in l'vlaillet-Nijhoff (1989). Several 3D systems possessing this structure were found in Kashaev (1996). The functional tetrahedron equation was introduced in Kashaev-Korepanov-Sergeev (1998) as one of the versions of the 4D consistency. Note that their notation is different from the one in formula (6.136): their indices 1 ::; i,j, k ::; 6 of Sijk numerate two-dimensional coordinate planes. This paper contains also a list of solutions of this equation possessing local Yang-Baxter representations with a certain Ansatz for the participating tensors. The discrete Darboux system was derived in Bogdanov-Konopelchenko (1995). The fact that the star-triangle map satisfies the functional tetrahedron equation was observed in Kashaev (1996). In discrete differential geometry the star-triangle map appeared in Konopelchenko-Schief (2002a). The discrete BKP equation goes back to Miwa (1982). Its double crossratio form is due to Nimmo-Schief (1997). Its 4D consistency was observed in Adler-Bobenko-Suris (2003). Theorem 6.47 is due to Tsarev-Wolf (2007). The first works on quantization of discrete differential geometry appeared recently. Quantum versions of the discrete Darboux system and its reduction for circular nets were investigated in Sergeev (2007) and BazhanovMangazeev-Sergeev (2008). A quantization of circle patterns is proposed in Bazhanov-Mangazeev-Sergeev (2007).
Section 6.17: Exercises. Ex. 6.8: This result is due to Adler-Sokolov; see Adler-Bobenko-Suris (2007). Ex. 6.9, 6.10, 6.11: See Adler-Bobenko-Suris (2003). Ex. 6.12: See Hietarinta (2004). Ex. 6.13: See Adler-Bobenko-Suris (2007). Ex. 6.14: See Adler-Bobenko-Suris (2007) and Atkinson (2008). Ex. 6.16: Unpublished result by Adler. Ex. 6.18: See Adler-Suris (2004). Ex. 6.19: In this generality the result seems to be new.
290
6. Consistency as Integrability
Ex. 6.20, 6.21: See Hoffmann (2008), Schief (2007), and Pinkall-Springborn-Weifimann (2007). Ex. 6.25: See Adler-Bobenko-Suris (2004). The map FI appeared also in a different context in Tongas-Tsoubelis-Xenitidis (2001). Ex. 6.27: See Kashaev-Korepanov-Sergeev (1998). Ex. 6.28: A related material can be found in Konopelchenko-Schief (2005). Ex. 6.29: From a work in progress with Adler.
Chapter 7
Discrete Complex Analysis. Linear Theory
7.1. Basic notions of discrete linear complex analysis Many constructions in discrete complex analysis are parallel to discrete differential geometry in the space of real dimension 2. Recall that a harmonic function U : ~2 c:::: C ---+ ~ is characterized by the relation 02U 02u b.u = ox2 + oy2 = o. A conjugate harmonic function v : ~2 c:::: C Riemann equations ov ou ov ox oy ox' Equivalently, j = u + iv : ~2 Cauchy-Riemann equation
c::::
C
---+
---+ ~
is defined by the Cauchy-
ou oy·
C is holomorphic, i.e., satisfies the
oj . oj oy = ~ ox· The real and the imaginary parts of a holomorphic function are harmonic, and any real-valued harmonic function can be considered as a real part of a holomorphic function. A standard classical way to discretize these notions is the following. A function u : 7!} ---+ ~ is called discrete harmonic if it satisfies the discrete Laplace equation
(b.u)m ,n =
Um+l ,n
+ Um-l ,n + Um,n+l + Um,n-l
-
4um,n = O.
-
291
7. Discrete Linear Complex Analysis
292
A natural domain of a conjugate discrete harmonic function v :
UZ 2 )*
---+ lR is the dual lattice; see Figure 7.1. The defining discrete Cauchy-Riemann
C(-- __ -yv I I I I
I I I I
U
0--- ---0 Figure 7.1. Regular square lattice and its dual.
equations read: Vm+l/2,n+I/2 - v m+I/2,n-I/2
Um+l,n - um,n,
Vm+l/2,n+I/2 - Vm-I/2,n-l/2
-(Um,n+1 - um,n),
with the natural indexing of the dual lattice; cf. Figure 7.2.
The corre-
9 VI I
t-----;-:
Uo ••
Vo 0- - - - - -0
--------(e*))
= 1.
e*Estar(yo;9*)
These conditions should be compared with conditions characterizing the angles 4> : E(9) U E(9*) ~ (0, n) of a rhombic embedding of a quad-graph '.D, which consist of (7.11) and
4>(e*) = 2n,
4>(e) = 2n,
(7.13) eEstar(xo;9)
e* Estar(y();S*)
for all Xo E V(9) and all Yo E V(9*). Thus, the integrability condition (7.12) says that the system of angles 4> : E(9) U E(9*) ~ (0, n) comes from
7.3. Integrable discrete Cauchy-Riemann equations
299
a realization of the quad-graph 'D as a rhombic ramified embedding in C. Flowers of such an embedding can wind around its vertices more than once. Another formulation of the integrability conditions is given in terms of the edges of the rhombic realizations.
Theorem 7.6. (Integrable Cauchy-Riemann equations in terms of rhombic edges) Integrability condition (7.9) for the weight function 1/ : E(9) U E(9*) -----; lR+ is equivalent to the following: there exists a labelling of directed edges of'D, (): E('D) -----; §l, such that, in the notation of Figure 7.4,
(7.14)
l/(Xo, Xl) =
1 ( 1/
Yo, YI
. ()o
) = ~ ()
0
- ()l
+ (). I
Under this condition, the 3D consistency of the discrete Cauchy-Riemann equations is assured by the following values of the weights 1/ on the diagonals of the vertical faces of D:
(7.15)
1/
( +) - . () - >.. x, y - ~ () + >..'
where () = ()(x, y), and>" E C is an arbitrary number which is interpreted as the label assigned to all vertical edges of D: >.. = ()(x, x+) = ()(y, y+).
So, integrable discrete Cauchy-Riemann equations can be written in a form with parameters assigned to directed edges of 'D:
(7.16)
f(yt) - f(yo) f(xI) - f(xo)
()l - ()o
()l
+ ()o'
where
and p : V(9) -----; C is a rhombic realization of the quad-graph 'D. Since ()l - ()o
()l
+ ()o
p(yt) - p(yo) p(xt) - p(xo) ,
we see that for a discrete holomorphic function f : V(9) -----; C, the quotient of diagonals of the f - image of any quadrilateral (xo, Yo, Xl, yt) E F ('D) is equal to the quotient of diagonals of the corresponding rhombus. A standard construction of zero curvature representation for 3D consistent equations, given in Theorem 6.4, leads in the present case to the following result.
Theorem 7.7. (Zero curvature representation of discrete CauchyRiemann equations) The discrete Cauchy-Riemann equations (7.16) admit a zero curvature representation with spectral parameter dependent 2 x 2
300
7. Discrete Linear Complex Analysis
matrices along (x, y) E E ('D) given by (7.17)
L(y, x, ex;
A) = (A + () o
-2()(f(x) + f(y))) , A-()
where () = p(y) - p(x). Linearity of the discrete Cauchy-Riemann equations is reflected in the triangular structure of the transition matrices. Also, all constructions of Section 6.8 can be applied to integrable discrete Cauchy-Riemann equations. In particular, for weights coming from a quasicrystallic rhombic embedding of the quad-graph 'D, with labels 8 = {±()I, ... , ±()d}, discrete holomorphic functions can be extended from the corresponding surface D,]) C Zd to its hull, preserving discrete holomorphy. Here we have in mind the following natural definition: Definition 7.8. (Discrete holomorphic functions on Zd) A function f : Zd - t C is called discrete holomorphic if it satisfies, on each elementary square of Zd, the equation
f(n f(n
(7.18)
+ ej + ek) - f(n) + ej) - f(n + ek)
For discrete holomorphic functions in Zd, the transition matrices along the edges (n, n + ek) of Zd are given by
(7.19)
Lk(n;
A) = (A + ()k o
IJ,
-2()k(f(n + ek) A - ()k
+ f(n)))
.
All results of this section hold also in the case of generic complex weights which leads to () E C and to parallelogram realizations of 'D.
7.4. Discrete exponential functions An important class of discrete holomorphic functions is built by discrete exponential functions. We define them for an arbitrary rhombic embedding p : V('D) - t C. Fix a point Xo E V('D). For any other point x E V('D), choose some path {Cj }j=1 C E('D) connecting Xo to x, so that Cj = (Xj-I, Xj) and Xn = x. Let the slope of the j-th edge be ()j = p(Xj) - p(xj-d E §1. Then n z+(). e(x; z) = z_/ .
II
j=1
J
Clearly, this definition depends on the choice of the point Xo E V('D), but not on the path connecting Xo to x.
7.4. Discrete exponential functions
301
An extension of the discrete exponential function from n'D to the whole of Zd is given by the following simple formula: (7.20)
e(n;z)
=
+ fh)nk . rrd (Zz-fh k=l
The discrete Cauchy-Riemann equations for the discrete exponential function are easily checked: they are equivalent to a siq'lple identity
( Z + Bj z - Bj
.
z + Bk _ l)j(Z + Bj z - Bk z - Bj
_
z + BI) = Bj Z - ~k Bj
+ Bk -
.
Bk
At a given n E Zd, the discrete exponential function is rational with respect t? the parameter z, with poles at the ~. ElB l , ... , EdBd, where Ek = sIgn nk. / Equivalently, one can identify the discrete exponential function by its initial values on the axes:
z+Bk)n e(nek;z)= ( Z-Bk .
(7.21)
Another characterization says that e ( .; z) is the Backlund transformation of the zero solution of discrete Cauchy-Riemann equations on Zd, with the "vertical" parameter z. We now show that the discrete exponential functions form a basis in some natural class of functions (growing not faster than exponentially).
Theorem 7.9. (Discrete exponentials form a basis of discrete holomorphic functions) Let f be a discrete holomorphic function on V(1)) rv V(n'D), satisfying
'In E V(n'D),
(7.22)
with some C E R Extend it to a discrete holomorphic function on the hull 9{(V(n'D)). There exists a function g defined on the disjoint union of small neighborhoods around the points ±Bk E C and holomorphic on each of these neighborhoods, such that
(7.23)
1. f(n) - f(O) = -2
rg(.\)e(n; .\)d.\,
1n ir
where r is a collection of 2d small loops, each running counterclockwise around one of the points ±Bk.
Proof. The proof is constructive and consists of three steps. (i) Extend f from V(n'D) to 9{(V(n'D)); inequality (7.22) propagates in the extension process, if the constant C is chosen large enough.
7. Discrete Linear Complex Analysis
302
(ii) Introduce the restrictions I~k) of I : 9-C(V(f22»))
-----t
C to the coor-
dinate axes:
(iii) Set g(A) = 'L%=1 (gk(A) + g-k(A)), where the functions g±k(A) vanish everywhere except in small neighborhoods of the points ±Ok, respectively, and are given there by convergent series (7.24)
1 ((k) gk(A) = 2A 11 - 1(0)
~(A-Ok)n (k) +~ A + Ok (fn+1
(k») - In-I) ,
and a similar formula for g-k(A). Formula (7.23) is then easily verified by computing the residues at A = ±Ok (see Exercise 7.5).
o It is important to observe that the data I~k), necessary for the construction of g(A), are not among the values of I on V(1)) rv V(f22») known initially, but are encoded in the extension process.
7.5. Discrete logarithmic function We now define the discrete logarithmic function on a rhombic quad-graph 1). Fix some point Xo E V(1)), and set
(7.25)
£(x) =
2~i
Ir 10;i
A) e(x; A)dA,
'\Ix
EV(1)).
Here the integration path r is the same as in Theorem 7.9, and fixing Xo is necessary for the definition of the discrete exponential function on 1). To make (7.25) a valid definition, one must specify a branch of 10g(A) in a neighborhood of each point ±Ok. This choice depends on x, and is done as follows. Assume, without loss of generality, that the circular order of the points ±Ok on the positively oriented unit circle §,1 is the following: 01, ... , 0d, -01, ... , -Od. We set Ok+d = -Ok for k = 1, ... , d, and then define Or for all r E Z by 2d-periodicity. For each r E Z, assign to Or = exp(i[r) E §,1 a certain value of the argument "Ir E JR.: choose a value "II of the argument of 01 arbitrarily, and then extend it according to the rule
"Ir+1-"Ir
E
(0,7f),
'\Ir E Z.
Clearly, "Ir+d = "Ir + 7f, and therefore also "Ir+2d = "Ir + 27f. It will be convenient to consider the points Or, supplied with the arguments "ITl as belonging to the Riemann surface A of the logarithmic function (a branched covering of the complex A-plane).
7.5. Discrete logarithmic function
303
For each Tn E Z, define the "sector" Urn on the plane C carrying the quad-graph 1> as the set of all points of V(1)) which can be reached from .TO along paths with all edges from {Om, ... , Om+d- d. Two sectors Unq and Um2 have a nonempty intersection if and only if ITnI - Tn21 < d. The union U = U:=-= Urn is a branched covering of the quad-graph 1>, and it serves as the domain of the discrete logarithmic function. The definition (7.25) of the latter should be read as follows: for x E E A. The integration path r consists of d small loops on A around these points, and arg('\) = S'log('\) takes values in a small open neighborhood (in JR) of the interval
Urn, the poles of e(x;'\) are exactly the points Om, ... ,Om+d-l
(7.26) of length less than 7f. If Tn increases by 2d, the interval (7.26) is shifted by 27f. As a consequence, the function e is discrete holomorphic, and its restriction to the set V(9) of "black" points is discrete harmonic everywhere on U except at the point .TO: (7.27) Thus, the functions gk in the integral representation (7.23) of an arbitrary discrete holomorphic function, defined originally in disjoint neighborhoods of the points Qr, in the case of the discrete logarithmic function are actually restrictions of a single analytic function log('\) / (2,\) to these neighborhoods. This allows one to deform the integration path r into a connected contour lying on a single leaf of the Riemann surface of the logarithm, and then use standard methods of complex analysis to obtain asymptotic expressions for the discrete logarithmic function. In particular, one can show that at the "black" points of V(9), (7.28)
€(x) '"" log Ix
- xol,
.T --) 00.
Properties (7.27), (7.28) characterize the discrete Green's function on Thus:
9.
Theorem 7.10. (Discrete Green's function) The discrete logarithmic function on 1>. restricted to the set of vertices V(9) of the "black" graph 9, coincides with discrete Green's function on 9. Now we extend the discrete logarithmic function to Zd, which will allow us to gain significant additional information about it. In addition to the unit vectors ek E Zd (corresponding to Ok E §l), we introduce their opposites ek+d = -ek, k E [1, d] (corresponding to Ok+d = -Ok), and define e r for all
7. Discrete Linear Complex Analysis
304
r E Z by 2d-periodicity. Then m+d-l
8m =
(7.29)
EB
Ze r C Zd
r=m
is a d-dimensional octant containing exactly the part of n1) which is the P-image of the sector Um C 'D. Clearly, only 2d different octants appear among the 8 m (out of 2d possible d-dimensional octants). Define 8m as the octant ~m equipp!d with the interval (7.26) of values for 'Slog(Br ). By definition, 8 m1 and 8 m2 intersect if the underlying octants 8 m1 and 8 m2 have a nonempty intersection spanned by the common coordinate semi axes Ze r , and the 'S log( Br) for these common semiaxes match. It is easy to see that 8 m1 and 8 m2 intersect if and only if iml - m2i < d. The union 8 = U~=-OO 8m is a branched covering of the set U~=l 8 m C Zd. ~
~
Definition 7.11. (Disc~ete logarithmic function on Zd) The discrete logarithmic function on 8 is given by the formula (7.30)
£( ) = _1 n 27ri
( . A)dA irr10g(A) 2A e n, ,
'in E 8,
where for n E 8 m the integration path r consists of d loops around the points Bm, . .. , Bm+d- 1 on A, and'S 10g(A) on r is chosen in a small open neighborhood of the interval (7.26).
The discrete logarithmic function on 'D ca~ be described as the restriction of the discrete logarithmic function on 8 to a branched covering of n1) rv 'D. This holds for an arbitrary quasicrystallic quad-graph 'D whose set of edge slopes coincides with e = {±Bl, ... , ±Bd}. Now we are in a position to give an alternative definition of the discrete logarithmic function. Clearly, it is completely characterized by its values ~ner), r E [m, m + d -1], on the coordinate semiaxes of an arbitrary octant 8 m . Let us stress once more that the points ner do not lie, in general, on the original quad-surface n1). Theorem 7.12. (Values of discrete logarithmic function on coordinate axes) The values ~~) =~£(ner), r E [m,m + d - 1], of the discrete logarithmic function on 8 m C 8 are given by: (7.31)
£~) =
{
2(1+~+."'+ n~l)' 10g(Br)=Z"fr,
Here the values 10g(Br )
= i'"Yr
n even, n odd.
are chosen in the interval (7.26).
Proof. Comparing formula (7.30) with (7.24), we see that the values e~) can be obtained from the expansion of 10g(A) in a neighborhood of A = Br
7.5. Discrete logarithmic function
305
into the power series with respect to the powers of (A - Or) / (A + Or). This expansion reads:
~l-(-1)n(A-O )n n A+O:'
10g(A)=10g(Or)+~
Thus, we come to a simple difference equation n(e(r) _ e(r) ) n+l n-l
(7.32)
= 1 _ (_l)n ,
with the initial conditions (7.33)
£~r) = £(0) = 0,
o
which yield (7.31).
Observe that values (7.31) at even (resp. odd) points imitate the behavior of the real (resp. imaginary) part of the function 10g(A) along the half-lines arg(A) = arg(Or)' This can be easily extended to the whole of S. Restricted to the black points n E S (those with nl + ... + nd even), the discrete logarithmic function models the real part of the log~rithm. In particular, it is real-valued and does not branch: its values on 3 m depend on m (mod 2d) only. In other words, it is a well-defined function on 3 m . On the contrary, the discrete logarithmic function restricted to the white points n E S (those with nl + ... + nd odd) takes purely imaginary values, and increases by 27fi as m increases by 2d. Hence, this restricted function models the imaginary part of the logarithm. It turns out that recurrence relations (7.32) are characteristic for an important class of solutions of the discrete Cauchy-Riemann equations, namely for the isomonodromic solutions. In order to introduce this class, recall that discrete holomorphic functions in Zd possess a zero curvature representation with transition matrices (7.19). The moving frame we, A) : Zd --> GL(2, C) [A] is defined by prescribing some w(O; A), and by extending it recurrently according to the formula
(7.34) Finally, define the matrices A(-; A) : Zd
(7.35)
A(n; A) =
-->
gl(2, C)[A] by
dW~~; A) w- 1 (n; A).
These matrices satisfy a recurrence relation, which is obtained by differentiating (7.34),
(7.36)
A(n + ek; A) =
dLk(n; A) -1 -1 dA Lk (n; A) + Lk(n; A)A(n; A)Lk (n; A),
and therefore they are determined uniquely upon fixing some A(O; A).
7. Discrete Linear Complex Analysis
306
Definition 7.13. (Isomonodromy) A discrete holomorphic function f : Zd -+ C is called isomonodromic if, for some choice of A(O; .\), the matrices A(n;.\) are meromorphic in .\, with poles whose positions and orders do not depend on n E Zd. This term originates in the theory of integrable nonlinear differential equations, where it is used for solutions with a similar analytic characterization.
It is clear how to extend Definition 7.13 to functions on the covering 5. In the following statement, we restrict ourselves to the octant 51 = (Z+)d for notational simplicity.
Theorem 7.14. (Discrete logarithmic function is isomonodromic) For a proper choice of A(O; .\), the matrices A(n;.\) at any point n E (Z+)d have simple poles only: . _ A(O)(n) A(n,.\) .\
(7.37)
~ (B(l)(n)
+L
.\
1=1
with
(~
(7.39)
n, (~ C(l)(n)
~
'"
C(l)(n))
+ .\ _ ()
I
,
I
(-I)n;: +"'),
(7.38)
(7.40)
+
()
-(f(n)
(~
f(n
+;(n -
e,))) .
+ e,; H(n))
.
At any point n E 5, the following constraint holds: d
(7.41)
Lnl(e(n+q) -£(n-e/))
=
1- (-lt 1 +··+ nd •
/=1
Proof. The proper choice of A(O;.\) mentioned in the Theorem, can be read off formula (7.38): A(O;.\)
=
1(0 1)
>:
0 0 .
The proof consists of two parts. (i) First, one proves the claim for the points of the coordinate semiaxes. For any r = 1, ... , d, construct the matrices A(ne r ;.\) along
7.6. Exercises
307
the r-th coordinate semi-axis via formula (7.36) with transition matrices (7.19). This formula shows that the singularities of A(ner ; A) are poles at A = 0 and at A = ±()r, and that the pole A = 0 remains simple for all n > O. By a direct computation and induction, one shows that it is exactly the recurrence relation (7.32) for fAr) = f(ne r ) which assures that the poles A = ±()r remain simple for all n > O. Thus, (7.37) holds on the r-th coordinate semiaxis, with B(l)(ne r ) = C(l) (ne r ) = 0 for I =1= r.
(ii) The second part of the proof is conceptual, and is based upon the multidimensional consistency only. Proceed by induction, filling out the hull of the coordinate semiaxes: each new point is of the form n + ej + ek, j =1= k, with three points n, n + ej, and n + ek known from the previous steps, where the statements of the proposition are assumed to hold. Suppose that (7.37) holds at n + ej, n + ek. The new matrix A(n + ej + ek; A) is obtained by two alternative formulas, (7.42)
A(n+ej+ek;A)=
dLk(n + ej; A) -1 dA Lk (n+ej;A)
+Lk(n + ej; A)A(n + ej; A)L};l(n + ej; A),
and the other with k and j interchanged. Equation (7.42) shows that all poles of A(n + ej + ek; A) remain simple, with the possible exception of A = ±()k, whose orders might increase by 1. The same statement holds with k replaced by j. Therefore, all poles remain simple, and (7.37) holds at n + ej + ek. Formulas (7.38)-(7.40) and constraint (7.41) follow by direct computations based on (7.42). D
7.6. Exercises 7.1. Let 'D be a bipartite quad-graph, with black vertices Xj and white vertices Yj. Let J-l : E('D) -> C be a function such that, for any elementary quadrilateral (xo, Yo, Xl, yd E F('D), J-l(Xo, YO)J-l(Xl, yd = J-l(Xo, Yl)J-l(Xl, yo).
Show that there exists a function () : V('D) -> C such that for every edge (x, y) E E('D) we have iJ-l(x, y) = ()(y)/()(x). If J-l is real-valued, then one can assume that () takes real values at black points and imaginary values at white points.
7.2. Prove by induction that the entries of the matrix
7. Discrete Linear Complex Analysis
308
are given by
A=
~(II(1+ivk)+ II(l-ivk)), k
k
B=
~i(II(1+ivk)- II(1-ivk))' k
k
7.3. Check that the function f : 7i} ----+ C given by f(m, n) = (m(h satisfies the discrete Cauchy-Riemann equation
f(m f(m
+ 1, n + 1) - f(m, n) + 1, n) - f(m, n + 1)
+ n(h)2
(h + (h 01 - O2'
Generalize this function ("discrete z2,,) for 7l,d and for arbitrary quad-graphs 1).
7.4. Find the "discrete z3", i.e., the function f : 7l,2 ----+ C which is polynomial in m, n of degree 3, with cubic terms (mOl +n02)3, and satisfying the discrete Cauchy-Riemann equations.
7.5. Prove that for the functions 9k(>') from (7.24),
>. + Ok)n _ (k) Res,X=li k ( >. _ Ok 9k(>') - fn - f(O). 7.6. Estimate the difference £~k) -logn for the values given in (7.31), for n even. 7.7. Bibliographical notes Section 7.1: Basic notions of discrete linear complex analysis. The standard discretization of harmonic and holomorphic functions on the regular square grid goes back to Ferrand (1944) and Duffin (1956). This discretization of the Cauchy-Riemann equations apparently preserves the majority of important structural features. A pioneering step in the direction of further generalization of the notions of discrete harmonic and discrete holomorphic functions was undertaken by Duffin (1968), where the combinatorics of 7l,2 was given up in favor of arbitrary planar graphs with rhombic faces. A far reaching generalization of these ideas was given by Mercat (2001), who extended the theory to discrete Riemann surfaces. Section 7.2: Moutard transformation for discrete Cauchy-Riemann equations. For general Moutard transformations see the bibliographical note to Section 2.3 and Exercise 2.27. A further discussion of the Darboux transformation for discrete Laplace operators induced by the Moutard transformation for discrete Cauchy-Riemann equations can be found in DoliwaGrinevich-Nieszporski-Santini (2007). Section 7.3: Integrable discrete Cauchy-Riemann equations. Condition (7.13) on the system of angles ¢ : E(9) U E(9*) ----+ (0,71') characterizing rhombic embedding was given in Kenyon-Schlenker (2004). Theorems
7.7. Bibliographical notes
309
7.5, 7.6 characterizing 3D consistent (integrable) Cauchy-Riemann equations and their zero curvature representation from Theorem 7.7 are from Bobenko-Mercat-Suris (2005). Section 7.4: Discrete exponential functions. A discrete exponential function on 7!} was defined and studied in Ferrand (1944) and Duffin (1956). It was generalized for quad-graphs 'D in Mercat (2001) and Kenyon (2002). The question whether discrete exponential functions form a basis in the space of discrete holomorphic functions on 'D (Theorem 7.9) was posed in Kenyon (2002) and answered in Bobenko-Mercat-Suris (2005). Section 7.5: Discrete logarithmic function. The discrete logarithmic function on a rhombic quad-graph 'D was introduced in Kenyon (2002). Also the asymptotics (7.28) as well as Theorem 7.10 were proven in that paper. All other results in this section, starting with the extension of the discrete logarithmic function to Zd, are from Bobenko-Mercat-Suris (2005). For the theory of isomonodromic solutions of differential equations and its application to integrable systems see Fokas-Its-Kapaev-Novokshenov (2006). Isomonodromic constraint (7.41) was found in Nijhoff-RamaniGrammaticos-Ohta (2001), with no relation to the discrete logarithmic function.
Chapter 8
Discrete Complex Analysis. Integrable Circle Patterns
8.1. Circle patterns The idea that circle packings and, more generally, circle patterns serve as a discrete counterpart of analytic functions is by now well established. We give here a presentation of several results in this area, which treat the interrelations between circle patterns and integrable systems.
Definition 8.1. (Circle pattern) Let 9 be an arbitrary cell decompos'ition of an open or closed disk in C. A map Z : V(9) ---> C defines a circle pattern with combinatorics of 9 if the following condition is satisfied. Let y E F(9) '" V(9*) be an arbitrary face of 9, and let Xl, X2, . .. , Xn be its consecutive vertices. Then the points Z(XI), Z(X2), ... , z(xn) E C lie on a circle, and their circular order is just the listed one. We denote this circle by C(y), thus putting it into a correspondence with the face y, or, equivalently, with the respective vertex of the dual cell decomposition 9*.
As a consequence of this condition, if two faces yo, YI E F(9) have a common edge (xo. :rd, then the circles C(Yo) and C(yI) intersect in the points z(:r:t), Z(X2)' In other words, the edges from E(9) correspond to pairs of neighboring (intersecting) circles of the pattern. Similarly, if several faces Yl, Y2, ... , Ym E F (9) meet in one point Xo E V (9), then the corresponding circles C (Yl ), C (Y2), ... , C (Ym) also have a cornman intersection point Z (xo). A finite piece of a circle pattern is shown in Figure 8.1.
-
311
312
8. Integrable Circle Patterns
Figure 8.1. Circle pattern.
Given a circle pattern with combinatorics of 9, we can extend the function z to the vertices of the dual graph, setting
z(y) = center of the circle C(y),
Y E F(9) c::: V(9*).
After this extension, the map z is defined on all of VeD) = V(9) U V(9*), where']) is the double of 9. Consider a face of the double. Its z-image is a quadrilateral of the kite form, whose vertices correspond to the intersection points and the centers of two neighboring circles Co, C 1 of the pattern. Denote the radii of Co, C 1 by TO, Tl, respectively. Let xo, Xl correspond to the intersection points, and let Yo, Y1 correspond to the centers of the circles. Give the circles Co, C 1 a positive orientation (induced by the orientation of the underlying q, and let ¢ E (0, 7f) stand for the intersection angle of these oriented circles. This angle ¢ is equal to the kite angles at the "black" vertices z(xo), Z(X1); see Figure 8.2, where the complementary angle ¢* = 7f - ¢ is also shown. It will be convenient to assign the intersection angle ¢ = ¢( e) to the "black" edge e = (xo, xd E E(9), and to assign the complementary angle ¢* = ¢(e*) to the dual "white" edge e* = (Yo, yt) E E(9*). Thus, the function ¢ : E(9) U E(9*) -----7 (0,7f) satisfies (7.11). The geometry of Figure 8.2 yields following relations. First of all, the cross-ratio of the four points corresponding to the vertices of a quadrilateral face of']) is expressed through the intersection angle of the circles Co, C 1 :
(8.1)
q(z(xo), z(yo), z(xd, Z(Y1)) = exp(2i¢*).
8.2. Integrable cross-ratio and Hirota systems
313
Co
Figure 8.2. Two intersecting circles.
Furthermore, running around a "black" vertex of 'D (a common intersection point of several circles of the pattern), we see that the sum of the consecutive kite angles vanishes (mod 21f), hence:
rr
(8.2)
exp(i¢(e)) = 1,
Vxo
E
V(9).
eEstar(xo;9)
Finally, let 1/;01 be the angle of the kite (z(xo), z(Yo), Z(Xl), z(yI)) at the "white" vertex z(yo), i.e., the angle between the half-lines from the center z(yo) of the circle Co to the intersection points z(xo), z(xI) with its circle C 1 . It is not difficult to calculate this angle:
'oi,) TO+Tlexp(i¢*) exp (to/Ol = . TO + Tl exp( -i¢*)
(8.3)
Running around the "white" vertex of 'D, we come to the relation
(8.4)
rr rn
" )=1
TO
+
Tj
exp(i¢j)
-----,----"---,- = 1 TO + T)" exp( -i¢)~) ,
VyO
E
V(9*),
where the product is extended over all edges ej = (Yo, Yj) E star(yo; 9*), and ¢j = ¢(ej), while Tj are the radii of the circles Cj = C(Yj).
8.2. Integrable cross-ratio and Hirota systems Our main interest is in the circle patterns with prescribed combinatorics and with prescribed intersection angles for all pairs of neighboring angles. According to formula (8.1), prescribing all intersection angles amounts to prescribing cross-ratios for all quadrilateral faces of the quad-graph 'D. Thus, we come to the study of cross-ratio equations on arbitrary quad-graphs.
8. Integrable Circle Patterns
314
Let there be given a function Q condition
(8.5)
E(9)
Q(e*) = l/Q(e),
u E(9*)
----t
C satisfying the
\Ie E E(9).
Definition 8.2. (Cross-ratio system) The cross-ratio system on 1> corresponding to the function Q consists of the following equations for a function z: V(1)) ----t C, one for any quadrilateral face (:ro,yo,:rl,yd of1>:
(8.6)
q(z(xo), z(Yo), z(xt), Z(Yl))
= Q(xo, xd = l/Q(yo, yd·
An important distinction from the discrete Cauchy-Riemann equations is that the cross-ratio equations actually do not depend on the orientation of quadrilaterals. We have already encountered 3D consistent cross-ratio systems on Zd in Section 6.7 (see equation (6.33)), in the version with labelled edges. A natural generalization to the case of arbitrary quad-graphs is this: Yl
Xo
Yo Figure 8.3. Quadrilateral, with a labelling of undirected edges.
Definition 8.3. (Integrable cross-ratio system) A cross-ratio system is called integrable if there exists a labelling 0: : E(1)) ----t C of undirected edges of 1) such that the function Q admits the follow'ing factorization (in the notation of Figure 8.3): (8.7) Clearly, integrable cross-ratio systems are 3D consistent (see Theorem 4.26), admit Backlund transformations, and possess zero curvature r-epr-esentation with the transition matrices (6.47). It is not difficult to give an equivalent reformulation of the integrability condition (8.7). Theorem 8.4. (Integrability condition of a cross-ratio system) A cross-ratio system with the function Q : E(9) U E(9*) ----t C is integrable if
315
8.2. Integrable cross-ratio and Hirota systems
and only if for all Xo are fulfilled: (8.8)
E
V(9) and for all Yo
II
E
II
Q(e) = 1,
eEstar(xo;9)
V(9*) the following conditions Q(e*)
= 1.
e*Estar(yo;9*)
For a labelling of undirected edges a : E('D) --t C, we can find a labelling {} : E('D) --t C of directed edges such that a = (}2. The function P : V('D) --t C defined by p(y) - p(x) = {}(x, y) gives, according to (8.8), a parallelogram realization (ramified embedding) of the quad-graph 'D. The cross-ratio equations are written as (}2
(8.9)
q(z(xo), z(yo), z(xd, Z(Yl)) =
(}g = q(p(xo),p(Yo),p(Xd,p(Yl)); 1
in other words, for any quadrilateral (xo, Yo, Xl, Yl) E F('D), the cross-ratio of the vertices of its image under the map z is equal to the cross-ratio of the vertices of the corresponding parallelogram. In particular, one always has the trivial solution z(x) == p(x) for all x E V('D). A very useful transformation of the cross-ratio system is given by the following construction.
Definition 8.5. (Hirota system) For a given labelling of directed edges {} : E('D) --t C, the Nirota system consists of the following equations for the function W : V('D) --t C, one for every quadrilateral face (xo, Yo, Xl, Yl) E F('D):
(8.10) (}ow(xo)w(yo)
+ (}lW(YO)W(Xl)
- {}ow(xdw(Yd - (}lW(Yl)W(XO)
= o.
Note that the Hirota equation coincides with equation (6.30) of Section 6.7 (by the way, this shows that also in that previous version it was natural to assign parameters to directed edges). In terms of the parallelogram realization p : V('D) --t C of the quad-graph 'D corresponding to the labelling {}, equation (8.10) reads: (8.11)
+ W(YO)W(Xl) (p(xd - p(yo)) P(Xl)) + W(Yl)W(XO) (p(xo) - p(Yl))
W(xo)W(yo) (p(yo) - p(xo)) +w(xl)w(yd(p(yd -
= O.
Obviously, a transformation W f---> cw on V(9) and W f---> c-lw on V(9*) with a constant c E C, hereafter called a black-white scaling, maps solutions of the Hirota system into solutions. A relation between the cross-ratio and the Hirota system is based on the following observation:
Theorem 8.6. (Relation between cross-ratio and Hirota systems) Let W : V('D) --t C be a solution of the Hirota system. Then the relation (8.12)
z(y) - z(x) = (}(x,y)w(x)w(y) = W(X)W(y) (p(y) - p(x))
8. Integrable Circle Patterns
316
for all directed edges (x, y) E E('D) defines a unique (up to an additive constant) function z : V(1)) ----+ C which is a solution of the cross-ratio system (8.9). Conversely, for any solution z of the cross-ratio system (8.9), relation (8.12) defines a unique (up to a black-white scaling) function w : V(1)) ----+ C; this function w solves the Hirota system (8.10).
In particular, the trivial solution z(x) = p(x) of the cross-ratio system corresponds to the trivial solution of the Hirota system, w(x) == 1 for all x E V(1)). By a direct computation one can establish the following fundamental property.
Theorem 8.7. (Integrability of Hirota system) The Hirota system (8.10) is 3D consistent. As a usual consequence, the Hirota system admits Backlund transformations and possesses zero curvature representation with transition matrices along the edge (x, y) E E(1)) given by (8.13)
L(y, x, (); >.) = (
1 ->.()/w(x)
-()w(y) ) w(y)/w(x)
,
where () = p(y) - p(x).
8.3. Integrable circle patterns Returning to circle patterns, let {z(x) : x E V(9)} be the intersection points of the circles of a pattern, and let {z(y) : y E V(9*)} be their centers. Due to (8.1), the function z : V(1)) ----+ C satisfies a cross-ratio system with Q : E(9) U E(9*) ----+ §1 defined as Q(e) = exp(2i¢(e)). Because of (8.2), the first of the integrability conditions (8.8) is fulfilled for an arbitrary circle pattern. Therefore, integrability of the cross-ratio system for circle patterns with prescribed intersection angles ¢ : E(9*) ----+ (0,7f) is equivalent to (8.14)
II
exp(2i¢(e*))
= 1,
Vyo E V(9*).
e* Estar(yo; 9*)
This is equivalent to the existence of the edge labelling a : E(1)) that, in the notation of Figure 8.2, (8.15)
.
exp(2~¢*)
----+
C such
ao
= -. al
Moreover, one can assume that the labelling a takes values in §l. Our definition of integrable circle patterns will require somewhat more than integrability of the corresponding cross-ratio system.
8.3. Integrable circle patterns
317
Definition 8.8. (Integrable circle pattern) A circle pattern with prescribed intersection angles ¢ : E(9*) ---> (0,7r) is called integrable if
II
(8.16)
exp(i¢(e*))
= 1,
\:Iyo E V(9*),
e* Estar(yo; 9*)
i. e., if for any circle of the pattern the sum of its intersection angles with all neighboring circles vanishes (mod 27r).
This requirement is equivalent to a somewhat sharper factorization than (8.15), namely, to the existence of a labelling of the directed edges () : £('D) ---> §l such that, in the notation of Figure 8.2, (8.17)
exp( i¢) =
()l ()o
.A.*) = -
exp (Z'I'
()o
()l .
(Of course, the last condition yields (8.15) with 0: = ()2.) The parallelogram realization p : V(1)) ---> CC corresponding to the labelling () E §l is actually a rhombic one.
Theorem 8.9. (Isoradial integrability criterion) Combinatorial data 9 and intersection angles ¢ : E(9) ---> (0, 7r) belong to an integrable circle pattern if and only if they admit an isoradial realization. In this case, the dual combinatorial data 9* and intersection angles ¢ : E(9*) ---> (0, 7r) admit a realization as an isoradial circle pattern, as well. Proof. The rhombic realization p : V(1)) ---> CC of the quad-graph 1) corresponds to a circle pattern with the same combinatorics and the same intersection angles as the original one and with all radii equal to 1, and, 0 simultaneously, to an analogous dual circle pattern. Consider a rhombic realization p : V(1)) ---> CC of 1>. Solutions z : ---> CC of the corresponding integrable cross-ratio system which come from integrable circle patterns are characterized by the property that the z-image of any quadrilateral (xo, Yo, Xl, YI) from F(1)) is a kite with the prescribed angle ¢ at the black vertices z(xo), Z(Xl) (cf. Figure 8.2). It turns out that the description of this class of kite solutions admits a more convenient analytic characterization in terms of the corresponding solutions w : V('D) ---> CC of the Hirota system defined by (8.12).
V(1))
Theorem 8.10. (Circle pattern solutions of Hirota system) The solution z of the cross-ratio system corresponds to a circle pattern if and only if the solution w of the Hirota system, corresponding to z via (8.12), satisfies the condition (8.18)
w(X) E
§1,
w(y) E lR+,
\:Ix E V(9), y E V(9*).
8. Integrable Circle Patterns
318
The values w(y) E lR+ have then the interpretation of the radii of the circles C(y), while the (arguments of the) values w(x) E §l measure the rotation of the tangents to the circles intersecting at z (x) with respect to the isoradial realization of the pattern. Proof. As is easily seen, the kite conditions are equivalent to
lw(xo)l = 1 and Iw(xI}1
o
This yields (8.18), possibly upon a black-white scaling.
The conditions (8.18) form an admissible reduction of the Hirota system with () E §l, in the following sense: if any three of the four points w(xo), w(yo), w(xI}, W(Yl) satisfy the condition (8.18), then so does the fourth one. This is immediately seen, if one rewrites the Hirota equation (8.10) in one of the two equivalent forms: (8.19)
w(xI} w(xo)
w(yI) w(yo)
(}lW(Yl) - (}ow(yo) (}lW(YO) - (}ow(yI)
(}ow(xo) (}ow(xI)
+ (}lW(XI) + (}l w(xo)"
As a consequence of this remark, we obtain Backlund transformations for integrable circle patterns.
Theorem 8.11. (Backlund transformations of integrable circle patterns) Let all () E §l, and let p : V ('1» --- O
2a
This function is called the discrete logarithmic function; it should not be confused with the namesake function £(n) in the linear theory (Section 7.5). From (8.42) the following characterization is found: the discreteJogarithmic function L is the solution of the discrete cross-ratio system on S defined by
8. Integrable Circle Patterns
324
the values on the coordinate semiaxes L};) = L(ne r ), r E [m,m which solve the recurrence relation (Ln+l - Ln)(Ln - Ln-d 1 (8.43) n = Ln+1 - Ln- 1 2 with the initial conditions
Lg·) = L(O) =
(8.44)
+d -
1],
00,
where log er is chosen in the interval (7.26). Explicit expressions: (8.45)
(r)
L2n = loger
~1 1 + ~ k + 2n '
(r)
_
L2n+1 -loger
~1 + ~ k·
k=l
k=l
L
Theorem 8.17. (Circle pattern logarithm is isomonodromic The discrete logarithm is isomonodromic and satisfies, at any point n E S, the following constraint:
L nj (L(n + ej) d
(8.46)
L(n))(L(n) - L(n - ej)) = ~. L(n + ej) - L(n - ej) 2
j=l
By restriction to quad-surfaces On, we come to the discrete logarithmic function on arbitrary quasicrystallic quad-graphs 'D. By construction, they all correspond to circle patterns. A conjecture that these circle patterns are embedded seems plausible (see Figure 8.5).
Figure 8.5. Discrete logarithm circle patterns with combinatorics of the regular square and hexagonal lattices.
8.5. Linearization Let e : E('D) ---t C be an edge labelling, and let p : V('D) ---t C be the corresponding parallelogram realization of'D defined by p(y)-p(x) = e(x, y). Consider the trivial solutions zo(x) = p(x),
wo(x) = 1,
Vx E V('D)
8.5. Linearization
325
of the cross-ratio system (8.9) and the corresponding Hirota system (8.11). Suppose that Zo : V('D) ~ C belongs to a differentiable one-parameter family of solutions ZE : V('D) ~ C, E E (-EO, EO), of the same cross-ratio system, and denote by WE : V('D) ~ C the corresponding solutions of the Hirota system. Denote (8.47)
g -dZ-E I - dE E=O'
f =
(W;l dWE) dE
. E=O
Theorem 8.18. (Discrete derivative for discrete holomorphic functions) Both functions f,g : V('D) ~ C solve discrete Cauchy-Riemann equations (7.16). Proof. By differentiating (8.12), we obtain a relation between the functions f, 9 : V('D) ~ C:
(8.48)
g(y) - g(x) = (J(x)
+ f(y)) (p(y) - p(x)),
\i(x, y)
E
E('D).
The proof of the theorem is based on this relation solely. Indeed, the exactness condition for the form on the right-hand side on an elementary quadrilateral reads
(J(xo) + f(yo)) (p(yo) - p(xo)) + (J(yo) + f(Xl)) (p(Xl) - p(yo)) +(J(Xl) + f(Yl)) (p(Yl) - P(Xl)) + (J(Yl) + f(xo)) (p(xo) - p(Yl)) which is equivalent to (7.16) for the function condition for f, that is,
= 0,
f. Similarly, the exactness
(J(xo) + f(yo)) - (J(yo) + f(Xl)) + (J(xI) + f(Yl)) - (J(yI) + f(xo)) = 0, yields
g(yo) - g(xo) _ g(xI) - g(yo) p(yo) - p(xo) p(xI) - p(yo)
+ g(Yl) - g(xI) _ g(xo) - g(yI) p(yI) - p(xI)
= 0.
p(xo) - p(Yl)
Under the conditionp(yo)-p(xo) = p(Xl)-p(Yl), this is equivalent to (7.16) for g. D Remark. This proof shows that, given a discrete holomorphic function f : V('D) ~ C, relation (8.48) correctly defines a unique, up to an additive constant, function 9 : V('D) ~ C, which is also discrete holomorphic. Conversely, for any 9 satisfying the discrete Cauchy-Riemann equations (7.16), relation (8.48) defines a function f uniquely (up to an additive black-white constant); this function f also solves the discrete Cauchy-Riemann equations (7.16). Actually, formula (8.48) expresses that the discrete holomorphic function f is the discrete derivative of g, and so 9 is obtained from f by discrete integration. Summarizing, we have the following statement.
8. Integrable Circle Patterns
326
Theorem 8.19. (Linearization of circle patterns) a) A tangent space to the set of solutions of an integrable cross-ratio system, at a point corresponding to a rhombic embedding of a quad-graph, consists of discrete holomorphic functions on this embedding. This holds in both descriptions of the above set: in terms of variables z satisfying the crossratio equations, and in terms of variables w satisfying the Hirota equations. The corresponding two descriptions of the tangent space are related via the discrete derivative (resp. antiderivative) of discrete holomorphic functions. b) A tangent space to the set of integrable circle patterns of a given combinatorics, at a point corresponding to an isoradial pattern, consists of discrete holomorphic functions on the rhombic embedding of the corresponding quad-graph, which take real values at white vertices and pure imaginary values at black ones. This holds in the description of circle patterns in terms of circle radii and rotation angles at intersection points (Hirota system). A spectacular example of this linearization property is delivered by the isomonodromic discrete logarithm studied in Section 7.5 and isomonodromic z2a circle patterns of Section 8.4.
Theorem 8.20. (Linearization of w 2a - 1 circle patterns is the discrete logarithm) The tangent vector to the space of integrable circle patterns along the curve consisting of patterns w 2a - 1 , at the isoradial point corresponding to a = 1/2, is the discrete logarithmic function f defined in Section 7.5. Proof. We have to prove that the discrete logarithm f and the discrete power function w 2a - 1 are related by
f(n) =
(~~ w 2a - 1 (n)) 2 da
. a=1/2
Due to Theorem 8.18, it is enough to prove this for the initial data on the coordinate semiaxes. But this follows by differentiating with respect to a the initial values (8.31) at the point a = 1/2, where all w = 1: the result coincides with (7.31). 0
8.6. Exercises 8.1. Check that formulas (8.30), (8.31) give solutions to the corresponding difference equations (8.26), (8.28).
8.2. Prove asymptotic relations (8.32), (8.33). 8.3. Fill in the details of the proofs of Theorems 8.15, 8.16.
8.7. Bibliographical notel:i
327
8.4. For every solution z : 'lL d ---t C of the cross-ratio system (8.21), define the dual solution z* : 'lL d ---t C by z*(n+ej)-z*(n)= ( . z n
(]2 + ej) - z (n)
The dual solution is defined uniquely up to translation, and this freedom can be fixed by prescribing z*(O). Show that for a E (0,1) the dual solution to the discrete z2a, normalized to vanish at n = 0, coincides with the discrete z2(I-al.
8.5. Show that the limit a ---t 1 in Definition 8.12 leads to the discrete z2 as a solution of the cross-ratio equation, satisfying the recurrence relations (8.26) with a = 1 on the coordinate semiaxes, and with the initial data 2
z(O) = 0,
z(ej)
= 0,
z(2ej)
= ()j)
z(ej
+ ek) =
In particular, one sector of the discrete z2, defined ()l = 1, B2 = i, is characterized by the initial data z(O,O) = z(1, 0) = z(O, 1) = 0,
z(2,O) = 1,
Oil
()J - ()~ 2(10g()j -log()k) . ('lL+) 2 , in the case of
z(O,2) = -1,
8.6. Show that the dual solution to the discrete L.
z2
z(1, 1)
=
.2
2- • 1r
is the discrete logarithm
8.7. Show that for the cross-ratio system on ('lL+)2 with Bl = 1, B2 = 'i, the dual solution to z(m, '11.) = l/(m + in) is given by
z*(m,n)
=
1 3((m +in)3 - ('In -in)).
This can be regarded as the discrete z3.
8.7. Bibliographical notes Section 8.1: Circle patterns. The idea that circle packings and, more generally, circle patterns serve as a discrete counterpart of analytic functions is by now well established; see the monograph by Stephenson (2005). The origin of this idea is connected with the approach by Thurston (1985) to the Riemann mapping theorem via circle packings. Since then the theory bifurcated to several areas.
One of them is dealing mainly with approximation problems. The most popular are hexagonal packings, for which the convergence to the Riemann mapping was established in Rodin-Sullivan (1987). In He-Schramm (1998) it was shown that this convergence actually holds in the class Coo, that is, all higher derivatives are approximated. Sirnilar results are available also for
328
8. Integrable Circle Patterns
circle patterns with combinatorics of the square grid introduced in Schramm (1997), and even for more general circle patterns; see Bucking (2007). Another area concentrates around the uniformization theorem of KoebeAndreev-Thurston, and is dealing with circle packing realizations of cell complexes of prescribed combinatorics, rigidity properties, constructing hyperbolic 3-manifolds, etc.; see Thurston (1997), He (1999), Stephenson (2005). A variational description of circle packings was initiated by Colin de Verdiere (1991). Further progress is due to Bragger (1992), Rivin (1994), and Bobenko-Springborn (2004). The extremals of the functional used in the last paper are described by equation (8.4). An application of this approach in discrete differential geometry is the construction of discrete minimal surfaces through circle patterns in Bobenko-Hoffmann-Springborn (2006). The main topic of this chapter is interrelations of circle patterns with integrable systems. See the notes to Section 8.3.
Section 8.2: Integrable cross-ratio and Hirota systems. In this generality (for arbitrary quad-graphs) this material is due to Bobenko-Suris (2002a). On 71} the relation between the cross-ratio and Hirota systems is considered in Capel-Nijhoff (1995). Our presentation follows BobenkoMercat-Suris (2005). Section 8.3: Integrable circle patterns. Orthogonal circle patterns with combinatorics of the square grid were studied in Schramm (1997). Hexagonal circle patterns with fixed intersection angles were investigated in BobenkoHoffmann (2003), and with the multiratio property, in Bobenko-HoffmannSuris (2002). The general theory presented here is formulated in BobenkoMercat-Suris (2005). Section 8.4: za and log z circle patterns. The circle patterns za on the square lattice were introduced in Bobenko (1999) and studied in BobenkoPinkall (1999) and Agafonov-Bobenko (2000). The conjecture that these patterns are embedded, i.e., the interiors of different kites are disjoint, was formulated in the first of these papers. The study was extended to the regular hexagonal grid in Bobenko-Hoffmann (2003). The fact that the circle patterns za are immersed, i.e., the neighboring kites do not overlap, was proven in Agafonov-Bobenko (2000) for the square grid and in AgafonovBobenko (2003) for the hexagonal grid combinatorics. The embeddedness was proven in Agafonov (2003) for the case of the square grid combinatorics. The isomonodromic constraint (8.37) was obtained first for a = 1/2 in Nijhoff (1996), with no geometric interpretation. For the Hirota system, the isomonodromic constraint (8.41) was derived in Nijhoff-RamaniGrammaticos-Ohta (2001), also with no relation to geometry. Our presentation here follows Bobenko-Mercat-Suris (2005).
8.7. Bibliographical notes
329
Section 8.5: Linearization. The operation of discrete integration for discrete holomorphic functions was considered in Duffin (1956, 1968), and Mercat (2001). Linearization of circle patterns was studied in BobenkoMercat-Suris (2005); in particular, the derivation of Green's function from the za circle pattern is taken from this paper. Section 8.6: Exercises. Ex. 8.3: See Bobenko-Mercat-Suris (2005). Ex. 8.5: See Agafonov-Bobenko (2000). Ex. 8.6: See Agafonov-Bobenko (2000) in the case of the regular square grid. Ex. 8.7: See Bobenko-Pinkall (1999).
Chapter 9
Foundations
For the reader's convenience we give here a brief introduction to projective geometry and the geometries of Lie, Mobius, Laguerre and Plucker. We also include a number of classical incidence theorems relevant to discrete differential geometry. For extensive presentations of these classical results we recommend, in particular, the textbooks: Blaschke (1954) and Pedoe (1970) on projective and Plucker line geometry, Blaschke (1929) on sphere geometries, Cecil (1992) on Lie geometry, and Hertrich-Jeromin (2003) on Mobius geometry.
9.1. Projective geometry Projective geometry studies properties of geometric objects which remain invariant under the group of projective transformations, which is generated by Euclidean motions, homotheties, and central projections. A suitable analytical framework for doing projective geometry is given by the notion of homogeneous coordinates. The main space of real projective geometry is
which is the set of equivalence classes of equivalence relation: X rv Y
{:}
X = AY,
]RN+1 \
X,Y E ]RN+l \
{O} modulo the following {O},
A E ]R*.
On a general note, building the set of equivalence classes with respect to the relation rv is called a projectivization. Thus, points of JP>]RN are projectivizations of I-dimensional vector subspaces of ]RN+1.
-
331
332
9. Foundations
The equivalence class of x = (Xl, ... , XN, xN+d E JR.N+l \ {O} is denoted by [x] = [Xl: ... : XN : XN+I]. The space JR.N+l is called the space of homogeneous coordinates on JR.lP'N. One says that X E JR. N+I \ {O} is a lift of [x] to the space of homogeneous coordinates, or a representative of [x] in the space of homogeneous coordinates. The usual space JR.N can be identified with the subset of equivalence classes of elements of JR.N+I with XN+I # 0:
This subset is called an affine part of lP'JR. N . The complement of an affine part, i.e., the set of equivalence classes [Xl : ... : XN : 0], is called the hyperplane at infinity, and its elements are called infinitely remote points. Of course, XN+I plays a distinguished role in this construction. One can single out a coordinate other than XN+I and will then obtain different affine parts. The N + 1 affine parts obtained in this way build an atlas of lP'JR. N as a real manifold, consisting of N + 1 charts. More generally, projective hyperplanes in lP'JR. N are projectivizations of hyperplanes, that is, of vector N-spaces in JR.N+I. Any hyperplane U can be described by an equation N+I (u,x)
=
L
UiXi =
0,
i=l
where (UI, ... ,UN,UN+d E (JR.N+l)* \ {O}, and (-,.) denotes the pairing between the dual spaces JR.N+I and (JR.N+I)*. Actually, only the equivalence class [UI : ... : UN : UN+I] is relevant in this description, and a hyperplane can be identified with this equivalence class. One calls [UI : ... : UN : UN+I] the homogeneous coordinates of a hyperplane u. For instance, the hyperplane at infinity has homogeneous coordinates [0 : ... : 0 : 1]. Thus, the set (JR.lP'N)* of all projective hyperplanes is isomorphic to JR.lP'N, again. Interchanging the roles of points from JR.lP'N with hyperplanes from (JR.lP'N)* is the projective duality. For any 1 ~ d ~ N -1, a projective d-space in JR.lP'N is a projectivization of a vector (d + I)-space in JR.N+I. There are two dual ways to describe a projective d-space . • Let Xl, . .. ,Xd+l be d + 1 points of JR.lP'N in general position with representatives Xl, ... , Xd+l in the space of homogeneous coordinates. The general position condition means that the vector space L: = span(xI, ... ,Xd+l) has dimension d + 1. Then lP'(L:) is the d-dimensional space through Xl, ... , Xd+l. The points of lP'(L:) are given, in homogeneous coordinates, by all possible linear combinations X = cdl + ... + Cd+IXd+1 with (Cl l • • • , cd+d # (0, ... ,0).
333
9.1. Projective geometry
• Alternatively, let U1, ... , UN-d be N - d hyperplanes of (IRIP'N)* in general position, with representatives '111, ... , UN-d in the space of homogeneous coordinates. Again, the general position condition means that the vector space 1;~ = span( '11 1, ... , UN -d) has dimension N - d. Then the vector space
1; =
{x
E
IR N+ 1 : ('111, x) = ... = (UN-d,X) =
o}
has dimension d + 1, and 1P'(1;) is a projective d-space defined as the intersection of the hyperplanes U1,···, UN-d' If d 1 + d 2 2: N, then the intersection of a d 1-space with a d 2 -space in IRIP'N is a projective space of dimension 2: d1 + d2 - N, with an equality in the case of general position. For instance, intersection of a projective line with a projective hyperplane, in the general position, is a point. Here "general position" means that the line does not belong to the hyperplane.
The projective duality extends from the relation between points and hyperplanes to the relation between projective spaces of any two complementary dimensions d 1 and d 2 such that d 1 + d2 = N - 1. A famous and striking duality principle of projective geometry says that to any statement about incidence of projective spaces corresponds the dual statement where every object is replaced by its dual, with a simultaneous inversion of all incidence relations. For instance, in the plane projective geometry, the dual statements are obtained by replacing the notions according to the following (incomplete) vocabulary: a point a line a line a point a point on a line a line through a point the intersection point of two lines the line connecting two points three lines have a common point three points are collinear In the three-dimensional projective geometry, the incomplete vocabulary looks as follows: a plane a point a line a line a plane a point a line in a plane a line through a point the line connecting two points the intersection line of two planes the intersection point of a line the plane through a line with a plane and a point the intersection point of three planes the plane through three points four points are coplanar four planes have a common point
9. Foundations
334
Projective transformations or collineations of ~JlDN are induced by nondegenerate linear maps on the space of homogeneous coordinates: fj = rx,
r
E
GL(N + 1,~).
Theorem 9.1. (Fundamental theorem of projective geometry) a) Let, : ~JlDN --t ~JlDN be an injective map such that ,(~JlDN) does not lie in a hyperplane, and for any three collinear points Xl, X2, X3 their images ,(Xl), ,(X2), ,(X3) are also collinear. Then, is a projective transformation. b) For any two sets {Xl, ... , XN+2} C ~JlDN and {YI, ... ,YN+2} C ~JlDN such that in each set no N + 1 points lie in a hyperplane, there is a unique projective transformation, such that Yk = ,(Xk) for all k = 1, ... ,N + 2. A projective transformation of a line is characterized by the property of preserving the cross-ratios of four points. Finally, we briefly discuss the notion of quadric in a projective space. Let Q : ~N+l X ~N+I --t ~ be a nondegenerate symmetric bilinear form; we will denote the matrix of this form by Q. The set of points X E ~JlDN with homogeneous coordinates E ~N+I satisfying the quadratic equation
x
N+I
(9.1)
L
Q(x, x) =
QjkXjXk = 0
j,k=l is called a (nondegenerate) quadric Q C ~JlDN. Of course, only those nonempty quadrics are interesting which correspond to indefinite bilinear forms. In particular, a nondegenerate quadric in ~JlD2 is called a conic. Two points x, Y E ~N with homogeneous coordinates x, fj E ~N+l are called conjugate with respect to a quadric if N+I
L
Q(x, fj) =
QjkXjYk =
o.
j,k=l The points conjugate to a given point X E ~JlDN build the polar hyperplane of x. Thus, the polar hyperplane is defined as JlD(x~), where the orthogonal complement is taken according to the scalar product Q: x~ = {fj E ~N+I : Q(x, fj) =
O}.
Homogeneous coordinates of the polar hyperplane can be chosen as
U= Qx. Thus, the polar point
X
of a hyperplane u has homogeneous coordinates A
X
=
Q-I u. A
9.2. Lie geometry
335
Two hyperplanes u, v with homogeneous coordinates it, v E ~N+l are called conjugate with respect to the quadric Q if N+l
L
(Q-l )jkUjVk = O.
j,k=l
Each of them contains the polar point of the other. A tangent hyperplane to the quadric Q is self-conjugate, so its homogeneous coordinates satisfy the quadratic equation N+l
L
(9.2)
(Q-l )jkUjUk = O.
j,k=l
A quadric can be viewed either as the set of points satisfying (9.1) or as the envelope of its tangent hyperplanes satisfying (9.2). The polarity relation can be generalized from points and hyperplanes to projective spaces of arbitrary dimensions. For a projective d-space U = JP>(U) one defines a polar subspace as JP>(UJ..), where the orthogonal complement is understood with respect to the scalar product Q. Polarity can be regarded as a generalization of duality.
9.2. Lie geometry 9.2.1. Objects of Lie geometry. The following geometric objects in the Euclidean space ~N are elements of Lie geometry: • Oriented hyperspheres. A hypersphere in ~N with center c E ~N and radius r > 0 is described as S = {x E ~N : Ix - cl 2 = r2}. It divides ~N into two parts, inner and outer. Declaring one of the two parts of ~N to be positive, we come to the notion of an oriented hypersphere. Thus, there are two oriented hyperspheres S± for any S. One can take the orientation of a hypersphere into account by assigning a signed radius ±r to it. For instance, one can assign positive radii r > 0 to the hyperspheres with the inward field of unit normals and negative radii r < 0 to the hyperspheres with the outward field of unit normals . • Oriented hyperplanes. A hyperplane in ~N is given by the equation p = {x E ~N : (v, x) = d}, with a unit normal v E §N-l and d E R Clearly, the pairs (v, d) and (-v, -d) represent one and the same hyperplane. It divides ~N into two half-spaces. Declaring one of the two half-spaces to be positive, we arrive at the notion of an oriented hyperplane. Thus, there are two oriented hyperplanes p± for any P. One can take the orientation of a hyperplane into
336
9. Foundations
account by assigning the pair (v, d) to the hyperplane with the unit normal v pointing into the positive half-space.
• Points. One considers points x E radius.
]RN
as hyperspheres of a vanishing
• Infinity. One compactifies the space ]RN by adding the point at infinity 00, with the understanding that a basis of open neighborhoods of 00 is given, e.g., by the outer parts of the hyperspheres Ixl 2 = r2. Topologically the compactification so defined is equivalent to a sphere §N. • Contact elements. A contact element is a pair consisting of a point x E ]RN and an (oriented) hyperplane P through x; alternatively, one can use a normal vector v to P at x. A contact element represents the (equivalence class of) hypersurfaces through the point x with the tangent hyperplane P at x. In the framework of Lie geometry, a contact element can be identified with a set (a pencil) of all hyperspheres S through x which are in oriented contact with P (and with one another), thus sharing the normal vector v at x; see Figure g.1.
p
Figure 9.1. Contact element.
9.2.2. Projective model of Lie geometry. All the above elements are modelled in Lie geometry as points, resp. lines, in the (N + 2)-dimensional projective space lP'(]RN+l,2) with the space of homogeneous coordinates ]RN+l,2. The latter is the space spanned by the N + 3 linearly independent vectors el, ... , eN +3 and equipped with the pseudo-Euclidean scalar
9.2. Lie geometry
337
product
1, (ei, ej) = { -1, 0, It is convenient to introduce two (9.3)
eo
= ~(eN+2 -
i=jE{1, ... ,N+1}, i=jE{N+2,N+3}, i =1= j.
isotropic vectors e OCl
eN+l),
= ~(eN+2 + eN+d,
for which (eo, eo) = (e OCl , e OCl ) = 0, (eo, e oo ) = -~. The models of the above elements in the space ~N+l,2 of homogeneous coordinates are as follows: • Oriented hypersphere with center c E ~N and signed radius r E R' (9.4)
S=
c+ eo + (lcl 2 -
• Oriented hyperplane (v, x)
p=
(9.5)
v
r2)e oo
= d with v
+ reN+3. E §N-l and d E ~:
+ 0 . eo + 2de oo + eN+3.
• Point x E ~N :
x = x + eo + Ixl 2e oo + O· eN+3'
(9.6) • Infinity
00:
(9.7)
00 = e oo . • Contact element (x, P):
(9.8)
span(x,p).
In the projective space lP'(~N+l,2) the first four types of elements are represented by the points which are equivalence classes of (9.4)-(9.7) with respect to the relation ~ '" 'rJ {:} ~ = A'rJ with A E ~* for ~, 'rJ E ~N+l,2. A contact element is represented by the line in lP'(~N+l,2) through the points with representatives x and p. We mention several fundamentally important features of this model: (i) All the above elements belong to the Lie quadric lP'(lI.P+l,2), where (9.9)
JL N +1,2 = {~E ~N+l,2: (~,~) = O}. Moreover, the points of lP'(JL N+ 1,2) are in a one-to-one correspondence with the oriented hyperspheres in ~N, including degenerate cases: proper hyperspheres in ~N correspond to points oflP'(JL N+ 1,2) with both eo- and eN+3-components nonvanishing, hyperplanes in ~N correspond to points of lP'(JLN+l,2) whose eo-component vanishes, points in ~N correspond to points of lP'(JLN+l,2) with vanishing eN+3-component, and infinity corresponds to the only point of lP'(JLN+l,2) with both eo- and eN+3-components vanishing.
9. Foundations
338
(ii) Two oriented hyperspheres 5 1 ,52 are in oriented contact (i.e., are tangent to each other with the unit normals at tangency pointing in the same direction) if and only if
leI -
(9.10)
c21 2 = (rJ - r2) 2 ,
and this is equivalent to (S],S2)
=
O.
(iii) An oriented hypersphere 5 = {x E IRN : Ix-cl 2 = r2} is in oriented contact with an oriented hyperplane P = {x E IRN : (v, :r) = d} if and only if
(c,v) - r - d = O.
(9.11)
Indeed, equation of the hyperplane P tangent to 5 at Xo E 5 reads: (xo - c, x - c) = r2. Denoting by v = (c - xo) / r the unit normal vector of P (recall that the positive radii are assigned to spheres with inward unit normals), we can write the above equation as (v, x) = d with d = (c, (c - XO)/T) - T = (c, v) - T, which proves (9.11). Now, the latter equation is equivalent to (s,p) = O. (iv) A point x can be considered as a hypersphere ofradius r = 0 (in this case the two oriented hyperspheres coincide). An incidence relation x E 5 with a hypersphere 5 (resp. x E P with a hyperplane P) can be interpreted as a particular case of oriented contact of a sphere of radius T = 0 with 5 (resp. with P), and it takes place if and only if (x,t,) = 0 (resp. (x,p) = 0). (v) For any hyperplane P, (oo,p) = O. One can interpret hyperplanes as hyperspheres (of an infinite radius) through 00. More precisely, a hyperplane (v, x) = d can be interpreted as a limit, as r -+ 00, of the hyperspheres of radii T with centers located at c = TV + 1t, with (v, u) = d. Indeed, the representatives (9.4) of such spheres are
s
+ u) + eo + (2dT + (u, u) )e oo + TeN+3 (v + O(l/r)) + (l/T)eo + (2d + O(l/T))e oo + eN+3 p + O(l/T). (rv
Moreover, for similar reasons, the infinity 00 can be considered as a limiting position of any sequence of points x with Ixl -+ 00. (vi) Any two hyperspheres 5 1 ,52 in oriented contact determine a contact element (their point of contact and their common tangent hyperplane). For their representatives SI, S2 in IR N + l ,2, the line in IP'(IR N + I ,2) through the corresponding points in IP'(JLN+1.2) is isotropic, i.e., lies entirely on the Lie quadric IP'(JLN+1,2). This follows from
9.2. Lie geometry
339
Such a line contains exactly one point whose representative x has vanishing eN+3-component (and corresponds to x, the common point of contact of all the hyperspheres), and, if x of- 00, exactly one point whose representative p has vanishing eo-component (and corresponds to P, the common tangent hyperplane of all the hyperspheres). In case when an isotropic line contains 00, all its points represent parallel hyperplanes, which constitute a contact element through 00. Thus, if one considers hyperplanes as hyperspheres of infinite radii, and points as hyperspheres of vanishing radii, then one can conclude that: ~
Oriented hyperspheres are in a one-to-one correspondence with points of the Lie quadric JP>(IL,N+l,2) in the projective space JP>(~N+l,2).
~
Oriented contact of two oriented hyperspheres corresponds to orthogonality of (any) representatives of the corresponding points in JP>(~N+l,2).
~
Contact elements of hypersurfaces are in a one-to-one correspondence with isotropic lines in JP>(~N+l,2). We will denote the set of all such lines by J:~+1,2.
9.2.3. Lie sphere transformations. According to F. Klein's Erlangen Program, Lie geometry is the study of properties of transformations which map oriented hyperspheres (including points and hyperplanes) to oriented hyperspheres and, moreover, preserve the oriented contact of hypersphere pairs. In the projective model described above, Lie geometry is the study of projective transformations of JP>(~N+l,2) which leave JP>(IL,N+l,2) invariant, and, moreover, preserve orthogonality of points of JP>(lI...N+ 1,2) (which is understood as orthogonality of their lifts to 1I...N+ 1,2 C ~N+l,2; clearly, this relation does not depend on the choice of lifts). Such transformations are called Lie sphere transformations. Theorem 9.2. (Fundamental theorem of Lie geometry) a) The group of Lie sphere transformations is isomorphic to the factor O(N + 1, 2)/{±I}. b) Every line preserving diffeomorphism of JP>(lI....N+ 1,2) is the restriction to JP>(lI.,N+l,2) of a Lie sphere transformation.
Since vanishing of the eo- or eN+3-component of a point in JP>(lI.,N+l,2) is not invariant under a general Lie sphere transformation, there is no distinction between oriented hyperspheres, oriented hyperplanes and points in Lie geometry.
9. Foundations
340
9.2.4. Planar families of spheres; Dupin cyclides. Considerations of this subsection hold for the geometrically most significant case N = 3. Definition 9.3. (Planar family of spheres) A planar family of (oriented) spheres in]R3 is a set of spheres whose representatives s E JP'(JL4,2) are contained in a projective plane JP'(~), where ~ is a three-dimensional vector subspace of ]R4,2 such that the restriction of (-, .) to ~ is nondegenerate. Thus, a planar family of spheres is a conic section JP'(~ there are two possibilities:
n JL 4,2).
Clearly,
(a) The signature of (', ')IE is (2,1), and so the signature of (-, also (2,1).
')IE~
is
(b) The signature of (-, .) IE is (1, 2), and so the signature of (-, .) IE~ is (3,0). It is easy to see that a planar family is a one-parameter family, parametrized by a circle §1. Indeed, if e1, e2, e3 is an orthogonal basis of ~ such that (e1, e1) = (e2, e2) = -(e3, e3) = 1 (say), then the spheres of the planar family come from the linear combinations s = a1 e1 + a2e2 + e3 with (a1e1
+ a2e2 + e3, a1e1 + a2e2 + e3) = 0
¢}
ai + a~ = 1.
In the second case mentioned above, the space ~~ has only a trivial intersection with JL 4,2, so the spheres of the planar family JP'(JL4,2 n ~) have no common touching spheres. Definition 9.4. (Cyclidic family of spheres) A planar family of spheres is called cyclidic if the signature of (-, ')IE is (2,1), so that the signature of (-, ')IE~ is also (2,1). For any cyclidic family JP'(JL 4 ,2 n ~) there is a dual cyclidic family JID(JL 4,2 n ~~) such that any sphere of the first family is in oriented contact with any sphere of the second. The family JP'(JL 4,2 n~), as anyone-parameter family of spheres, envelopes a canal surface in ]R3, and this surface is an envelope of the dual family JP'(JL 4,2 n ~~), as well. Such surfaces are called Dupin cyclides. Examples: a) Points of a circle build a planar cyclidic family of spheres (of radius zero). The dual family consists of all (oriented) spheres through this circle, with centers lying on the line through the center of the circle orthogonal to its plane; see Figure 9.2, left. The corresponding Dupin cyclide is the circle itself. It can be shown that any Dupin cyclide is an image of this case under a Lie sphere transformation. b) Planes tangent to a cone of revolution build a planar cyclidic family of spheres, as well. The dual family consists of all (oriented) spheres tangent
9.3. Mobius geometry
341
to the cone, with centers on the axis of the cone; see Figure 9.2, right. The corresponding Dupin cyclide is the cone itself.
Figure 9.2. Left: A cyclidic family of spheres through a circle. Right: A cyclidic family of spheres tangent to a cone.
9.3. Mobius geometry 9.3.1. Objects ofM6bius geometry. Mobius geometry is a subgeometry of Lie geometry, with points distinguishable among all hyperspheres as those of radius zero. Thus, Mobius geometry studies properties of hyperspheres in ]RN invariant under the subgroup of Lie sphere transformations preserving the set of points. The following geometric objects are elements of Mobius geometry of ]RN: • Points x E ]RN. • Infinity
00
which compactifies
]RN
into
§N.
• (Nonoriented) hyperspheres S = {x E ]RN : centers c E ]RN and radii r > o. • (Nonoriented) hyperplanes P normals v E §N -1 and d E lR..
= {x
E ]RN :
Ix - cI2
= r2} with
(v, x) = d}, with unit
The Mobius group Mob(N) of ]RN consists of point transformations generated by reflections in hyperplanes P = {x E ]RN : (v, x) = d}:
(9.12)
x
f--r
x- 2
(v, x) - d ( ) v, v,v
and by inversions in hyperspheres S = {x
(9.13)
E]RN :
Ix - cl 2 = r2}:
r2
Xf--rC+
Ix-c 12 (x-c).
Clearly, Mob(N) contains as a subgroup the group E(N) of Euclidean motions of ]RN, which is generated by reflections in hyperplanes. It contains
9. Foundations
342
also dilations, since they can be represented as compositions of inversions in two concentric hyperspheres. For N 2: 3, the Liouville theorem says that Mob(N) coincides with the group of conformal diffeomorphisms. Yoo
]RN
Figure 9.3. Stereographic projection.
One compactifies ]RN by adding the point 00, thus arriving at the Nsphere §N. It is convenient to model §N as embedded in ]RN +1 : §N
= {Y E ]RN+l : (y,y) = 1}
(we use one and the same notation for the scalar products in ]RN and in ]RN+l; its meaning in each case should be clear from the context). The (inverse) stereographic projection 0" : ]RN - t §N \ {Yoo} from the north pole Yoo = eN +1 is defined by (9.14)
see Figure 9.3. The formula
2
O"(x)
= eN+l + 1x 12 + 1 (x
- eN+d
= eN+l + 1x -
2 eN+l
12
(x - eN+d
shows that one can view the stereographic projection 0" also as the restriction to ]RN of the inversion of ]RN+l in the hypersphere with center eN+l and radius.)2. Setting 0"(00) = Yoo makes 0" to a diffeomorphism 0" : ]RN U {(X)} - t §N. Hyperplanes and hyperspheres in ]RN are mapped by the stereographic projection 0" to hyperspheres in §N, the images of hyperplanes being hyperspheres through Yoo' Thus, hyperplanes in ]RN can be interpreted as hyperspheres through 00.
9.3. Mobius geometry
343
Elements of Mobius geometry of §N are: • Points y E
§N .
• (Nonoriented) hyperspheres S C
§N.
Any hypersphere S C §N, except for great ones, may be described as the intersection of §N with an affine hyperplane {y E IRN+l : (s, y) = 1}. The point s E IR N +1 , which is the pole of this hyperplane with respect to §N, lies outside of §N, and S C §N is the contact set of §N with the tangent cone to §N with apex s. Also, S c §N is the intersection of §N and the orthogonal N-sphere S C IRN +1 with center s and radius p such that p2 = (s, s) - 1, see Figure 9.4. (For a great hypersphere S C §N, which is the intersection of §N with a hyperplane {y E IR N +1 : (s, y) = O}, the latter hyperplane also plays the role of the orthogonal N-sphere S, and the tangent cone becomes a cylinder.)
(8, y) = 1
S
Figure 9.4. Hypersphere S C §N and the corresponding point s E with an orthogonal N-sphere S through S.
]RN+l,
The Mobius group Mob(N) of spheres S C §N, given by
§N
is generated by inversions in hyper-
(9.15) Transformation (9.15) coincides with the restriction to §N of the inversion of IR N + 1 in the N-sphere S, which is orthogonal to §N and intersects §N along the hypersphere S. A hypersphere S in IRN (or in §N) can also be interpreted as the set of points XES. This allows us to introduce lower-dimensional spheres:
9. Foundations
344
• Spheres. A k-sphere is a (generic) intersection of N -k hyperspheres Si (i = 1, ... ,N - k).
They are further objects of Mobius geometry (in contrast to Lie geometry). This means that the class of k-spheres is preserved by Mobius transformations.
9.3.2. Projective model of Mobius geometry. In modelling elements of Mobius geometry (of either of the spaces jRN U{ oo} or §N), one can use the Lie-geometric description and just omit the eN+3-component. The resulting objects are points of the (N + I)-dimensional projective space IP(jRN+l,l) with the space of homogeneous coordinates jRN+l,l. The latter is the space spanned by N + 2 linearly independent vectors el, ... , eN+2 and equipped with the Minkowski scalar product
1, ~=~E{I, ... ,N+I}, { (ei,ej)= -1, z=]=N+2, 0,
i
=1=
j.
We continue to use notation (9.3) in the context of Mobius geometry. Elements of Mobius geometry of jRN are modelled in the space jRN+l,l of homogeneous coordinates as follows: • Point x E
jRN: X~
(9.16) • Infinity
~ = XEue = x
+ eo + Ix 12 eCX).
00:
00 = eCX).
(9.17)
• Hypersphere with center c E jRN and radius r
S=
(9.18)
SEue
= C + eo + (lcl 2 - r2)eCX).
• Hyperplane (v,x) = d with v E
P = PEue =
(9.19)
> 0:
V
§N-l
and d E R
+ 0 . eo + 2deCX).
In the projective space IP(jRN+l,l) these elements are represented by points which are equivalence classes of (9.16)-(9.19) with respect to the usual relation ~ rv TJ ¢} ~ = ATJ with A E jR* for ~,TJ E jRN+1,l. Fundamental features of these identifications are the following: (i) The infinity 00 can be considered as a limit of any sequence of x for x E jRN with Ixl - t 00. The points x E jRN U {oo} are in a one-toone correspondence with the points of the projectivized light cone IP(lLF+l,l), that is, with the straight line generators of
(9.20)
rrP+l,l
=
{~E jRN+l,l : (~,~)
= o}.
345
9.3. Mobius geometry
The points x E ]RN correspond to the points of lP'(IL,N+l,l) with a nonvanishing eo-component, while 00 corresponds to the only point of lP'(lL N +1,l) with the vanishing eo-component. Euclidean representatives (9.16) have an important property:
(9.21)
(Xl, X2) =
(ii) Hyperspheres
-!IX1 - x21 2,
s and hyperplanes p belong to lP'(]R:%t 1,1), where
]R:%t 1,1 = {~ E
(9.22)
VX1, X2 E ]RN.
]RN+1,1 :
(~,~)
> o}
is the set of space-like vectors of the Minkowski space ]RN+1,1. Hyperplanes can be interpreted as hyperspheres (of an infinite radius) through 00. (iii) Two hyperspheres 8 1 ,82 with centers Cl, C2 and radii r1, r2 intersect orthogonally if and only if
(9.23) which is equivalent to (Sl' S2) = O. Similarly, a hypersphere 8 intersects orthogonally with a hyperplane P if and only if its center lies in P:
(9.24)
(c, v) - d = 0, which is equivalent to (s,p)
=
O.
(iv) A point x can be considered as a limiting case of a hypersphere with radius r = o. An incidence relation x E 8 with a hypersphere 8 (resp. x E P with a hyperplane P) can be interpreted as a particular case of an orthogonal intersection of a sphere of radius r = 0 with 8 (resp. with P). We have: x E8
~
x E P
~
(x, s) = 0, (x, p) = o.
Switching from the Euclidean space ]RN to the sphere §N corresponds to a different choice of representatives for the points of lP'(]RN+1,1): • Point y E
§N:
f) = f)Sph = Y + eN+2·
(9.25) • Hypersphere 8
(9.26)
=
{y E §N: (s,y)
s = SSph = s + eN+2. • Great hypersphere 8
(9.27)
= 1}:
S=
=
{y E §N : (8, y)
SSph = 8
= O}:
+ O· eN+2.
Features of this choice of representatives:
9. Foundations
346
(i) In formulas (9.25), (9.26), Y and 8 are points ofIR N+ 1 with (y, y) = 1 and (8,8) > 1, which is equivalent to f) E ll..N+ u and s E IR~l~l,l, respectively. Also elements (9.27) (still defined up to a real factor) l TlllN+l,l . L)e I ongt 0 8 E !No. out A
(ii) Incidence relation: yES
(f), ,~)
{::;>
= O.
Indeed, the relation (f), s) = 0 for f) from (9.25) and for s from (9.26) is equivalent to (8, y) = 1. Similarly, the relation (f), s) = o for elements ,9 with vanishing eN+2-component, as in (9.27), is equivalent to (8, y) = 0, which characterizes great hyperspheres. To sum up: in the Minkowski space IRN+l,l of homogeneous coordinates, points and hyperspheres (different from hyperplanes) of the Euclidean space IRN find their place in the affine hyperplane (~, e oo ) = in particular,
-!;
IRN
(9 .28) 7ro
TlllN '" :!No. ~
~ «J!~
x
f---t
=
{~ E lLN+l,l : ~N+2
X =X Euc A
A
=.T
= x
-
~N+l
=
I},
+ eo + 1x 12 e=
1 2 + 2(lxl -
1) eN+l
1 2 + 2(lxl + 1) eN+2
N
E «J!o
(Euclidean metric d~r + ... + d~Jv being induced from the ambient IRN+1,1). The model «J!~ of the Euclidean space IRN can be viewed as a paraboloid in an (N + I)-dimensional affine subspace through eo spanned by el,"" eN, e oo · Similarly, points and hyperspheres of §N (different from great hyperspheres) find their place in the affine hyperplane (~,eN+2) = -1 of the Minkowski space IR N +l,I; in particular, §N ~
(9.29)
7rl :
§N
«J!'i = {~E lL N+ 1,1 3 y
~N+2 = I},
Y = YSph = Y + eN+2 E A
f---t
:
A
rn,N
'1,£1 .
The model «J!'i of the N-sphere §N can be viewed as a copy of §N in the (N + I)-dimensional affine subspace through eN+2 spanned by el,"" eN+l. Note that the correspondence between «J!~ and «J!'i along the straight line generators of lLN+1,1 induces the stereographic projection (j (compare (9.28) with (9.29) and with (9.14)). In particular, the generators of lLN+l,l through the points eo and e oo correspond to the zero and the point at infinity in IR N , and to the south pole Yo = -eN+l and the north pole Yoo = eN+l on §N, respectively. Turning to projective models of lower-dimensional spheres, recall that a hypersphere Sin IRN (or in §N) can also be interpreted as the set of points xES, and therefore it admits, along with the representation 5, the dual
347
9.3. Mobius geometry
Figure 9.5. Projective model of Mobius geometry.
representation as a transversal intersection of IP(JL N + 1,1) with the projective N-space IP(s.l), polar to the point S with respect to IP(JL N+l,1); here, of course, s.l = {x E ]R N +1, 1 : (s, X) = O}. This can be generalized to model lower-dimensional spheres .
• Spheres. A k-sphere is a (generic) intersection of N - k hyperspheres Si represented by Si E ]R~t1,1 (i = 1, ... , N - k). Such an intersection is generic if the (N - k )-dimensionallinear subspace of ]RN+1,1 spanned by Si is space-like: "
LI
' ' ) = span (S1,··., SN-k C
lIll N
+1,1 .
ll'\,.out
As a set of points, this k-sphere is represented as IP(JLN+l,l where
n
n I;.l),
N-k
I;.l=
Sf={XE]RN+I'l: (SI,X)="'=(SN-k,X)=O}
i=l
is a (k + 2)-dimensional linear subspace of ]RN+I,1 of signature (k+1,1). Through any k + 2 points Xl, ... , Xk+2 E ]RN in general position one can draw a unique k-sphere. It corresponds to the (k + 2)dimensional linear subspace I;.l = span(x1, ... ,Xk+2), of signature (k + 1, 1), with k + 2 linearly independent isotropic vectors Xl, ... , xk+ 2 E JL N+ I, 1. In the polar formulation, this ksphere corresponds to the (N - k)-dimensional space-like linear
9. Foundations
348
subspace k+2
2: =
nxf
=
{s E JRN+l,l:
(S,Xl) = ... = (S,Xk+2) =
o}.
i=1
To conclude, we mention that for hyperspheres s yet another choice of representatives in JRN+l,1 is sometimes used: one fixes the Lorentz norm of S. For any K, > 0, introduce the quadric (9.30)
lL~+I,1 = {~E JRN+l,l: (~,~) = K,2},
and choose the representative of a hypersphere in lL~+I,I: (9.31)
s = sMob =
~(s + eN+2) p
=
~ (c + eo + (lcl2 - r2)e oo ) E lL~+l,1. r
Actually, equation (9.31) contains two representatives of any hypersphere, corresponding to opposite values of p, resp. r, and therefore it represents oriented hyperspheres, each choice of the sign corresponding to one of the two possible orientations of a given hypersphere. Strictly speaking, this choice leads us outside of the projective model of Mobius geometry, and is a remainder of the Lie-geometric approach. We call p E JR (resp. r E JR) the oriented spherical (resp. Euclidean) radius of the hypersphere. For any two (oriented) hyperspheres 8 1 , 8 2 , the scalar product of their representatives SMob is a Mobius invariant: if K, = 1, then (SI,S2) = _1_((SI,S2) -1) = - 12 (ri PIP2 rl r 2
+ r~ -ICI -
c212)
is the cosine of the intersection angle of 8 1 , 8 2 , if they intersect, and the inversive distance between 8 1 , 8 2 , otherwise.
9.3.3. Mobius transformations. Mobius geometry is the study of properties of (nonoriented) hyperspheres invariant with respect to projective transformations of JPl(JR N+1,I) which map points to points, i.e., which leave JPl(lL N+ l ,1) invariant. Such transformations are called Mobius transformations.
Theorem 9.5. (Fundamental theorem of Mobius geometry) a) The group of Mobius transformations is isomorphic to O(N + 1,1)/ ~ O+(N + 1, 1), the group of Lorentz transformations of JR N +l ,1 preserving the time-like direction.
{±I}
b) Every conformal diffeomorphism of §N ~ JRN U {oo} is induced by the restriction to JPl(lLN+l,l) of a Mobius transformation.
The group O+(N + 1, 1) is generated by reflections, (9.32)
As : JRN+l,l -. JR N +l ,l,
As(x)
=
x _ 2((!, ~)) s. s,s
9.3. Mobius geometry
349
These reflections preserve the light cone lI.P+1,l and map straight line generators to straight line generators. Therefore, they induce some transformations on lP'(JLN+1,l) c::::: Ql'i', resp. on Ql~. The induced transformations on Ql~ c::::: ffi.N are obtained from (9.32) by direct computations with representatives (9.16) for points and representatives (9.18) for hyperspheres, and are given by (9.13) (inversion in the hypersphere S = {x E ffi.N : Ix-cl 2 = r2}); similarly, if s = P is the hyperplane (9.19), then the transformation induced on ffi.N by Ap is easily computed to be as in (9.12) (reflection in the hyperplane P = {x E ffi.N : (v, x) = d}). Similarly, the induced transformations on Ql'i' c::::: §N are obtained by a straightforward computation with representatives (9.25) for points and (9.26) for hyperspheres:
~
=
As(Y)
(
Y-8+
IY - 812 ) p2
and so the induced transformation on
§N
8
+ IY -p2 81
2
eN+2,
is given by (9.15).
Since (non)vanishing of the eo-component of a point in lP'(ffi.N+1,l) is not invariant under a general Mobius transformation, there is no distinction in Mobius geometry between hyperspheres and hyperplanes. The elements of the isotropy subgroup Oto(N + 1,1) of Lorentz transformations which fix e oo are generated by reflections in the hyperspheres (9.19), which induce reflections in the hyperplanes of ffi.N. Therefore, Oto(N + 1,1) is identified with E(N), the group of Euclidean motions of ffi.N. It is convenient to work with spinor representations of these groups. Recall that the Clifford algebra e£( N +1, 1) is an algebra over ffi. with generators el, ... ,eN+2 E ffi.N+1,l subject to the relation
e
e
This implies that 2 = -(e, e); therefore any vector E ffi.N+l,l \ JLN+l,l l = -e/(e,e). The multiplicative group generated by has an inverse the invertible vectors is called the Clifford group. We need its subgroup generated by the unit space-like vectors:
e-
9 = Pin+(N + 1,1) = {7/J = 6·· ·en: el = -I}, and its subgroup generated by the vectors orthogonal to e oo :
900 = Pin~(N + 1,1) = {7/J = 6 ... en :
e; = -1,
(ei, e oo ) = O}.
These groups act on ffi.N+1,l by twisted conjugations: A1/I('1]) = (_l)n7/J-l'1]7/J. In particular, for a vector with = -lone has:
e
Ae('1]) =
e
-e- l '1]e = e'1]e = '1] -
2(e, '1])e,
e.
which is the reflection in the hyperplane orthogonal to Thus, 9 is generated by reflections, while 900 is generated by reflections which fix eo, and
9. Foundations
350
therefore leave Ql{j invariant. Actually, 9 is a double cover ofO+(N +1,1) :::::: Mob(N), while 900 is a double cover of Ot,(N + 1,1) :::::: E(N). Orientation preserving transformations from 9, 900 form the subgroups J(
= Spin+(N + 1,1),
J(oo
=
Spin~(N + 1,1),
which are singled out by the condition that the number n of vectors ~i in the multiplicative representation of their elements 7jJ = 6 ... ~n is even. The Lie algebras of the Lie groups J( and J(oo consist of bivectors:
~ ~oo
spin(N+I,I)
= span{eiej: i,jE{O,I, ... ,N,oo}, i#j},
spinoo(N+I,I) = span{eiej: i,jE{I, ... ,N,oo},
i#j}.
9.4. Laguerre geometry Laguerre geometry is a sub geometry of Lie geometry, with hyperplanes distinguished among all hyperspheres, as the hyperspheres through 00. Thus, Laguerre geometry studies properties of hyperspheres invariant under the subgroup of Lie sphere transformations which preserve the set of hyperplanes. The following objects in JRN are elements of Laguerre geometry. • (Oriented) hyperspheres S = {x E JRN : Ix - cl 2 = r2} with centers c E JRN and signed radii r E JR, can be put into correspondence with (N + I)-tuples (c, r). • Points x E JRN are considered as hyperspheres of radius zero, and are put into correspondence with (N + I)-tuples (x,O). • (Oriented) hyperplanes P = {x E JRN : (v, x) = d}, with unit normals v E §N-l and d E JR, can be put into correspondence with (N + I)-tuples (v, d).
In the projective model of Lie geometry, hyperplanes are distinguished as elements of J!D(lL,N+l,2) with vanishing eo-component. Thus, Laguerre geometry studies the subgroup of Lie sphere transformations preserving the subset of J!D(lLN +1,2) with vanishing eo-component. There seems to exist no model of Laguerre geometry where hyperspheres and hyperplanes would be modelled as points of one and the same space. Depending on which of the two types of elements is modelled by points, one comes to the Blaschke cylinder model or to the cyclographic model of Laguerre geometry. We will use the first model, which has the advantage of a simpler description of the distinguished objects of Laguerre geometry, which are hyperplanes. The main advantage of the second model is a simpler description of the group of Laguerre transformations.
9.4. Laguerre geometry
351
The scene of both models consists of two (N + I)-dimensional projective spaces with dual spaces of homogeneous coordinates, ]RN,l,l and (]RN,l,l)*, which arise from ]RN+1.2 by "forgetting" the eo-, resp. eoo-components. Thus, ]RN,l,l is spanned by N + 2 linearly independent vectors el" .. , eN, eN+3, e oo , and is equipped with a degenerate bilinear form of signature (N, 1, 1) in which the above vectors are pairwise orthogonal, the first N being space-like: (ei' ei) = 1 for 1 :s: i :s: N, while the last two being time-like and isotropic, respectively: (eN+3,eN+3) = -1 and (eoo,e oo ) = 0, Similarly, (]RN,l,l)* is assumed to have an orthogonal basis consisting of el",· ,eN, eN+3, eo, again with an isotropic last vector: (eo, eo) = O. Note that one and the same symbol (', ,) is used to denote two degenerate bilinear forms in our two spaces, We will overload this symbol even more and use it also for the (nondegenerate) pairing between these two spaces, which is established by setting (eo, eexJ = - ~, in addition to the above relations. (Note that a degenerate bilinear form cannot be used to identify a vector space with its dual.) In both models mentioned above, • Hyperplane P = (11, d) is modelled as a point in the projective space JPl(]RN,l,l) with a representative
P= v + 2de oo + eN+3'
(9.33)
• Hypersphere S = (c, r) is modelled as a point in the projective space JPl( (]RN,l,1 )*) with a representative
(9.34)
Each of the models appears if we consider one of the spaces as a preferred (fundamental) space, and interpret the points of the second space as hyperplanes in the preferred space, In the Blaschke cylinder model, the preferred space is the space JPl(]RN,l,l) whose points model hyperplanes P C ]RN. A hypersphere S C ]RN is then modelled as a hyperplane {~E JPl(]RN,I,l) : (s,~) = O} in the space JPl(]RN,I,l). Basic features of this model are the following: (i) Oriented hyperplanes P C ]RN are in a one-to-one correspondence with the points p of the quadric JPl(lLN,I,I), where (9.35)
lLN,I,l = {~ E ]RN,I,I : (~,~) = O}.
(ii) Two oriented hyperplanes PI, P2 C ]RN are in oriented contact (parallel) if and only if their representatives PI, 'P2 differ by a vector parallel to e OC)) that is, if (PI, P2) = O. (iii) An oriented hypersphere S C ]RN is in oriented contact with an oriented hyperplane P C ]RN if and only if if (p, s) = O. Thus,
352
9. Foundations
a hypersphere 8 is interpreted as the set of all its tangent hyperplanes. The quadric lP'(IL N,l,l) is diffeomorphic to the Blaschke cylinder (9.36)
Two points of this cylinder represent parallel hyperplanes if they lie on one straight line generator of Z parallel to its axis. In the ambient space ]RN+l of the Blaschke cylinder, oriented hyperspheres 8 C ]RN are in a one-to-one correspondence with the hyperplanes nonparallel to the axis of Z:
8
(9.37)
rv
{(v,d) E
]RN+l:
(c,v) - d - r =
a}.
An intersection of such a hyperplane with Z consists of points in Z which represent tangent hyperplanes to 8 C ]RN, as follows from (9.11). In the cyclographic model, the preferred space is the space of hyper-
spheres (]RN,l,l)*, so hyperspheres 8 C ]RN are modelled as points s E lP'( (]RN,l,l )*), while hyperplanes P C ]RN are modelled as hyperplanes {e : (p, e) = o} c lP'( (]RN,l,l )*). Thus, a hyperplane P is interpreted as the set of hyperspheres 8 which are in oriented contact with P. Basic features of this model are the following: (i) The set of oriented hyperspheres 8 C ]RN is in a one-to-one correspondence with the points
a
(9.38)
=
(c, r)
of the Minkowski space ]RN,l spanned by the vectors eI, ... ,eN, eN+3. This space has interpretation as an affine part oflP'((]RN,l,l )*).
(ii) Oriented hyperplanes P
C
]RN can be modelled as hyperplanes in
]RN,l:
(9.39)
7r
= {(c, r)
E ]RN,l :
((v, 1), (c, r)) = (v, c) - r = d}.
Thus, oriented hyperplanes P E ]RN are in a one-to-one correspondence with the hyperplanes 7r C ]RN,l which make angle 7r / 4 with the subspace]RN = {(x,D)} C ]RN,l.
(iii) An oriented hypersphere 8
C ]RN is in oriented contact with an oriented hyperplane P C ]RN if and only if a E 7r.
(iv) Two oriented hyperspheres 8 1 ,82 C ]RN are in oriented contact if and only if their representatives in the Minkowski space aI, a2 E ]RN,l differ by an isotropic vector: ia1 - a2i = 0. In the cyclographic model, the group of Laguerre transformations admits
a beautiful description:
9.5. Plucker line geometry
353
Theorem 9.6. (Fundamental theorem of Laguerre geometry) The group of Laguerre transformations is isomorphic to the group of affine transformations of ffi,.N,l: y f---4 AAy + b with A E O(N, 1), A > 0, and b E jRN.l.
9.5. Pliicker line geometry In this section we denote the homogeneous coordinates of a point x E ffi,.1P'3 by x = (xO,x 1 ,x2 ,x3 ) E ffi,.4. For the sake of notational convenience, we abbreviate V = JR.4. In the standard way, projective subspaces of ffi,.1P'3 are projectivizations of vector subspaces of V. In particular, let x, y E JR.1P'3 be any two different points, and let x, fj E V be their arbitrary representatives in the space of homogeneous coordinates. Then the line g = (xy) C ffi,.1P'3 is the projectivization of the two-dimensional vector subspace span(x, fj) C V. After H. Grassmann and J. Plucker, the latter subspace can be identified with (a projectivization of) the decomposable bivector (9.40) We choose a basis of A2V to consist of ei 1\ ej with 0 ::; i < j < 3. A coordinate representation of the bivector (9.40) in this basis is (9.41)
9=
2::gij
ei 1\ ej,
(ij)
The numbers (gOl, g02, g03, g12 , g13, g23) are called PlUcker coordinates of the line g. They are defined projectively (up to a common factor). Indeed, changing the choice of the two points defining 9 from x, y to x, y with the homogeneous coordinates xj = ax j + byj, yj = cx j + dyj, ad - bc =1= 0, would lead to a simultaneous multiplication of all gij by a common factor: gij
=
(ad - bc)gij.
Not every bivector represents a line in ffi,.1P'3, since not every bivector is decomposable, as in (9.40). An obvious necessary condition for a non-zero 9 E A2 V to be decomposable is (9.42)
9 1\ 9 = O.
It can be shown that this condition is also sufficient. In Plucker coordinates, this condition can be written as
(9.43) Summarizing, we have the following description of £}, the set of lines in JR.1P'3, within Plucker line geometry. The six-dimensional vector space A2V
354
9. Foundations
with the basis e.i I\ek is supplied with a nondegenerate scalar product defined by the following list of nonvanishing scalar products of the basis vectors: (eo 1\ e1, e2 1\ e3)
=
-(eo 1\ e2, el 1\ e3)
=
(eo 1\ e3, e1 1\ e2)
=
l.
It is not difficult to verify that the signature of this scalar product is (3,3), so that we can write A 2 V ':::' lR 3 ,3. Denote
(9.44) The points of the PlUcker' quadric lP'(IL3,3) are in a one-to-one correspondence with elements of I.., 3 . A fundamental feature of this model is the following: • Two lines g, h in lRlP'3 intersect if and only if their representatives in A2V are polar to one another: (9.45)
(9, h) =
gOlh 23 - g02h 13
+ g03h l2 + g 23 h Ol
_ g 13 h 02
+ g 12 ho:3 = O.
In this case the line P C lP'(A2V) through [9] and [h] is isotropic:
P c lP'(IL 3,3). To prove this, note that if the lines g, h intersect at the point z, then 9 = x 1\ Z and 11 = Y 1\ z, and then 91\12 = O. Conversely, if the lines g, h do not intersect, then their lifts to V span the whole of ~/, and so 91\11, =1= o. It remains to observe that 91\ h = (9, h.) eo 1\ e1 1\ e2 1\ e3. Next, we turn to important linear subsets of the Plucker quadric. • Any isotropic line P C lP'(IL3.3) corresponds to a one-parameter family of lines in lRlP'3 through a common point, which lie in one plane. Such a family of lines is naturally interpreted as a contact element (a point and a plane through this point) within the line geometry. • Other than in Lie geometry, in the present case of signature (3,3) there exist also isotropic planes, which are projectivizatiolls of 3dimensional vector subspaces of A2V that belong to IL 3.:3. There are two sorts of isotropic planes in the Pliicker quadric lP'(IL·3,:3). An isotropic plane can represent: 0:) a two-parameter family of all lines in lRlP'3 through some common point; such a family is naturally identified with that common point; (3) a two-parameter family of all lines in some plane in lRlP'3; such a family is naturally identified with that common plane. To see why the latter statement holds, consider three noncollinear points in the isotropic plane. Their pairwise connecting lines are all isotropic. Therefore these three points represent three pairwise intersecting lines in lRlP'3. If all three are concurrent, then we are in the situation 0:). Otherwise they lie in a plane in lRlP'3, and we are in the situation ;3).
355
9.5. PlUcker line geometry
Projective transformations of W'(lR3,3) which leave the Plucker quadric W'(1L 3 ,3) invariant can be distinguished depending on their action on the two types of isotropic planes.
Theorem 9.7. (Fundamental theorem of Plucker line geometry)
a) The group of projective transformations of lRW'3 is isomorphic to the subgroup of 0(3, 3)/(±I) consisting of transformations which preserve the types 0;) and (3) of the three-dimensional vector subspaces in 1L3 ,3. b) The group of correlative transformations of lRW'3 is isomorphic to the subgroup of 0(3, 3)/(±I) consisting of transformations which interchange the types 0;) and (3) of the three-dimensional vector subspaces in 1L3 ,3. Next, we discuss planar families of lines. Such a family of lines is represented by a conic section W'CEnIL3,3), where L; stands for a three-dimensional vector subspace of lR3 ,3 such that the restriction (-, .) I~ is nondegenerate. It is not difficult to realize that four pairwise nonintersecting (skew) lines in lRW'3 belong to a planar family (have linearly dependent representatives in 1L3 ,3) if and only if they belong to a regulus (one family of generators of a ruled quadric in lRW'3, i.e., of a one-sheet hyperboloid or of a hyperbolic paraboloid). The complementary regulus is represented by the dual planar family of lines lJD(L;..l n 1L3,3). Finally, we briefly mention the duality in Plucker line geometry. One can describe any projective subspace W'(L;) c JRW'3 through its dual subspace W'(L;..l) c (lRW'3) *, where L;..l C V* is the annihilator of the vector subspace L; c V. As a set of points, W'(L;) is the intersection of planes represented by lJD(L;..l). Thus, a plane U C JRW'3 can be described through an element of lJD(L;..l) C (JRlJD3)* with homogeneous coordinates fl = (UO,UI,U2,U3) = L:?=o ulei E V*. As a set of points, this plane consists of x E JRlJD3 with homogeneous coordinates x = (x O, xl, x 2, x 3) E V satisfying 3
(9.46)
LU1X1 =0. [=0
This description of U is dual to the description as the projectiivization of the three-dimensional vector subspace span(x, fj, z) C V, where x, fj, z are homogeneous coordinates of any three noncollinear points x, Y, z E u. In the spirit of the Grassmann-Plucker approach, the latter vector subspace can be represented by (a projectivization of) the decomposable three-vector it = x 1\ fj 1\ z E A3V. In the basis of A3V consisting of ei 1\ ej 1\ ek, o ::; i < j < k ::; 3, one has:
it =
L (ijk)
) ei u t"k
1\ ej 1\ ek ,
u ijk =
Xi
x")
Xk
Yi
Yj
Yk
Zi
-)
'Y"
Zk
9. Foundations
356
It is easy to see that the homogeneous coordinates that Uo
=
u 123 ,
U1
=
_u0 23 ,
U2
= u0 13 ,
Ul U3
can be normalized so
=
_u0 12
Similarly, in the dual description, any line 9 C lR.1fD3 can be viewed as an intersection of two planes u, v C lR.1fD3, and thus can be described through span( il, v) E V*, which, in turn, can be represented by (a projectivization of) the bivector (9.47) In coordinates: (9.48)
9=
il /\
v=
L gij e; /\ ej, (ij)
The sextuple of numbers (g01, g02, g03, g12, g13, g23) is called dual PlUcker coordinates of the line g. Remarkably, this new set of coordinates is related to the previously introduced Pliicker coordinates in a fairly simple way: if u, v are any two planes in lR.1fD3 intersecting along the line g, then their homogeneous coordinates Ui, Vi can be so normalized that the dual Plucker coordinates (9.48) of the line 9 = un v coincide, after a suitable reordering, with its coordinates (9.41): (9.49)
gOl g23
= g23, = gOl,
g02 g13
= =
_g13,
g03
_g02,
g12
= g12; = g03.
To see this, take 9 = (xy) = un v and choose points p E u, q E v so that p (j. v and q (j. u. We can normalize homogeneous coordinates of the planes u, v so that
(~~ ~i ~: ~:) (~: ~:) qO
q1
q2
q3
11,3
=
v3
(H) . 1 0
Now (9.49) follows from a well-known generalization of the Cramer rule, which says that the 2 x 2 determinants gij = xiyj - xjyi are proportional to the 4 x 4 determinants obtained from the matrix of the latter linear system by replacing the i-th and j-th columns by the columns on the right-hand side of the system. To conclude, we mention a couple of useful relations for the usual and dual Plucker coordinates of lines. They follow directly from definitions. • Homogeneous coordinates of the plane point p (j. 9 are given by
11,
through a line 9 and a
3
(9.50)
Uj
= L9jkpk k=O
(j = 0,1,2,3);
9.6. Incidence theorems
357
homogeneous coordinates of the intersection point x of a line 9 with a plane v not containing 9 are given by 3
x j = L9jkvk
(9.51)
(j=O,1,2,3).
k=O
• A point x belongs to a line 9 if and only if 3
(9.52)
L9jk Xk
=
k=O
°
(j = 0, 1, 2, 3);
a line 9 lies in a plane u if and only if 3 "~gJ"k Uk =
(9.53)
k=O
°
(j=O,1,2,3).
By the way, the last statement allows us to give a simple argument for the claim that any 9 E lL3,3 corresponds to a line in IRlP3 . Indeed, (9,9) = and 9 i= is equivalent to the fact that the rank of system (9.53) is equal to 2, and so the solution of this system delivers two different planes. They intersect along the line we are looking for.
°
°
9.6. Incidence theorems This section contains a collection of classical incidence theorems which lie in the basis of discrete differential geometry. We will often use the cross-ratio of four collinear points a, b, c, d, defined as (9.54)
l (a, b) l (c, d) q(a, b, c, d) = l(b, c) . l(d, a)'
and the fact that the cross-ratio is invariant under projective transformations. 9.6.1. Menelaus' and Ceva's theorems. Theorem 9.8. (Menelaus' theorem) Consider a triangle L::.(AIA2A3) in the plane. Let P12, P23 , P31 be some points on the side lines (AIA2), (A 2A 3), (A3Ad, respectively, different from the vertices Ai of the triangle. These three points are collinear if and only if
(9.55) Theorem 9.9. (Ceva's theorem) Consider a triangle L::.(AIA2A3) in the plane. Let P12 , P23 , P31 be some points on the side lines (A 1A 2), (A 2A3), (A3Ad, respectively, different from the vertices Ai of the triangle. The three
9. Foundations
358
Figure 9.6. Menelaus'theorem.
lines (A 1P23), (A 2P3t) , (A 3P 12) have a common intersection point if and only if (9.56)
Figure 9.7. Ceva's theorem.
Both Menelaus' and Ceva's theorems have a similar flavor: their hypotheses are of a seemingly affine-geometric nature (the left-hand sides of equations (9.55), (9.56) are expressed in terms of quotients of directed lengths), while their conclusions are projectively invariant. Actually, one can show that the numeric value of the cyclic product of the quotients of directed lengths on the left-hand sides of (9.55), (9.56) is itself a projectively invariant quantity. This is a consequence of the following theorem. Theorem 9.10. (Projective invariance of a cyclic product of directed lengths ratios for a triangle) Consider a triangle 6(AIA2A3) in
9.6. Incidence theorems 359
the piane, and let P 12 , P", P" be Some points on the lines (A,A,), (A,A ), 3 (A"A,J, respectively, different fmm the vertices A; of the bia"qle, a) Drno9.8. te byThen Q12 the intersection point of the line (A, A,) with (P P ); 8('e Figurr: 23 31
() 9,57
"~_q (A
I(A I , P12) I(A 2, P2.3 ) I(A:l , P3J) I(P12 , A 2) I(P2:l, A 3) 1(P..11, Ad
A Q )
P 1,
[2,
"
.
~
12,
b) Set 0 (A,P",) n (A,p,,) and denote by Ii" the intersection point of the line (A,A 2) with (A:30); 8ee Figure 9.9. Then
() 9,58
I(A" P,,) i(A" P2J ) I(A" P31) I(P", A,) - I(P", A;} - I(P", AJ)
~ q(A
PAR ) 1,
12,
"
Figure 9.8. A projectively invariant meaning of the cyclic product from J\.Iendaus' theorem.
Figuretheorem. 9.9. A projectiwly invariant meaning of the cyclic product frorn Ceva's
-12-
360
9. Foundations
Proof. Clearly, this theorem yields Menelaus' and Ceva's: the cross-ratios on the right-hand sides of equations (9.57), (9.58) are equal to 1 if and only if Ql2 = P 12 , resp. Rl2 = P 12 . For a proof, note that, since both sides of (9.57), (9.58) are invariant under affine transformations, it is enough to consider Al = (0,0), A2 = (1,0), A3 = (0,1), and then P l2 = (XI'O)' P 23 = (1- X2,X2), P31 = (0,1- X3) with some XI,X2,X3 E IR \ {O, I}. Then a straightforward computation confirms both claims of the theorem. 0 9.6.2. Generalized Menelaus' theorem. Upon using the results of Theorem 9.10 to "cut off vertices", one can prove the following result. Theorem 9.11. (Projective invariance of cyclic products of directed lengths ratios) Let AI, A 2, ... , An be n 2': 3 points in IRm such that no three consecutive points in cyclic order are collinear. Let P 12 , P 23 , ... , P nl be points on the lines (AIA2)' (A2A3),"" (AnAl), respectively. Then the product of ratios of directed lengths
rr n
l(Ai, Pi,i+l)
i=l
l(Pi,HI, AHI )
is invariant under projective transformations.
The geometric meaning of the situation where the cyclic product in Theorem 9.11 takes a special value (_l)n is given by the following result. Theorem 9.12. (Generalized Menelaus' theorem) Let AI"'" An be n points in general position in IRn - l , so the affine space through the points Ai is (n - I)-dimensional. Let Pi,i+l be some points on the lines (AiAi+d different from A, AHI (indices are taken modulo n). The n points Pi,i+l lie in an (n - 2)-dimensional affine subspace if and only if the following relation for the quotients of the directed lengths holds:
IT
l(Ai,Pi,i+l) .z=l l(Pi ' HI, AHd
= (-It.
Proof. The points Pi ,Hl lie in an (n - 2)-dimensional affine subspace if there is a nontrivial linear dependence n
L
n
J1i P i,i+1 = 0
with
i=l
L J1i = O. i=l
Substituting Pi,i+l = (1 - ~i)A + ~iAHI' and taking into account the general position condition, which can be read as linear independence of the --+ vectors AlAi, we come to a homogeneous system of n linear equations for n coefficients J1i: ~iJ1i
+ (1 -
~i+I)J1HI = 0,
i = 1, ... , n
361
9.6. Incidence theorems
(where indices are understood modulo n). Clearly it admits a nontrivial solution if and only if
n~ rr 1 c. i=l
-