Combinatorial Games MSRI Publications Volume 29, 1995

Multilinear Algebra and Chess Endgames
LEWIS STILLER
of multilinear algebraic formalism for high-performance computing. (2) To describe an application of this formalism in the analysis of chess endgames, and results obtained thereby that would have been impossible to compute using earlier techniques, including a win requiring a record 243 moves. (3) To contribute to the study of the history of chess endgames, by focusing on the work of Friedrich Amelung (in particular his apparently lost analysis of certain six-piece endgames) and that of Theodor Molien, one of the founders of modern group representation theory and the rst person to have systematically numerically analyzed a pawnless endgame.

Abstract. This article has three chief aims: (1) To show the wide utility

1. Introduction
Parallel and vector architectures can achieve high peak bandwidth, but it can be di cult for the programmer to design algorithms that exploit this bandwidth e ciently. Application performance can depend heavily on unique architecture features that complicate the design of portable code Szymanski et al. 1994; Stone 1993]. The work reported here is part of a project to explore the extent to which the techniques of multilinear algebra can be used to simplify the design of highperformance parallel and vector algorithms Johnson et al. 1991]. The approach is this: De ne a set of xed, structured matrices that encode architectural primitives of the machine, in the sense that left-multiplication of a vector by this matrix is e cient on the target architecture. Formulate the application problem as a matrix multiplication. Factor the matrix corresponding to the application in terms of the xed matrices using addition, tensor product, and matrix multiplication as combining operators. Generate code from the matrix factorization.
Supported by U.S. Army Grant DAAL03{92{G{0345.

1

2

LEWIS STILLER

This approach has been used in signal processing algorithms Granata and Tolimieri 1991; Granata et al. 1991; Johnson et al. 1990; Tolimieri et al. 1989; Tolimieri et al. 1993]. The success of that work motivates the present attempt to generalize the domain of application of the multilinear-algebraic approach to parallel programming Stiller 1992b]. I have used the methodology presented in this paper in several domains, including parallel N -body codes Stiller 1994], Fortran 90 communication intrinsic functions Stiller 1992a], and statistical computations Garc a and Stiller 1993]. In each case, signi cant speedup over the best previous known algorithms was attained. On the other hand, it is clear that this methodology is intended to be applicable only to a narrow class of domains: those characterized by regular and oblivious memory access patterns. For example, parallel alpha-beta algorithms cannot be formulated within this paradigm. This paper describes the application of the multilinear algebraic methodology to the domain of chess endgames. Dynamic programming was used to embed the state space in the architecture. By successively unmoving pieces from the set of mating positions, the set of positions from which White could win can be generated. This application presents an interesting challenge to the multilinearalgebraic parallel-program design methodology: The formalism for the existing multilinear algebra approach had been developed to exploit parallelization of linear transformations over a module, and had to be generalized to work over Boolean algebras. The symmetry under a noncommutative crystallographic group had to be exploited without sacri cing parallelizability. The state-space size of 7.7 giganodes was near the maximum that the target architecture could store in RAM. Two main results are reported here: Table 1 on page 18 gives equations de ning the dynamic programming solution to chess endgames. Using the techniques described in this paper, the factorizations can be modi ed to produce e cient code for most current parallel and vector architectures. Table 2 on page 25 presents a statistical summary of the state space of several six-piece chess endgames. This table could not have been generated in a practicable amount of time using previous techniques. Section 2 below provides the background of the chess endgame problem. A survey of some human analysis of chess endgames is given, followed by a survey of previous work in the area of computer endgame analysis. Readers interested only in chess and not in the mathematical and programming aspects of this work can skip from this section to Section 8. Section 3 introduces some basic concepts of parallel processing.

MULTILINEAR ALGEBRA AND CHESS ENDGAMES

3

Section 4 describes previous work in the area of tensor product formalism for signal-processing computations. Section 5 develops a generalized version of the formalism of Section 4, and describes the chess endgame algorithm in terms of this formalism. Section 6 presents equations de ning the dynamic programming solution to chess endgames (Table 1). Section 6.2 describes how these equations are modied to exploit symmetry. The derivation of crystallographic FFTs is used as a motivating example in the derivation of symmetry-invariant equations. Section 7 discusses very brie y some possible re nements to the algorithm. Section 8 presents some of the chess results discovered by the program, including the best play from a position that requires 243 moves to win. Section 9 gives some implementation details and discusses run times.

2. Endgame Analysis by Humans and Computers
We start with a brief historical survey of the analysis of endgames, particularly those containing at most six pieces. In listing the pieces of an endgame, the order will be White King, other White pieces, Black King, other Black pieces. Thus, KRKN is the same as KRkn, and comprises the endgame of White King and White Rook against Black King and Black Knight. `Adl , plates 105 and 112]. However, the rules were slightly di erent in those days, as stalemate was not necessarily considered a draw. The oldest extant collection of compositions, including endgames, is the Alfonso manuscript, ca. 1250, which seems to indicate some interest during that time in endgame study Perez 1929, pp. 111{112]. Modern chess is generally considered to have begun roughly with the publication of Luis Ramirez de Lucena's Repeticion de amores y arte de ajedrez Lucena 1497?]. Ruy Lopez de Sigura's book Lopez 1561] brie y discusses endgame theory, but its main impact on this work would be the introduction of the controversial fty-move rule pp. 55{56], under which a game that contains fty consecutive moves for each side without the move of a pawn or a capture could be declared drawn Roycroft 1984]. Pietro Carrera's Il gioco de gli scacchi Carrera 1617] discussed a number of fundamental endgames such as KQKBB, and certain six-piece endgames such as KRRKRN and KRRKRB Book 3, pp. 176{178]. A number of other authors began developing the modern theory of endgames Greco 1624; Stamma 1737; Philidor 1749?]. Giovanni Lolli's monumental Osservazioni teorico-pratiche sopra il giuoco degli scacchi Lolli 1763] would be one of the most signi cant advances in endgame theory for the next ninety years Roycroft 1972]. Lolli

2.1. Human analysis. Endgame analysis appears to date from at least the ninth century, with al-`Adl 's analysis of positions from KRKN and KRNKR

4

LEWIS STILLER

analyzed the endgame KQKBB, and he agreed with the earlier conclusion of Salvio 1634] that the endgame was a general draw. This assessment would stand substantially unchanged until Kenneth Thompson's computer analysis demonstrated a surprising 71-move win Thompson 1986]. Notwithstanding this error, Lolli did discover the unique KQKBB position in which White to play draws but Black to play loses pp. 431{432]. Bernhard Horwitz and Josef Kling's Chess Studies Kling and Horwitz 1851, pp. 62{66] contained a number of in uential endgame studies, although their analysis of KBBKN was questioned in Roycroft 1972, p. 207]. The Horwitz and Kling assessment was de nitively shown to be incorrect by two independent computer analyses Thompson 1983; Roycroft 1984]. Alfred Crosskill 1864] gave an analysis of KRBKR in which he claimed that more than fty moves were required for a win; this was con rmed by computer analysis of Thompson. The Crosskill analysis was the culmination of a tradition of analysis of KRBKR beginning at least from the time of Philidor Philidor 1749?, pp. 165{169]. Henri Rinck and Aleksei Troitzky were among the most in uential endgame composers of the next two generations. Troitzky is well-known for his analysis of KNNKP, in which he demonstrated that more than fty moves are at times required for a win Troitski 1934]. Rinck was a specialist in pawnless endgames, composing more than 500 such studies Rinck 1950; Rinck and Malpas 1947], including some with six pieces. Troitzky summarized previous work in the area of KNNKP, beginning with a problem in KNKP from the thirteenth-century Latin manuscript Bonus Socius Anonymous], and reserved particular praise for the systematic analysis of this endgame in an eighteenthth-century manuscript Chapais 1780]. An early version of the program reported in the present paper resulted in the rst published solution for this entire endgame Stiller 1989]. The twentieth century saw the formal codi cation of endgame theory in works such as Berger 1890; Cheron 1952; Euwe 1940; Fine 1941; Averbakh 1982], and many others. Some work focusing particularly on pawnless six-piece endings has also appeared, for example, Roycroft 1967; Kopnin 1983; Berger 1922]. Currently the Informator Encyclopedia of Chess Endings series Matanovic 1985], which now uses some of Thompson's computer analysis, is a standard reference. The books Nunn 1992; Nunn 1994] are based on that work. Additional historical information can be found in Hooper and Whyld 1992; Golombek 1976; Murray 1913; Roycroft 1972].

2.2. Friedrich Amelung and Theodor Molien. Friedrich Ludwig Amelung

(born March 11, 1842; died March 9, 1909) was a Latvian chess player and author who edited the chess column of the Riga newspaper Duna-Zeitung. He studied philosophy and chemistry at the University of Dorpat from 1862 to 1879, and later became a private teacher and director of a mirror factory Lenz 1970, p. 11; Duna-Zeitung 1909a; 1909b]. He published a number of endgame studies and

|l bil govorit~ on pozdnee. and began a systematic study of pawnless endgames. Th. Amelung variously used Mollin. to my knowledge. although it appeared in a widely read and in uential publication. However. is Amelung's comment that an even earlier. see also Bashmakova 1991].MULTILINEAR ALGEBRA AND CHESS ENDGAMES 5 analyses of endgames. 1941. . This table listed the approximate number of positions in KRKN from which White could win and draw in 2{5 moves. Greek. Dutch. requiring up to 46 moves to win. including Hebrew. his main interest to our work actually lies in two major projects: an analysis of the four-piece endgame KRKN Amelung 1900]. of course. Latin. the Deutsche Schachzeitung. 1908a. He studied celestial mechanics at Dorpat University (1880{1883) and took courses from Felix Klein in Leipzig (1883{1885). was a philologist and teacher. Stiller 1989]. der Mathematiker von Fach ist". 1861 in Riga. 1 <Proqita$ ite sto romanov na kakom-libo zyke. His doctoral dissertation. including the decomposability of group algebras into direct sums of matrix algebras. 1908b. Mollien and Molien. however. Italian. which was translated for me by Boris Statnikov. of a pawnless endgame or. or Fyodor Eduardovich Molin. \and you will know that language. English. Even more intriguing. and Norwegian. 265{266]. For example. and 20{30 moves. Portuguese. and his studies of certain pawnless six-piece endgames Amelung 1902. that is. 10{20 moves. Spanish. Amelung's 1900 analysis of KRKN was signi cant because it contained the rst histogram. he explored the endgame KQKRN in detail Amelung 1901]. p.) His father. Mollien. in Stiller 1989]. Mollien. French. this endgame was shown to have unexpected depth. pp. for that matter. of course. (The biographical information here has been taken from Kanunov 1983]. we write Theodor Molien. and was due to \Dr. numerical analysis containing the number of win-in-k moves for each k of a four-piece chess endgame was known. combinations that were not exhaustively analyzed until the 1980s Thompson 1986. exact."1 He studied celestial mechanics at Dorpat University (1880{1883) and also took courses from Felix Klein in Leipzig (1883{1885). of any endgame Amelung 1900. and died on December 25. Kanunov gives his name Fedor duardoviq Molin.|i Vy budete znat~ tot zyk. Theodor Molien was born on September 10.> Kanunov 1983. and Theodor eventually became uent in a number of languages. published in Mathematische Annalen Molien 1893]. The existence of this early analysis does not appear to have been known to contemporary workers. 1908c]. Eduard." Molien liked to say. 9]. \Read a hundred novels in a language. to the professor Th. proved a number of the fundamental structure theorems of group representation theory. 5{10 moves. Swedish. Such tables have been a mainstay of computer-age endgame analysis. He also published an article Amelung 1893] on KBNKN and KBBKN. as well as German and Russian. in conformity with his publications.

for example. 1898]. 35{36]. After studying medieval mathematical manuscripts at the Vatican Library in 1899. in which he investigated non-commutative multiplication and obtained important general results of which the properties of the group determinant are special cases. at least in one case. 1974] on the history of group representation theory. the classic text Wussing 1969] barely mentions him. and more recent historical appraisals gives him due credit van der Waerden 1985. Referring to Molien 1893]. where he was cut o from the mathematical mainstream and became embroiled in obscure administrative struggles (he was. Er hat im 41.) Despite these results. it is quoted here with the permission of Springer-Verlag. von denen die Eigenschaften der Gruppendeterminant specielle Falle sind. because the committee considered his work too theoretical and without practical applications Kanunov 1983. 1995)." . 1972. 1897b. Theodor Molien in Dorpat. despite their importance. He has published : : : a very beautiful but hard-to-read work \On systems of higher complex numbers". aber schwer zu lesende Arbeit `Ueber Systeme hoherer complexer Zahlen' vero entlicht. he was one of the strongest chess players in Dorpat and was particularly known for his blindfold play (Ken Whyld. and several of his games were published in 2 \Sie werden bemerkt haben. pp. 206{209. With the publication of Hawkins' series of articles 1971. May 6. Frobenius wrote to Dedekind: You will have noticed that a young mathematician. They anticipated Frobenius' classic paper on the determinant of a group-circulant matrix Frobenius 1896]. 1897a. and despite Frobenius' support. pp. 237{238]. were obscure and di cult to understand. personal communication.2 (This letter of February 24.6 LEWIS STILLER Molien's early papers on group representation theory 1893. Molien was rejected from a number of Russian academic positions. a fact that Frobenius readily admitted Hawkins 1974]. 1898 was kindly supplied by Thomas Hawkins. He was president of the Dorpat chess club. before Molien moved to Tomsk. Not mentioned in Kanunov's biography is that. has independently of me considered the group determinant. As a consequence. in a transcription made by Walter Kaufmann-Buhler. I have bene ted from Hawkins' translation in supplying mine. he accepted a post at the Tomsk Technological Institute in Siberia. da sich ein jungerer Mathematiker Theodor Molien in Dorpat unabhangig von mir mit der Gruppendeterminante beschaftigt hat. partly because of the Czarist politics of the time (according to Kanunov) and. although he had tremendous di culty understanding Molien's work (letter to Alfred Knezer. 1898). in fact. His remaining mathematical work had little in uence and he spent most of his time teaching. the signi cance of his work became better-known. Molien's work was unknown or underestimated in the West for a long while. brie y red). worin er die nicht commutative Multiplication untersucht hat und wichtige allgemeine Resultate erhalten hat. Bande der Mathematischen Annalen eine sehr schone.

In 1898 Molien published four chess studies Molien 1898a] based on his research into the endgame KRKB Amelung 1900. Diaconis and Rockmore 1990. We now continue with our discussion of chess endgame history proper. some mathematical a nity between these areas as well. 223{233]. although his work in both areas was mostly ignored for a long time. but I have not been able to locate a publication of his complete results. 5]. pp. A solution was published the next year. Amelung's interest in this endgame was so great that he held a contest in Duna-Zeitung for the best solution to a particular example of this endgame Amelung 1908c]. p. and in 1908 he published a short article on the topic in Fur Haus und Familie. 5. a biweekly supplement to the Riga newspaper Duna-Zeitung. Rook and minor piece (N or B) against King and two minor pieces Amelung 1902]. Volume 5. despite the historical signi cance of such a document. the chess move operators can be encoded by a group-equivariant matrix.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 7 a Latvian chess journal. Karpovsky 1977]. analyzing the endings of King. It is an interesting coincidence that within a span of a few years Molien performed groundbreaking work in the two apparently unrelated areas of group representation theory and quantitative chess endgame analysis. perhaps. Consequently. and his work seemed to have been largely forgotten. in general. of which he was the chess editor Amelung 1908a]. rapid multiplication of a group-equivariant matrix by a vector. There is. particularly Amelung's work on pawnless endgames. in 1902 Amelung published a three-part series in Deutsche Schachzeitung. Molien 1901]. Schachbl. 8]. Partly in response to the rst edition of the in uential manual of endings Berger 1890. 1909. by Amelung Balt. perhaps the premier chess journal of its time. pp. massively parallel implementations are described in Stiller 1995. Baltische Schachblatter. Diaconis 1988. edited. since. These numerical studies are alluded to several times in the chess journals of the time Amelung 1898. but it was criticized by the mathematician and chess champion Machgielis (Max) Euwe in his titanic study of pawnless endgames Euwe 1940. 1902. one of the games he lost fetched a \bestgame" prize in the main tournament of the Jurjewer chess club in 1894 Molien 1900]. 50{53]: . succeeding commentators dismissed many of his more extreme claims. Amelung indicated that the case KRBKNN was particularly interesting. as we shall see. Chapter 7]. and representing a continuation of Amelung's earlier work with Molien on the endgame KRKN Amelung 1900]. pp. p. for a time. 1893. It is discussed in Berger 1922. but Amelung died that year and was unable to continue or popularize his research. relies on the algebra-isomorphism between a group algebra and a direct sum of matrix algebras rst noted by Molien Clausen and Baum 1993. p. of which his work on KRBKNN deserves special mention. 167{169]. 265.

even by the game of chess. Additional theoretical work was done by John von Neumann and Oskar Morgenstern 1944. pp.8 LEWIS STILLER \This endgame KRBKNN] o ers the stronger side excellent winning chances. Torres-Quevedo's automaton could move its own pieces. like that of Huberman 1968]. In addition to the remark about Molien. 467] Automated endgame play appears to date from the construction by Leonardo Torres-Quevedo of an automaton to play KRK endings. F." (Translation from the Dutch by Peter Jansen. Simons 1986]. in which the value of each node of the state-space is computed by backing up the game-theoretic values of the leaves. we are concerned with exhaustive analysis of endgames. but this assessment seems to be untrue. maar deze opvatting schijnt ojuist te zijn. Although some have dated computer chess from Charles Babbage's brief discussion of automated game-playing in 1864. Amelung ging zelfs zoo ver. and in substantiation of his opinions presents his automatic chess-playing machine. 2. Allowing one hundred moves on each side for the longest game at chess. pp. Although some sources give 1890 as the date in which this machine was designed. which de nes the eld of retrograde endgame analysis. 298]. It used a rule-based approach Scienti c American 1915. I was not able to locate any references to it in domestic catalogues and indices. p. Amelung went so far as to say that the defense was hopeless. and is not included in the comprehensive bibliography by van den 3 \Dit eindspel biedt de sterkste partij zeer goede winstkansen. dat hij de verdediging als kansloos beschouwde. p. F. it was exhibited at about 1915 Bell 1978. Torres-Quevedo 1951]. this article is not generally known to the computer game community. The only copy I was able to nd was archived at the National Library of Latvia. \Torres believes that the limit has by no means been reached of what automatic machinery can do. the research reported here argues for a renewed appreciation of the accuracy and importance of Amelung's work. 8{11. Computer endgame analysis. was discovered by Richard Bellman 1965].) ." Unlike most later work. Babbage 1864. (Strangely enough. his conclusion suggests that he did not appreciate the complexities involved: In consequence of this the whole question of making an automaton play any game depended upon the possibility of the machine being able to represent all the myriads of combinations relating to it. According to Scienti c American 1915."3 The Duna-Zeitung has turned out to be an elusive newspaper. I found that the combinations involved in the Analytical Engine enormously surpassed any required.3. By contrast. The contemporary dynamic programming methodology. 124{125]. The mathematical justi cation for the retrograde analysis chess algorithm was already implicit in Zermelo 1913].

and heuristic evaluation could be used to solve much larger state spaces than could be tackled by either technique alone. 1994. 3]. J. and probably all endings involving a minor piece and pawns. 1996]. In checkers. for checkers Lake et al. Bellman 1954. 1987. Whether or not this is desirable is another matter Bellman 1961. where he observed that the entire state-space could be stored and that dynamic programming techniques could then be used to compute whether either side could win any position. Hans Berliner and Murray S. Bellman predicted that checkers could be solved by his techniques. 1983] analyzed KRKN from the point of view of a machine learning testbed. and the utility of his algorithms for solving very large state spaces has been validated by Jonathan Schae er et al. Jaap van den Herik and coworkers have produced a number of retrograde analysis studies of various four-piece endgames. A. but his work came to fruition with Bellman 1965]. 1957] had considered game theory from a classical perspective as well. 1988]. On the other hand. 1988]. H. the general situation is still rather complex. Bellman's work was the culmination of work going back several years: Checkers and Chess. In chess. the number of possible moves in any given situation is so small that we can con dently expect a complete digital computer solution to the problem of optimal play in this game. although they were not able to solve the general instance of such endgames. E. p. Campbell 1984] studied the Szen position of three connected passed pawns against three connected passed pawns by simplifying the promotion subgames. Schae er et al. The rst retrograde analysis of general ve-piece endgames with up to one pawn was published in Thompson 1986]. The signi cance of this work was . 1992. but we can use the method described above to analyze completely all pawn-king endings. van den Herik and Herschberg 1985a. Bellman also sketched how a combination of forward search. or of endgames with more than 4 pieces whose special structure allows the state-space size to be reduced to about the size of the general four-piece endgame Dekker et al. 1992c] has studied endgame play when the opponent is presumed to be fallible. 1992b. but where the deviations are reasonably small in number. Futer 1974] studied certain special cases of KQPKQ. Danny Kopec has written several papers in the area as well Kopec et al. S. Komissarchik and A. Campbell 1988] has begun to extend this idea to wider classes of endgames. Jansen 1992a. 4 4 4 tic-tac-toe has been solved by Patashnik 1980] using forward search and a variant of isomorph-rejection based on the automorphism group computation of Silver 1967]. Herschberg 1986]. are checkers and chess.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 9 Herik and I. dynamic programming. L. Schae er and Lake 1996] and Ralph Gasser for Nine Men's Morris Gasser 1991. Interesting examples of processes in which the set of all possible states of the system is indescribably huge. Ross Quinlan 1979.

10 LEWIS STILLER twofold. Valiant 1990a. in fact.. it is hoped that the so-called \von Neumann bottleneck" between the CPU and the memory of a standard serial computer could be alleviated by massive parallelism Hwang and Briggs 1985]. with processors connected if and only if their addresses di er in exactly one bit location. although we shall not discuss the techniques required for that here. Two processors connected by an edge can communicate directly in a single timestep. First. Second. and imagine each edge of the cube as a wire that directly connects the processors at its endpoints. The winning procedure proved to be quite di cult for humans to understand Michie and Bratko 1987]. such as the type of interconnection network. won. Thompson's work invalidated generally accepted theory concerning certain ve-piece endgames by demonstrating that certain classes of positions that had been thought to be drawn were. of course. Blelloch 1990]. and whether each processor executes the same instruction (SIMD) or di erent instructions (MIMD) Hillis 1985. 1990b. Higher-dimensional shift patterns are also easy to compute on a hypercube Ho and Johnsson 1990]. E8 shifts each element of the array down. It is also possible to perform arbitary permutations on a hypercube. see Figure 1.1) The computation is performed by Gray coding the coordinates of the elements of the array v so that vi and vi+1 are physically adjacent in the hypercube Gilbert 1958]. many trade-o s that can be made in the architecture of a computer. One common form of interconnection network is the hypercube Leighton 1992]. Consider a parallel computer with 2n processors. Each processor in an n-dimensional hypercube can be viewed as having a length n binary address. A @ . many more moves were required to win certain endgames than had previously been thought. Parallel Processing The motivation for using parallel processing is to achieve increased computation bandwidth by using large numbers of inexpensive processors van de Velde 1994. Place each processor at a distinct vertex of the unit cube in Euclidean n-space. The pawnless ve-piece work of Thompson was extended to all pawnless ve-piece endgames and many ve-piece endgames with one pawn by an early version of the program discussed in this paper.. each with some local memory. In practice general . Stone 1993]. lling the initial element with 0: 0 v1 1 0 0 1 B v2 C B v1 C C B C = E8 B B C B @ . In particular. C A v8 . (3. 3. the granularity of the processors. v7 . There are. Applied to an array. The hypercube can compute the e ect of the end-o shift matrix E8 .

particularly to the parallel and vectorized computation of fast Fourier transform. Let In be the n n identity transformation of Vn : Write diag(v 0 . are ordered by =1 4. 1992]. Let Vn be the space of length-n vectors n with entries in a eld F. In the following. Parallel Processing and Tensor Products This section brie y summarizes some previous work on the application of tensor products to parallel processing.j =0 . This gure illustrates a three-bit Gray code. Let Mn m be the space of n m matrices over F. Let Mn = Mn n . v1 . m n?1. 4. Left multiplication by an end-o shift matrix can be computed by processor permutation is typically performed with hardware assistance Nassimi and Sahni 1982. a linear transformation will be identi ed with its matrix representation in the standard basis. An element of Vn may be thought of as a length-n array whose elements are in F.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 11 111 101 100 001 011 010 000 Gray coding the coordinates of each element of the array. vectors (having one component 1 and all others 0). The chess algorithm will be developed in the next section by generalizing this approach. which can be thought of as an embedding of a cycle of length eight into a three-cube. or as an n 1 matrix over F Marcus 1973]. Leiserson et al.m?1 The mn basis elements of Vn Vm . Mathematical preliminaries.1. fen i ej gi=0. 110 Figure 1. : : : . vn?1 ) diag(v ) for the diagonal matrix in Mn whose diagonal elements are taken from the coordinates of v: . We let fen i gi be the standard basis consisting of en i em j 7! emn mi j : + In this manner an element of Vn Vm may be considered to be a vector of length mn with elements drawn from F.

allows the order of evaluation in a tensor product to be varied. taking m steps of size l. the Commutation Theorem will be the method by which one type of execution is traded o for another. A B C B . The e ect of Pln on a vector is to stride through the vector.. If A 2 Mm . Anm B C A (In 0 0v 11 B B B AC @ . C 0 1 0 1 2 1 1 ( 1) (4. for B 2 Mm Johnson et al... 1989]: Theorem 4.12 LEWIS STILLER n0 nn0 If A 2 Mn m and B 2 Mm0 . v nm? C B C B 1 0 B v n? m C C B B A @B B AC @ . C C B C B C B v m? B 0 vm 1 C C 0 v 1 B C B C B B C .2 that some evaluation orders naturally correspond to vectorization... For example. l = 2. .. taking m = 3. and n = 6. we have n = B A: Pln (A B )Pm This theorem. . where v 2 Vm and w 2 Vl . 1991]: An1 B An2 B .1) Suppose n = ml. C B = B) B C B C C v m? @ . which is easy to prove even when the entries are from a semiring. and n = ml.. A1m B A2m B C C .. . A . B 2 Ml . C B C B v C BB @ .. C B .1... we have: v nm?1 0v 1 0v 1 B B v C v C C C B B C B B v C Bv C B P B C C=B C B v C Bv C C B B A @v A @v C 0 1 2 3 4 0 2 4 1 3 6 2 v5 v5 Stride permutations are important due to the following Commutation Theorem Tolimieri et al. and some to parallelizations.. 21 22 A11 B A12 B The importance of the tensor-product to our work in parallel processing lies in the following identity. The n-point stride l permutation matrix Pln is the n n matrix de ned by Pln (v w) = w v . We shall see in Section 4. @ . the matrix of the tensor product A B 2 Mmm0 is given by 0 1 BA B A B A B=B B ..

encodes all such scheduling. v(n=m)?1 are stored in processor 0. elements v0 . The target architecture of the language is a machine comprising m parallel processors. We now 0 B B @ v0 v1 Assuming that m divides n. User data is always stored conceptually in the form of a vector 4. which can be important for time-critical applications. The most serious disadvantage is its narrowness of application. Whereas most high-level languages. Originally developed for signal processing codes. communication. so that computations within the processors should be vectorized. The matrix notation we use is nothing more than an informal notation for describing algorithms. Suppose B 2 Ml . On the other hand. and operation scheduling. It di ers from standard notations primarily in its explicit denotation of data distribution. : : : . each with shared memory. but there are many applications which would not easily fall under its rubric. Because many of the algorithms to be presented will be presented in the tensorial manner. the notation we use. . v n?1 . on local-memory machines. and even special-purpose parallel languages. 1 C C A: . and so on. We do not assume restrictions on the vector length capability of the processors. code(B ) is a program that takes as input an array v of l elements of F. it requires some conscious decisionmaking over data-distribution that is unnecessary in some other languages. That is.2. where vectors are identi ed with their coordinates in the standard basis. with an extra communication step or two. and let code(B ) be any sequence of machine instructions that computes the result of left-multiplication by B . However. It is assumed that certain general classes of speci c matrices are already implemented on the architecture. : : : . leave the distribution of data over the processors and the scheduling of operations within processors to the discretion of the compiler. in particular.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 13 describe the relationship between the matrix representation of a formula and the denoted machine code. the stride permutations and any speci c permutations corresponding to the interconnection network. this section is fundamental to this work. Each processor may also have vector capabilities. at least potentially. it is easy to see that the results go through also. elements vn=m .. Matrices are stored in column-major order. and returns as output the array B v of l elements of F. v(2n=m)?1 in processor 1. this work demonstrates its wider application. the functional nature of the notation does make it potentially amenable to compiler reordering. This has both advantages and disadvantages: although it gives the programmer a ner level of control. with the code-generation phase only represented implicitly. Code generation: Conversion from factorization to code.

is a bit more subtle.j =0 a primitive n-th root of unity. it is easy to nd code(M M 0 ). and then returns the coordinate-wise sum of the arrays that are returned by these two subroutine calls. 4. rst runs as a subroutine code(B ) on v (saving the result). Of course. Simply let code(B + B 0 ) be the program that.1 and Table 1. given code(M ) and code(M 0 ). Fast Fourier transforms. pi runs code(B ) on the l elements of v that are stored in its local memory.3. where ! is Let Fn be the n-dimensional Fourier transform matrix (!ij )i. Fn = (Fm Il )Tl (Im Fl )Pm . pm . Let n = ml. except that it operates on l-vectors rather than on scalars. one can interpret code(Im B ) as code that runs on this m-processor architecture. The exposition of the chess material. then runs code(B 0 ) on v. This corresponds closely to hardware primitives on certain vector architectures. one can de ne concise expressions that can be translated into e cient code for certain classes of functions Granata et al. assuming the dimensions of M and M 0 are compatible: run code(M ) on the result of running code(M 0 ) on the argument v. p1 . and some simple tensor product identities. for A 2 Mm . Similarly. each with some local memory. it is easy to compute some code(B + B 0 ). The presentation is intended to illustrate the parallel code development methodology used to describe parallelization of the chess endgame algorithm. We make the convention that a length ml array will be stored with its rst l elements in processor p1 . its second l elements in processor p2 . The interpretation of code(A Il ) is as the code corresponding to A. We discuss brie y the formulation of the FFT in the tensor product framework presented above Tolimieri et al. code(Il ) is the code that returns its argument. This code can be constructed (loosely speaking) from code(A) by interpreting the length ml argument array v as being an element of the m-module over the ring Fl . in Section 6. and outputs the result to its local memory. Similar rules can be derived when the number of processors is di erent from m. 1993. When called on a length ml array v . Given this convention.14 LEWIS STILLER Given code(B ) and code(B 0 ) for two matrices B and B 0 . n?1 . The code corresponding to A Il . : : : . To construct code(Im B ). and so on.1) shows that this will compute the tensor product. load code(B ) in each pi . The Singleton 1967] mixed-radix version of the Cooley{Tukey fast Fourier transform Cooley and Tukey 1965] can be expressed recursively by n. By combining a xed set of transformations re ecting the hardware primitives of the underlying architecture with combining rules like +. The relation A B = (A Il ) (Im B ) can be used to compute general tensor products. Consider a parallel processor with m processors. 16{ 20]. pp. and . given its input array v. does not depend on any of the results here. an l-vector. Equation (4. however. 1992].

Fn =Pm m l (4. The Im Fl term vector v.3). 1989]. !. each of length l. Abdali and Saunders 1985. Application to Chess This section describes the chess endgame algorithm in a generalization of the tensor product formalism described in Section 4. Tl just multiplies each element by a twiddle factor. We write + and for _ and ^. !l?1 ) j : This can be interpreted as a mixed parallel/vector algorithm.2). with di erent tradeo s between parallelization and vectorization Chamberlain 1988. matrix product. By using the commutation theorem and varying the factorization. Auslander and Tolimieri 1985]. Tarjan 1981a. and the relation between and parallelization still holds. ^g. as does. for the moment. 1981b]. therefore. of course. Imagine. In particular. or using the mathematics-of-arrays formalism of Mullin 1988]. The commutation theorem can be used to derive a parallel form n (Il Fm ) P n Tl (Im Fl ) P n . in the natural way. Pm performs m l-point FFTs in parallel on each segment. This generalization has been used for expressing graph algorithms Backhouse and Carre 1975. Finally.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 15 where Tl is a diagonal matrix encoding the twiddle factors: Tl = m ?1? M j =0 diag(1. and matrix sum remain essentially unchanged. Bird et al. many different FFT algorithms have been derived. Johnson 1989. and the vectorized Korn{Lambiotte FFT Korn and Lambiotte 1979] can be derived by unrolling (4. 1991. . Given an input n v forms a list of m segments. These ideas could. the Fm Il term performs an m-point FFT on vectors of size l. 5. Mn m . Granata et al. The notion of linear transformations then changes. the commutation theorem. 1. : : : . be presented categorically using the approach of Skillicorn 1993.3) and a vector form n (Fl Im ) : Fn =(Fm Il ) Tl Pm The parallel Pease FFT Pease 1968] can be derived by unrolling the recursion in (4. The de nitions of . _. 1990. the notion of code(M ). Lehmann 1977. that a matrix over GF2 is actually a Boolean matrix whose entries are taken from the Boolean algebra f0. Averbuch et al.2) (4.

k Let B C be the hyperboard corresponding to S .s . if k = 6 one could choose S = hk. This correspondence between sets of chess positions and elements of B forms the link between the chess algorithms and the tensor product formulation. stalemates. Ng is a piece. and therefore corresponds to a set of squares White 1975]. The ideas of Section 4. we need not worry about this distinction. pawns. R. It can be thought of as k . For example. given an S -position P returns the set of S -positions that could be formed by unmoving Ss in P as if Ss were a p. each of which is formed from pieces of S . : : : . We write S = hS1 . castling. (Technically. is associated with a unique element of B : the sum of the basis elements corresponding to each of the positions from the set. Let S be an ordered set of k chess pieces. C V V: N Let fei gi N be the standard basis for V . q. and thereby becomes an element of M64k . Q. Each basis element of C is of the form ei ej for 1 i.2 may then be used to derive e cient parallel code from this factorization. However. We denote by Vn the space of length-n Boolean vectors. j 8.16 LEWIS STILLER 5. ri: An S -position is a chess position that contains exactly the k pieces in S . captures. Each position that is formed from the pieces of S is thereby associated with a unique basis element of B. We now describe a factorization of Xp. the unmove operators are only quasilinear. Any set of positions. and the fty-move rule will be disregarded unless otherwise stated. For simplicity of exposition. Since each cs is a square on the chessboard. R.s in terms of primitive operators. B. each basis element of B can be thought of as a sequence of k squares of the chessboard.s is the function that. Any such basis element.) The core of the chess endgame algorithm is the e cient computation of the Xp. Any element of C is a sum of distinct basis elements. Each basis element of B is of the form c1 c2 ck . therefore. Q. and thus B is not a module. S2 . If p 2 fK. Xp. K. the unmove operator Xp. the distinction between sets of chess positions formed from the pieces in S and elements of the hyperboard B will be omitted when the context makes the meaning clear. since the Boolean algebra is not a ring. where each cs is some basis element of C. Sk i: An S -position can be viewed as an assignment of each piece Si 2 S to a distinct square of the chessboard (note that captures are not allowed). and j V the j -th tensor power of V. In the following. De nitions.1. The space of 8 8 Boolean matrices is thus 8 8 8 =1 8 a cube of side-length 8 in R 2 : Each of the 64k basis elements corresponds to a point with integer coordinates between 1 and 8.s can be extended to a linear function from elements of B to itself. denotes a unique square on the 8 8 chessboard.

There is a close correspondence between multilinear algebra. We will use the group-theoretic terminology both to give concise descriptions of certain move operators and to describe the exploitation of symmetry. Endgame Algorithm We now present the endgame algorithm using the notation developed in Section 5.3 describes the control structure of the algorithm. combinatorial enumeration. f(ei ej ) = ei e8?j+1: Thus. 6. D4 . p.2. Recall from (3. 1992. Section 6.1. R G x= X g2G gx: 6.1 gives the fundamental factorization. The symmetric on Sk acts on B by permuting the order of N group Nkk elements the factors: s k c = c . 1981. 6]. A group G acting on Vn and Vm acts on g(M g?1(v)).1) that E8 was de ned to be the unit one-dimensional 8 8 end-o shift matrix. and group actions which motivates much of this section Merris 1980. s=1 s s=1 ss The dihedral group of order 8. Merris and Watkins 1983]. .2 describes the modi cation of the equations of Table 1 to exploit symmetry. r rotates the chessboard counterclockwise 90 and f ips the chessboard about the horizontal bisector. Section 6. We let Z to The notation G x is intended to represent the group average of x withR respect G Fulton and Harris 1991. The unit multidimensional shift along dimension s is de ned by Us 2 M64k I64s?1 (E8 I8 ) I64k?s : Such multidimensional shifts are commonly used in scienti c computation. Group actions. Section 6. D4 acts diagonally on B by d k O s=1 cs = k O s=1 dcs Mm n by conjugation: (gM )v = Let C4 be the cyclic group generated by r. It is a xed point of the G action: g Gx = RGx for all g 2 G. is the group of symmetries of the square. It is generated by two elements r and f with relations r4 = f2 = e and r3 f = fr: It acts on C by r(ei ej ) = e8?j+1 ei.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 17 5. We introduce a few group actions Fulton and Harris 1991]. Factoring the unmove operator. for s 2 Sk and cs 2 C.

3).s = L XB. By modifying the factorizations. and de ne L 2 M64k diag Z X ! ci1 c ik Certain basis elements of B do to legal S -positions. 1993].s C4 Table 1. conjugation by r of the operation of moving the R right corresponds to moving the R up: if one rotates the chessboard clockwise 90 .2) and (4. The vectorized implementation of Table 1 by Burton Wendro et al. Other factorizations appropriate for combined vector and parallel architectures.s = (1 s)Xp. if the interconnection architecture is a two-dimensional grid.s = XR. Figure 2 illustrates the computation of the integrand in the expression for XR. For example.s = LEWIS STILLER Z C4 XN. the result will be the same as if the R had been moved up to begin with. by varying the factorization. equations appropriate for a grid architecture can be derived. can also be derived Kaushik et al. These equations de ne the core of a portable endgame algorithm. These equations are vectorizable as well Smitley 1991]. Sk i1 < <ik .18 XR. Fix a basis fci g64 i=1 of C.s = L Z XQ. one can derive code suitable for di erent architectures.s + XB. 1993].s = Z LUs(I64k + LUs )6 Us (r(Us2 )) LUs (I64k + LUs rUs )6 Us + Us rUs Z D4 D4 XK. The average over C4 means that the R must be moved in 4 directions. code suitable for execution on a wide range of high-performance architectures can be derived. As in the case of fast Fourier transforms (4. moves the R right. s0 for s=1 s which cs = cs0 . For example. Table 1 de nes the piece-unmove operators. only Us for s = 1 can be directly computed.1 in Table 1. These Nnot ccorrespond \holes" are elements of the form k such that there exist distinct s. If v 2 B then Lv is the projection of v onto the subspace of B generated by basis elements that are not holes. This corresponds to moving the R to the right. and then rotates the chessboard counterclockwise 90 . By using the relations Us = (1 s)U1 and Xp. where (1 s) 2 Sk interchanges 1 and s. such as a parallel network of vector processors. has supported this claim Wendro et al.1 .

S . we need Nk?of 1 C. Therefore. Therefore. because each Xp. of the chessboard. Unmoving the R to the right from the position on the left results in the three positions shown in the center. = hR Ki Each position e e e e corresponds to a point in the hyperboard.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 19 Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z S ZKZ Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z ZRZ ZKZ Z Z Z Z Z Z Z Z e e 3 e 3 e 6 e 3 2 e 3 e 6 e 3 Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z ZRZKZ Z Z Z Z Z Z Z Z e 4 e 3 e 6 e 3 Z Z Z Z 5 3 6 3 Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z SKZ Z Z Z Z Z Z Z Z Figure 2.2. The position e6 e3 e6 e3 is illegal and is zeroed out by L. Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z j Z Z Z jkZ Z jkj Z Z jkjkZ Z . shown on the right. Here. the class of positions considered can be restricted to those in which the k is in the lower left-hand octant. Algebraically. The game-theoretic value of a chess position with- out pawns is invariant under rotation and re ection of the chessboard. The chessboard may be rotated 90 or re ected about any of its bisectors without altering the value of a pawnless position. : 6. as shown in Figure 3. Exploiting symmetry. or fundamental region.s is a xed point the D4 action. the k may be restricted to lie in one of the ten squares shown. These positions correspond to points in a triangular wedge in the hyperboard. rather than the bigger only consider the 10 64k?1 -space B0 C=D4 Figure 3.

as in going from the top to the middle positions. is the k. These three positions correspond to three points in the hyperboard. Only an eighth of the hyperboard is physically stored. Xk. the resulting position must be transformed so that the k is in its fundamental region. only the rst and third of which are physically stored. X 0 Ok?1 X0 d.1 d 2 M10 . However. The Black-to-move position at the top requires 222 moves against best play for White to win (see Table 2). Figure 4. the induced motion in the hyperboard remains within the wedge. this is represented by the position at the bottom. the rst piece of S . the induced functions X0p. When pieces other than the k are moved. corresponding to the rst factor in the expression for B0 .20 LEWIS STILLER Z Z Z Z Z Z Z Z KZnm Z Z Z Z Z Z Z Z Z Z ZBZkZ Z S Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z KZnm Z Z Z Z Z Z Z Z Z Z ZBZ j Z S Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z mnZK Z Z Z Z Z Z Z Z Z j ZBZ Z Z Z S Z Z Z Z moved outside the squares marked in Figure 3. Algebraically. .s : B0 7! B0 have the same form as Table 1 when s 1.1 = d2D4 where X0k. we apply a symmetry transformation that puts the k back in the allowed area. By convention. when the k is moved outside its fundamental region. This transformation of the chessboard induces a transformation on the hyperboard (Figure 4). When the k is 64k -space B.1 d k. Thus.

and whose edges are labeled by one of the generators r or f. . The position shown arose during a game between Anatoly Karpov and Gary Kasparov in Tilburg. thus. Gray (diagonal) arrows correspond to f. October 1991. An edge labeled h connects node g to node g0 if hg = g0 . The communication complexity of the routing can be reduced by exploiting the Cayley graph structure Stiller 1991a]. and ip the board horizontally.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 21 The sum over d 2 D4 corresponds to routing along the pattern of the Cayley graph of D4 (see Figure 5). Each node is pictured by showing the e ect of its corresponding transformation on a position in KBNNKR. The Cayley graph for D4 . This is a graph whose elements are the eight transformations in D4 . The actual Z Z s Z Z Z Z Z Z j ZBZ Z Z Z Z J M Z Z Z ZNZ Z Z Z Z Z Z Z Z Z Z s Z Z Z Z Z Z ZBZ j Z Z Z Z Z Z Z M J Z ZNZ Z Z Z Z Z Z Z Z Z Z s Z Z Z Z Z Z ZBZ j Z Z Z Z Z Z Z M J Z ZNZ Z Z Z Z Z Z Z Z Z Z Z s Z Z Z Z Z Z j ZBZ Z Z Z Z J M Z Z Z ZNZ Z Z Z Z Z Z Z Z Z Z s Z Z Z Z Z Z ZBZ j Z Z Z Z Z Z Z M J Z ZNZ Z Z Z Z Z Z Z Z Z Z Z s Z Z Z Z Z Z j ZB Z Z Z Z Z J M Z Z Z ZNZ Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z ZNZ Z Z Z M J Z Z Z Z ZBZ j Z Z Z Z Z Z s Z Z Z s Z Z Z Z Z Z ZBZ j Z Z Z Z Z Z Z M J Z ZNZ Z Z Z Z Z Z Z Z Z Figure 5. Black arrows correspond to r. the chess value of each of these nodes is the same. and rotate the board counterclockwise 90 .

MAGMA. which is used to compute the FFT of a data set invariant under a crystallographic group An et al. v2 is the set of positions from which White can either checkmate Black in one move or can move to a position from which any Black reply allows a mate-in-one.. possibly in conjunction with more traditional move generators. The method for computing vi from vi?1 is called the backup rule. such as GAP. The method we use turns out to be similar to the orbital exchange method. pp. Lake et al. and AXIOM. The groups encountered here are so small. of the number of moves required to win that position Strohlein 1970. for performing the purely algebraic operations required. Although their terminology is di erent from our terminology. and i is the maximin value of the set S : the maximum. Several backup rules have been used in various domains Schae er et al. in general. 472{475].s : The backup rule used is vi+1 = XWhite(XBlack (vi )). The overall structure of the algorithm is to iteratively compute the sets v1 . Thus. Control structure. i White moves and i ? 1 Black moves).3. Sims 1994]. where Symj is the j -th symmetric power of C Fulton and Harris 1991. For i 1 we de ne v i 2 B to be the vector of positions from which White to move can checkmate Black within i moves (i. Tolimieri and An 1990]. 1994]. plus some cycles White 1984]. where a vertical bar denotes complement. however.s . . 1992. 6. is Black g XSs . Then v = v i is the set of positions from which White can win. There are e cient algorithms. which looks like a number of disjoint copies of the Cayley graph. or move them backward.e. that are suitable for the denotation of the algebraic structures used Butler 1992. Thompson 1986]. over all positions from which White can win. their general ideas are similar Hillis and Taylor 1990]. and so on. until some i is reached for which v i = v i+1 . The problem of parallel application of a structured matrix to a data set invariant under a permutation group has been studied in the context of nite-element methods by Danny Hillis and Washington Taylor as well. : : : .22 LEWIS STILLER communication pattern used is that of a group action graph. that computer-assisted group-theoretic computation was not required. v 2 . They are all characterized by the use of an unmove generator to \unmove" pieces. v1 is the vector of positions from which White can checkmate Black on the next move. 1990. We let XWhite = XBlack = X fs : Ss fs : Ss X is White g XSs . 1992. as well as languages. It is interesting to note that exploiting symmetry under interchange of identical pieces can be handled in this notation: j identical pieces correspond to a factor Symj C in the expression for C. Sims 1971.

The broadcast is a tensor product of copies of an identity matrix with the 1 64 matrix of 1's. or SPREAD primitive Adams et al. of the distance-to-win of that position. Each pawn position induces a separate hyperboard. over all positions using those pieces from which White can win. it does not matter for most purposes how many moves are required to win the subgame KRK. The more usual distance-to-win metric is simply the number of moves required by White to force conversion into a winning subgame. a diagonal matrix. The distance-to-win of each point in the hyperboard can be stored so that 7. Each subset of the original set of pieces S induces a subgame. The max-to-win for a set of pieces is the maximum. but when captures are considered they are slightly di erent. Unmoving pieces cannot capture. Re nements 7. Pawn unpromotion induces communication between a quotient hyperboard and a full hyperboard.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 23 7. Database. Captures and pawns. moving and unmoving are the same. in KRKN. then against perfect play the win value would be altered by an m0 move rule for m0 < m. In practice. it can produce misleadingly high distance values because the number of moves to mate in trivial subgames. the results presented here use the more conservative distanceto-win metric. leaving a piece in their wake. The algorithms developed so far must be modi ed to account for captures and pawns. not to moving them Thompson 1986]. The uncapture operation can be computed by using outer products.s developed in the preceding section refer to unmoving pieces. and a broadcast. when White plays so as to checkmate Black as soon as possible.1. Although the distance-to-mate might seem like the natural metric to use. It also is the metric of relevance to the fty-move rule. There are two values that can be associated with a position: . If a particular position has a distance-to-win of m. again implemented by multiplication by D4 . like KRK. and each subgame has its own hyperboard Bellman 1965]. An uncapture is the product from left to right of an unmove operator in the parent game. once White captures the N. as long as the N is captured safely Rasmussen 1988]. corresponding to the parallel broadcast. Without captures. Although our program has implemented distance-to-mate metric for vepiece endgames. would be included in the count of something like KRKN.2. In fact. The distance-to-mate is the number of moves by White required for White to checkmate Black. This is simulated via interhyperboard communication. 1992. this metric is more useful when the position has no pawns. but they can uncapture. and Black tries to avoid checkmate as long as possible Zermelo 1913]. The equations for Xp. a sequence of stride matrices. distance-to-mate and distance-to-win. Stiller 1992a].

185. Thus. because games arising from permutations of the identical pieces are counted as di erent. although some special con gurations may not be wins for obvious reasons. with 223 moves required to win in the worst case. and most of the nontrivial ones had max-to-wins of approximately 50. Rosenberg 1988. Later.24 LEWIS STILLER a two-ply search permits optimum play. or it may be able to capture a piece in one move. The distance-to-mate results were not very illuminating. although it is not a general win. When there are repeated pieces. Gray codes. therefore. a number explained as follows: there are 462 arrangements of two nonadjacent kings. the aggregate statistics (though not the percentages) are also in ated. 385. Because many interconnection networks are Cayley graphs or group action graphs Annexstein 1990. but it is counted six times. like KQNKN. Curiously. a leading endgame expert. The max-to-win values were signi cantly higher than previously known endgames. although the size of each hyperboard is 462 644 7:7 109 nodes. Roycroft. for each such arrangement.) We remark that KRBKNN is a general win. Note that this in ates the statistics for wins. KRNKNN has the longest known max-to-win of 243. the motif of embedding Cayley graphs into Cayley graphs arises several times in this work. several pawnless six-piece endgames were also solved. By Gray coding this distance. such as the immediate loss of a piece by White due to a fork. For most endgames. we considered 6. Table 2 presents statistical information about these six-piece endgames. This work resulted in the rst publication of the 77-move KBNKN max-to-win. modulo the dihedral symmetry. for example. which can be viewed as embedding the Cayley graph for Z2n into that of Zn 2 . Thus the state space per endgame has about 6:2 109 nodes. Embedding the Cayley graph for D4 in that of Zn 2 arises during unpromotion and moving the K. 360 = 462 62 61 60 59 positions. 1991] this motif will reappear on other implementations. where all \normal" starting con gurations lead to a win. Some endgames were solved under the distance-to-mate metric as well. which at the time was the longest known pawnless max-to-win Stiller 1989]. said in 1972 that this . are used both for implementing U (and. Xp. the remaining four pieces are placed in any available square. 8. there is really only a single mutual zugzwang in KRRRkq.s ) and for maintaining the database. Chess Results The combinatorially possible pawnless ve-piece games and many ve-piece games with a single pawn were solved using an early version of the current program. the increment of the value can be done by modifying only one bit. No ve-piece endgame had a max-to-win over 100. (The phrase \general win" is not susceptible to precise de nition. Draper 1990. because of the advantage of the rst move in a random position: White is already won if the k is in check. It applies to any class of games.

and number Z of mutual zugzwangs. number and percentage of positions that are wins for White. Bishops linked with a bar are constrained to lie on squares of opposite colors. he called only \controversial or unknown". particularly when no pawns are involved. Cheron 1952. The 92-move win in KQRKQR is somewhat surprising too. Cheron. seems to reserve judgment. 5. mutual zugzwangs can sometimes be understood by humans. Most of the standard works concurred with the opinion that KRBKNN was not a general win Euwe 1940. Unlike the \maximin" positions such as the one in Table 3. Fine 1941. p. Such positions seem amusing because. Statistics gathered for six-piece endgames: description of endgame. Berger 1890. . The fty-move rule would a ect the value of each endgame listed with max-towin of fty or more. 50{53. maximum D of distance-to-win over all positions that are wins for White. this reduces the state space roughly by a factor of four. from 6:19 109 to 1:67 109 nodes. however. A mutual zugzwang is closely related to a game whose Conway value is zero: it is a position in which White to move can only draw. con guration was \known to be a draw". chess is a very hot game in the sense of Winning Ways Berlekamp et al. 417.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 25 KRNknn KRBknn KRNkbn KQNkrr KRNkbb KRRNkq KRBNkq KRBkbn KNNNkb KQRkqr KNNNkn KQBkrr KRRBkq KRBkbb KQRkqb KRRkrn KQNNkq KQRkqn KBBBkr KBBNkr KRRRkq pieces # wins 4821592102 5948237948 4433968114 5338803302 4734636964 5843483696 4242312073 5359504406 5324294232 5125056553 5834381682 5707052904 5935067734 1123471668 5365200098 5023789102 5808880660 5553239408 4944693522 1497242834 6054654948 % 78 96 72 86 77 94 69 87 86 83 94 92 96 72 87 81 94 90 80 95 98 D Z 243 223 190 153 140 101 99 98 92 92 86 85 82 75 73 73 72 71 69 68 65 18176 456 8030 1858 1634 1520 1010 1478 6300 243 12918 342 388 95 1410 1410 2228 1780 48 83 6 KQkbbn KRRkrb KRNkbb KQkbbb KBNNkr KQQkqr KQBkqb KQQkqq KRBBkq KBBknn KRRkbb KQBkqn KQknnn KQBkqr KQNkqb KNBBkn KQNkqn KQNkqr KRRkrr KBBNkq pieces # wins 5257968414 4529409548 1015903231 5058432960 3302327120 5689213742 4079610404 5122186896 1185941301 981954704 1483910528 4213729734 4626525594 3825698576 3789897570 6130532196 3920922433 3533291870 4136045492 970557572 % 85 73 65 82 53 92 66 83 75 63 94 68 75 62 61 99 63 57 67 62 D Z 63 54 52 51 49 48 46 44 44 38 37 36 35 32 32 31 29 27 18 12 6670 1030 256 2820 1270 32 22 32 396 1662 26 78 17688 6 35 58 152 3 16 18 Table 2. whose analysis is fairly impenetrable. 167{ 169. p. 1982]. while KBBKN. which was considered a draw by most players. pp. 521]. pp. Vol. but Black to move loses.

as the R would immediately capture the other N. 1. Ng6+ wins) 6. and the resulting subgame of KBNK would be winning for White. Kg7 and White wins. if White moves rst then Black can force the draw. Kd3 Kd6 5. This mutual zugzwang. Kc2 Kc5 (If 3. Rusinek 1994. it would be captured. : : : Kh2 2. is somewhat inelegant in that includes promoted pieces. any Black move loses. 1. For example. Qg7+ Not 1. shows that this general position in KBNNKR is drawn. : : : Qh3+ 3. : : : b1/Q loses to 3. Chess theory. but now 2. On the other hand. Ke4 Ke7! draws. 2 : : : Qb5+ . f8/Q Qh3+ 4. Black tries for perpetual check. f8/Q (interpolating further checks does not help) when 2. On the other hand. drawing. : : : Kc3 2. if the White K reaches g7 rst. Kb1 Kd4 3. The position seems to be a race between Kings to see who will reach the upper right corner area rst. Noam Elkies used it as the basis for a much more elegant composition. con rmed by the program. Ng6+ wins the R) 4 Kd3 Kd6 5. #546] is shown in Figure 7. Kg5 b1/Q with Kh1 and Be4 draws. Kg5 Qe3+ forces either perpetual check or a queen trade. Now consider Figure 7. but Black to play loses. 4. Qf4+ Kg1 4. If the Black K reaches g7 or e8 rst. Any other move of the B on move 1 allows Black to win the race. the Black R can sacri ce itself for a White N. If Black moves the Qa7 then Qhg1 or Qa2 mates. Note also that neither N can move. and not with 2 : : : Qd1+? 3. If the Qf6 moves then Qb2 or Qf1 will mate. We quote from Elkies' analysis: \1. leaving the drawn endgame KBK. 1. Be4+ and mate. Ke4 Kc7 (If Ke7? 6. f8/Q. since g8 is guarded by the B on a2. It is not di cult to see that Black to play loses: White gets in rst. Ba2 Kc2. Bb1+ Kc3 forces 2. This composition Elkies 1993. discovered by the program. The R is trapped on h8. For example. since other moves by White on the second move allow the Rh8 to escape via g7. Figure 6 shows an example. discovered by the computer. If Kb1 then Qc2 mates. White to move from the position in Figure 6 must move the B.26 LEWIS STILLER Z Z M s Z Z Z ZN Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z BZkZ Z Z J Z Z Z Figure 6. Kc2 Kc5 4. Bf3. Kf5 Kd8 7. right. 1. left. Bf7 Kc3 2. If 2. it simply captures the Black Rh8. one that follows accepted aesthetic practice by avoiding the use of promoted pieces in the original position and by appearing more natural. Thus. Mutual zugzwang: White to play draws. If the Black R were to capture a N. and then the Black K captures the other White N. However. Kb1 Kd4 3. Qd6+? Kxg2 2. Thus. : : : Ke5?. and the N's guard each other. Kg6 Ke8 8. Qe5+ Kxg2 3.

4. leading to a version of the position on the left. 6. Kxg5 b1/Q 7. White to play and win.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 27 Figure 7. . Kxh7 b1/Q+ 5. Z Z Z Z l Z Z Z Z Z l Z Z Z Z Z Z Z Z Z Z Z Z Z Z Z ZQL j Z Z ZK Z Z L Z Z Z ZPZb Z Z Z Z Z Z Z ZK Z Z Z Z ZqZ Z j o Z ZBZ Z Z Z Z 3. Kh6 Qb6+ 4. So Black does manage to give the rst check in the four-queen endgame. Kh8 Qb8! drawing. During an elite tournament in Tilburg. : : : Qe3+ 5. Starting point for the longest KRNKNN ght. : : : Kh1 White wins only with 7. Kxh7 b1/Q+. but they do arise sometimes. bringing about a rotated version of the KQQKQQ mutual zugzwang]. Recall that this means it takes 243 moves against optimal Black play in order for White to be able to safely capture a piece. Qg5 Qxg5+ 6. namely 243. Z Z ZNZ Z Z ZKS ZnZ Z Z Z Z Z Z Z Z Z Z Z Z Z Z ZnZ Z Z ZkZ Z Z Figure 8. but it was far from clear that a win could not be achieved given unlimited play. Kh8 Kh1! Black not only cannot continue checking." Other analyses of mutual zugzwang can be found in Elkies 1996] in this book. a second quiet move in this most tactical of endgames. : : : Qxc6+ 5. Bc6! Not yet 4. 6. : : : Qg2 7. Qfc5+ Kh1 9. Left: mutual zugzwang found by the program. Now Black must take the bishop because 4. But against the quiet 6. but he is still in mortal danger. Pawnless six-piece endgames are extremely rare in tournament play. but must play this modest move to avoid being himself checked to death! For instance. a game between Anatoly Karpov and Gary Kasparov reached the position shown in Figure 5. thereby ensuring an easy win. Qh5+ and the Black king soon perishes from exposure. An exhaustive analysis by the six-piece program Stiller 1991b] proved that it couldn't. Qf2+ mates. Right: Position constructed by Noam Elkies 1993]. and the line of play in Table 3. We conclude this section displaying the best play line for the KRNKNN position with maximal distance-to-win. See Table 3. The initial position is shown in Figure 8. After another fty moves a draw was reached. Qc7+ Kg1 8. Qfg8!!.

28 LEWIS STILLER 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 moves 1{40 Kf7-e6 Nc6-b4 Ke6-e5 Nb4-d3 Ke5-e4 Nd3-f2 Ke4-f3 Nf2-d3 Kf3-e2 Nc2-b4 Ke2-e3 Kb1-b2 Ke3-d4 Nd3-f4 Kd4-c4 Nb4-d5 Rg7-h7 Nd5-e3 Kc4-d4 Ne3-c2 Kd4-e4 Nf4-e6 Ke4-e5 Ne6-g5 Rh7-h5 Nc2-e1 Ke5-f5 Ng5-f3 Kf5-e4 Nf3-d2 Ke4-e3 Nd2-b3 Rh5-h1 Ne1-c2 Ke3-d3 Nb3-c1 Kd3-e4 Nc1-b3 Rh1-h3 Nb3-c5 Ke4-e5 Nc2-e1 Ng8-f6 Ne1-d3 Ke5-d6 Nc5-b7 Kd6-c7 Nb7-c5 Kc7-c6 Kb2-c2 Rh3-h2 Kc2-b3 Kc6-d5 Kb3-b4 Kd5-d4 Nd3-f4 Rh2-h4 Kb4-b5 Nf6-e8 Nc5-b3 Kd4-e4 Nf4-g6 Rh4-h7 Nb3-c5 Ke4-d4 Ng6-f4 Ne8-d6 Kb5-c6 Rh7-h6 Nc5-b3 Kd4-e4 Nf4-e6 Ke4-e5 Ne6-d4 Rh6-h3 Nb3-c5 Nd6-c8 Nd4-c2 Rh3-c3 Nc2-b4 moves 41{80 Ke5-d4 Nb4-a6 Rc3-c2 Kc6-d7 Nc8-b6 Kd7-d6 Nb6-c4 Kd6-c6 Nc4-e3 Kc6-d6 Ne3-f5 Kd6-e6 Nf5-g7 Ke6-f7 Ng7-h5 Nc5-e6 Kd4-e5 Na6-b4 Rc2-e2 Nb4-d3 Ke5-e4 Nd3-b4 Re2-b2 Kf7-g6 Nh5-g3 Ne6-g5 Ke4-d4 Ng5-e6 Kd4-c4 Nb4-a6 Rb2-f2 Ne6-g5 Rf2-f1 Na6-c7 Ng3-e2 Ng5-f7 Ne2-f4 Kg6-g5 Kc4-d4 Nc7-b5 Kd4-c5 Nb5-d6 Nf4-e6 Kg5-g6 Ne6-f8 Kg6-g5 Kc5-d5 Nd6-f5 Rf1-b1 Nf5-g3 Rb1-b7 Nf7-h6 Rb7-g7 Kg5-f4 Nf8-e6 Kf4-f3 Rg7-b7 Ng3-h5 Rb7-b4 Nh5-f6 Kd5-d4 Nf6-h5 Kd4-d3 Nh6-g4 Ne6-g5 Kf3-g3 Ng5-e4 Kg3-h4 Rb4-a4 Nh5-f4 Kd3-d4 Nf4-e6 Kd4-d5 Ne6-f4 Kd5-d6 Nf4-h3 Ra4-a8 Ng4-f2 Ne4-c5 Kh4-g5 moves 81{120 moves 121{160 Kd6-e5 Nf2-g4 Nd6-e4 Ne7-g6 Ke5-d4 Nh3-f4 Ke5-f5 Ng6-f8 Nc5-e4 Kg5-g6 Rh7-h6 Kc6-c7 Ra8-a6 Kg6-f5 Rh6-h1 Nf8-d7 Ra6-a5 Kf5-e6 Rh1-b1 Nd7-b8 Ne4-c5 Ke6-e7 Kf5-e5 Nd5-e3 Ra5-a7 Ke7-f6 Ke5-d4 Ne3-f5 Kd4-e4 Kf6-g5 Kd4-d5 Nf5-e3 Ra7-a5 Nf4-h5 Kd5-c5 Nb8-d7 Nc5-e6 Kg5-g6 Kc5-d4 Ne3-g4 Ra5-b5 Kg6-f7 Rb1-c1 Kc7-d8 Ne6-c5 Kf7-e7 Rc1-e1 Ng4-f6 Rb5-b2 Ke7-d6 Ne4-g5 Kd8-c7 Nc5-b7 Kd6-e7 Ng5-f7 Nd7-f8 Rb2-a2 Nh5-g7 Re1-f1 Nf6-g4 Ra2-e2 Ke7-d7 Rf1-g1 Ng4-f6 Re2-g2 Ng7-e8 Rg1-e1 Kc7-d7 Ke4-f4 Ng4-f6 Kd4-e5 Nf6-e8 Kf4-e5 Kd7-e7 Nf7-h8 Kd7-e7 Rg2-e2 Ke7-d7 Ke5-d5 Ke7-d7 Nb7-a5 Nf6-g4 Re1-f1 Ne8-c7 Ke5-f5 Ng4-h6 Kd5-e5 Nf8-e6 Kf5-g6 Nh6-g8 Nh8-g6 Ne6-c5 Na5-c4 Ne8-c7 Rf1-b1 Kd7-c6 Kg6-f7 Ng8-h6 Ng6-e7 Kc6-d7 Kf7-f6 Nh6-g8 Ne7-f5 Kd7-c6 Kf6-e5 Ng8-e7 Nf5-d4 Kc6-d7 Re2-d2 Kd7-c6 Rb1-d1 Nc7-a6 Rd2-c2 Nc7-a6 Nd4-f5 Kd7-c6 Nc4-e3 Kc6-d7 Rd1-h1 Na6-b4 Rc2-d2 Kd7-c6 Rh1-h6 Kc6-d7 Rd2-d6 Kc6-b5 Ke5-d4 Nc5-e6 Rd6-h6 Ne7-c8 Kd4-c4 Nb4-a6 Ke5-d4 Na6-b4 Rh6-h7 Kd7-c6 Rh6-h5 Kb5-c6 Rh7-h1 Na6-c7 Ne3-c4 Nc8-e7 Rh1-d1 Nc7-e8 Rh5-h6 Kc6-c7 Nf5-e7 Kc6-c7 Rh6-h7 Kc7-d7 Kc4-d5 Ne6-f8 Kd4-e5 Nb4-d5 Ne7-g8 Kc7-d7 Nc4-d6 Kd7-c6 Kd5-c5 Kd7-e6 Table 3. 15. An optimal line of play for the position of Figure 8. but are not shown. Occasional local variations are possible (for instance. . f5-f4).

Nefkens 1985.s is slower than either XR. Per-node time per endgame (time to solve the endgame divided by number of nodes in the state-space) is faster by a factor of approximately 6000 than timings of di erent endgames reported using classical techniques van den Herik and Herschberg 1985b. as is clear from Table 1. The processors were interconnected in a hypercube and clocked at 7MHz (10 MHz for the CM-200).s or XB. Implementation Notes and Run Times The implementation was on a 64K processor CM-2/200 with 8 GBytes RAM. XQ.s ) as well as run-time settable factorization choices and load on the front end. Exact timings depend on S (for instance. van den Herik et al. In unpublished personal communication Thompson has indicated that the per-node time of the fastest serial endgame code is currently only a factor of approximately 700 times slower than that of the code reported in this paper (depending on the endgame) Thompson 1990]. The CM-2 six-piece code required approximately 1200 seconds for initialization and between 111 and 172 seconds to compute Ki+1 from Ki . Thompson 1986.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 29 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 moves 161{180 moves 181{200 moves 201{220 moves 221{243 Rd1-e1 Ke6-d7 Rd5-a5 Nh6-g4 Nh6-f5 Kf8-f7 Kd5-d6 Nc7-e8 Re1-e7 Kd7-d8 Ke5-d4 Kf8-f7 Re1-e2 Nd5-b6 Kd6-e7 Ne8-c7 Re7-a7 Nf8-d7 Ra5-a7 Kf7-f6 Re2-e7 Kf7-f8 Rh7-h6 Na6-b8 Kc5-c6 Nd7-e5 Kd4-e4 Ng7-e8 Re7-e1 Nb6-d5 Na4-b6 Kc8-b7 Kc6-d5 Ne5-g6 Ra7-a6 Kf6-g7 Re1-e5 Nd5-b6 Nb6-c4 Nb8-c6 Ra7-h7 Ne8-c7 Ra6-b6 Ng4-f6 Kg5-g6 Ne8-c7 Ke7-d6 Nc6-b4 Kd5-c6 Ng6-e5 Ke4-f5 Nf6-d7 Nf5-d6 Nb6-d5 Rh6-h8 Nb4-a6 Kc6-d6 Ne5-c4 Nf4-e6 Kg7-f7 Re5-e1 Nc7-e6 Rh8-h7 Kb7-c8 Kd6-c5 Nc4-e5 Ne6-g5 Kf7-f8 Kg6-f5 Ne6-c7 Nc4-a5 Kc8-d8 Rh7-h5 Ne5-f7 Rb6-a6 Ne8-g7 Kf5-e5 Nd5-b4 Na5-c6 Kd8-c8 Kc5-c6 Nc7-e6 Kf5-g6 Nd7-e5 Re1-f1 Kf8-e7 Nc6-e7 Kc8-d8 Rh5-a5 Kd8-e8 Kg6-h7 Ng7-e8 Rf1-f7 Ke7-d8 Ne7-d5 Nc7-e8 Ng8-f6 Ke8-e7 Ra6-e6 Ne5-f7 Nd6-b7 Kd8-c8 Kd6-c6 Na6-b8 Nf6-d5 Ke7-f8 Ng5-f3 Nf7-d6 Nb7-c5 Nc7-b5 Kc6-b5 Ne8-d6 Kc6-d7 Ne6-d4 Kh7-g6 Nd6-f5 Rf7-g7 Kc8-d8 Kb5-c5 Nd6-c8 Nd5-f4 Nf7-h6 Re6-e1 Nf5-e7 Rg7-b7 Nb4-c6 Rh7-h8 Kd8-d7 Ra5-d5 Nd4-f5 Kg6-g5 Kf8-f7 Ke5-e6 Kd8-c8 Nd5-f6 Kd7-c7 Kd7-e6 Nf5-g7 Nf3-e5 Kf7-g7 Rb7-h7 Nc6-b4 Rh8-h7 Kc7-d8 Ke6-f6 Nh6-g8 Ne5-g4 Kg7-f8 Nc5-a4 Nb4-a6 Rh7-b7 Nb8-a6 Kf6-e5 Ng8-h6 Ng4-h6 Ne7-d5 Ke6-d5 Nb5-c7 Kc5-c6 Nc8-e7 Kc6-b6 Na6-b4 Rb7-d7 Kd8-c8 Rd7 Ne7 Table 3 (continuation) 9. 1988] based on the ve-piece timings of the code reported here. .

cs.jhu. six-piece endgames are in fact coming within reach of classical solution techniques. In addition to the work of Babbage. we remark that K. Future Work The main historical open question is to nd out Molien's exact contribution to the history of numerical chess endgame analysis. Kanunov 1983. not currently possible since six-piece endgames could not have been solved in a practicable amount of time using classical techniques on previous architectures. in a little-known 1925 article in Deutsche Schachzeitung. This would permit a more informative timing comparison. Schwarz. although per-node timing comparisons based on radically di erently sized state-spaces are not very meaningful. Berliner and Campbell 1984] and solve endgames with . Zermelo. However. 10. direct comparison of six-piece timing against other work is. Molien. The only program with per-node time of comparable speed to the author's CM-200 implementation is the vectorized implementation of Table 1 by Burton Wendro et al.edu/pub/stiller/snark.2. Bellman 1965. the large per-node timing di erential of the current program compared to classical programs does tend to support the hypothesis that the techniques reported here lend themselves to e cient parallel implementation. From a computational point of view. it might seem that the next logical step in the evolution of the current program should be the exhaustive solution of pawnless seven-piece endgames. Amelung himself is also a gure about whom little is known. Amelung. 6] refers to private papers held by Molien's daughter.30 LEWIS STILLER Unfortunately. and the remarks here would seem to suggest that a detailed reassessment of his contribution to the endgame study would be desirable. In fact. Thus. argued for a postional evaluation function similar to the squares-attacked heuristic used in some full-chess programs Schwarz 1925]. in my opinion a more promising approach would be to follow up on the suggestions rst made by Bellman Bellman 1961. and with enough spare cycles. and to locate and check his analysis of KRKB. 1993]. of course. with larger and faster serial machines. The question of Molien and Amelung's contributions to quantitative endgame analysis is part of the larger historical question of pre-digital precursors to computer chess algorithms. currently we are trying to locate these papers in the hope that they might shed light on the questions raised in Section 2. p. The CM-200 source code implementing Table 1 is currently available from ftp://ftp. although this implementation currently solves only a single four-piece endgame. and Quevedo.

the results of such a search would. Jonathan Schae er. Boris Statnikov. for his help in the formulation and presentation of these ideas. Charles Fiduccia. A. and the sta of the Cleveland Public Library. it is also possible to search for certain classes of endgames considered artistic by endgame composers. Cleveland. Assistance in translating many of the non-English sources was provided by Peter Jansen. George Krotko . Loranth and Motoko B. including the mutual zugzwang recognition algorithm. have considerable impact on the evaluation of many endgames arising in practical play. Steven Salzberg. Murray Campbell. I am also grateful to the following people for useful conversations and suggestions: Alan Adler. Ralph Gasser. Claudia Salzberg. . Los Alamos. Reece of the Cleveland Public Library for their assistance during his visits to the collection. in my opinion. Silvio Levy skillfully and scrupulously edited the manuscript. Harold van der Heijden graciously provided all studies in his database of approximately 36000 endgame studies in which certain pawnless six-piece endgames arose. Thomas Hawkins. and David Waltz. Access to most of the manuscripts. NM 87545. located at the Cleveland Public Library. Richard Draper. Burton Wendro made possible much of the work and developed and implemented the vectorized Cray algorithm. Throughout the six-piece project. Orientalia and Fine Arts. such endgames typically depend on a key mutual-zugzwang or domination position some moves deep in the tree. The National Library of Latvia in Riga (Latvijas Nacionala Biblioteka) assisted in providing copies of a number of articles by Friedrich Amelung. 325 Superior Avenue. Computing facilities were graciously provided by the Advanced Computing Laboratory of the Los Alamos National Laboratory. Michael Klepikov. Even more speculatively. Noam Elkies patiently provided invaluable chess advice and suggestions. Acknowledgments Thanks go to my advisor. Thanks to Alice N. Hans Berliner. Kenneth Whyld helped narrow down the search for them to the city of Riga.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 31 multiple pawns and minor pieces. John Roycroft. Although the use of heuristics would introduce some errors. A. Such an approach would combine heuristic evaluation of node values corresponding to promotions with the exhaustive search techniques described here. John Roycroft provided much of the historical information on human endgame analysis. Simon Kasif. Feng-Hsiung Hsu. OH 44414. rare books and chess journals cited in the paper was obtained during my visit to the John Griswold White Collection of Chess. Ken Thompson.

. An et al. Wagener.32 LEWIS STILLER References Abdali and Saunders 1985] Syed Kamal Abdali and Benjamin David Saunders. especially the endgame of rook and bishop against bishop and knight]. see Acknowledgements). Amelung 1908a] Friedrich Ludwig Amelung. Al-`Adl . 57 (1902) 265{268 (September). \Das Endspiel von zwei O zieren gegen einen Springer" The endgame of two pieces against a rook]. Deutsche Schachzeitung. 20{21. Amelung 1908c] Friedrich Ludwig Amelung. 290{297. Losungspreis-endspiel Nr. \Fast Fourier transforms for space groups containing rotation axes of order three and higher". 261{ 266 (September). 101{105 (April). Deutsche Schachzeitung. An et al. \Factorization method for crystallographic Fourier transforms". Advances in Applied Mathematics 11 (1990). \Auszuge aus den Briefen von A. insbesondere das Endspiel von Thurm und Laufer gegen Laufer und Springer" The endgames with exchange advantage. 87. Amelung 1898] Friedrich Ludwig Amelung. 297{300 (October). McGraw-Hill. Deutsche Schachzeitung. photographic `Adl ] . 330{332 (November). White Collection. 178 Prizeendgame 178]. 358{371. Baltische Schachblatter 4 (1893). Brainerd. Fortran 90 Handbook: Complete ANSI/ISO Reference. Chao Lu. \Die Endspiele mit Qualitatsvortheil. Book of Chess]. Adams et al. Amelung 1908b] Friedrich Ludwig Amelung. and Jerrold L. Y @ d 'Q @ H AJ . Amelung 1902] Friedrich Ludwig Amelung. 55 (1900) 1{5 (January). and Richard Tolimieri. 1992. 134{138 (May). \Transitive closure and related semiring properties via elimination". 37{41 (February). Amelung]. James William Cooley. 346{349. \Das Endspiel von Thurm gegen Springer" The endgame of rook against knight]. 1990] Myoung An. ninth century. Martin. Fur Haus und Familie (supplement to Duna-Zeitung ) 63 (15{28 March 1908). copy of Arabic manuscript (available at the John G. Acta Crystallographica A48 (1992). Theoretical Computer Science 40 (1985). Amelung" Excerpts from the letters of A. and Richard Tolimieri. 198{202 (July). Endspiel 1028. \Das Endspiel von Thurm und Springer gegen die Dame" The endgame of rook and knight against queen]. Baltische Schacblatter 5 (1898). Prince. Solution in 13 (17{30 January 1909). 24(14) (5 April 1908). Amelung 1900] Friedrich Ludwig Amelung. Amelung 1893] Friedrich Ludwig Amelung. 52{53. Amelung 1901] Friedrich Ludwig Amelung. Brian T. Deutsches Wochenschach. 130. 12{38. Fur Haus und Familie (supplement to Duna-Zeitung ) 40 (16{29 February 1908). Jeanne T. Smith. 1992] Myoung An. 1992] Jeanne C. Ascharin an F. Walter S. E. 225{229 (August). 257{274. \Das Endspiel von Turm und Laufer gegen zwei Springer" The endgame of rook and bishop against two knights]. New York. Adams. Ascharin to F. 56 (1901) 193{197 (July).

3. 19 (1990). Longman and Green. 61{74. Berger 1922] Johann Berger. 266{267. 15 April 1957. \The theory of games". Journal of the Institute of Mathematics and its Applications. ferzevye Chess endings: queen]. 1991. 54. Boaz Gordissky. 49{52. Berger 1890] Johann Berger. 15(2) (1975). and Yoav Medan. Technical Report P-1062. Baltische Schachblatter 4 (1893). Backhouse and B. Parallel Computing. Moscow (second edition). The RAND Corporation. 1893] \Baltische Schachpartien aus den Jahren 1858 bis 1892". Fizkul'tura i sport.nding problems". Theorie und Praxis der Endspiele: Ein Handbuch fur Schachfreunde Theory and practice of the endgame: a handbook for chessplayers ]. Averbuch et al. 1982. \Fedor Eduardovich Molin". 1990] Fred S. New York. Technical Report P-473. 7 March 1961. Anonymous] Nicholas de St Nicholai?]. The RAND Corporation. CA. Schachbl. Air Force Project RAND. Proceedings of the National Academy of Sciences of the United States of America. Score of consultation Theodor Molien and A. Technical report. 1990] Amir Averbuch. 465{471 in Passages From the Life of a Philosopher. Bashmakova. Annexstein. CA. \A parallel FFT on a MIMD machine". The RAND Corporation. Vol. Bell. Leipzig. 161{186. \On the reduction of dimensionality for classes of dynamic programming processes". 1{24. 1 June 1954. 1864. 15 (1990). 1739{1740 in Biographical Dictionary of Mathematicians. Eran Gabber. Bell 1978] Alex G. Santa Monica. Santa Monica. Babbage 1864] Charles Babbage. 1978. Backhouse and Carre 1975] Roland C. Berlin and Leipzig (second edition). Carre. 53 (1965). \Ring structure and the Fourier transform". Oxford. \Bonus Socius Good companion]. Partie 114 Baltic chess games from the years 1892 to 1858. Pergamon Chess Series. Xahmatnye okonqani . London. Marc Baumslag. pp. Veit. Hasselblatt v. 1890. The Machine Plays Chess?. Mathematical Intelligencer. G. 1902] Carl Behting and Paul Kerkovius?] \Das zweite Baltische Schachturnier" The second Baltic chess tournament]. Scribner's. Bellman 1954] Richard Ernest Bellman. Bellman 1961] Richard Ernest Bellman. \Group action graphs and parallel architectures". Bellman 1957] Richard Ernest Bellman. Friedrich Amelung. 544{569. A. de Gruyter. \On the application of dynamic programming to the determination of optimal play in chess and checkers". Game 114]. SIAM Journal on Computing. Pergamon Press. Balt. Theorie und Praxis der Endspiele: Ein Handbuch fur Schachfreunde Theory and practice of the endgame: a handbook for chessplayers ]. \Games of skill". 7(3) (1985). \On a new iterative algorithm for nding the solutions of games and linear programming problems". Baltische Schachblatter 9 (1902). 244{246. and Arnold Leonard Rosenberg. Bashmakova 1991] J. 13th century. Schachbl. Balt. U. Santa Monica. pp.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 33 Annexstein et al. Auslander and Tolimieri 1985] Louis Auslander and Richard Tolimieri. CA. Bellman 1965] Richard Ernest Bellman. S. 1922. \Regular algebra applied to path. Averbakh 1982] Yuri Lvovich Averbakh. .

Lille. Diaconis 1988] Persi Diaconis. Springer. Draper 1991] Richard Noel Draper. \The rook and bishop against rook". 297{332. Mannheim. and I. CA. Group Representations in Probability and Statistics. 1982. France. S. Lecture Notes in Computer Science 559. Arti cial Intelligence series].34 LEWIS STILLER Berlekamp et al. Campbell 1988] Murray S. \Complexity starts at ve". MIT Press. Berliner and Campbell 1984] Hans Jack Berliner and Murray S. Guy. \Using chunking to solve chess pawn endgames. 305{311. \Gray codes. 225{233. Jaap van den Herik. Jeremy Gibbons. H. de Rossi ?]. Cooley and Tukey 1965] James William Cooley and John Wilder Tukey. \An overview of supertoroidal networks". with illustrations ]. Herschberg. 1617. Winning Ways for Your Mathematical Plays I. Yves Demailly. July 1990. Del gioco de gli scacchi T he game of chess ]. Chapais 1780] Chapais. Dekker. Blelloch 1990] Guy E. MA. 1952. Technical Report CMU-CS-88-116. \E cient computation of the Fourier transform on nite groups". Chamberlain 1988] R. Chamberlain. Blelloch. Parallel Computing 6 (1988). Vector Models for Data-Parallel Computing. Draper 1990] Richard Noel Draper. Campbell. Science of Computer Programming. 1987] Sito T. The ChessPlayer's Magazine 2 (1864). \Bibliographisches Institut Wissenschaftsverlag. Butler 1992] Gregory Butler. Diaconis and Rockmore 1990] Persi Diaconis and Daniel Nahum Rockmore. Clausen and Baum 1993] Michael Clausen and Ulrich Baum. \An algorithm for the machine calculation of complex Fourier series". Berlin/New York. Bird et al. 1992. Technical report. M. 1989] Richard S. 1993. Maryland. \Chunking as an abstraction mechanism". 297{301. \Formal derivation of a pattern matching algorithm. 1982] Elwyn Ralph Berlekamp. Essais analytiques sur les echecs. Nouveau traite complet d'echecs: La n de partie New complete treatise of chess: The endgame ]. Campbell. 12 (1989). Fundamental Algorithms For Permutation Groups. Trento ?]. Supercomputing Research Center. Carnegie-Mellon University. Lecture Notes|Monograph Series 11. 22 February 1988. Fast Fourier Transforms. 97{120. avec gures Analytical essay on chess. 93{104. 1988. London. Journal of the American Mathematical Society 3 (1990). Arti cial Intelligence 23 (1984). Crosskill 1864] Alfred Crosskill]. Technical Report SRC-TR-91-035. Bowie. Cambridge. 1780]. Hayward. Maryland. \A fast distributed routing algorithm for supertoroidal networks". and hypercubes". fast Fourier transforms. Bowie. Academic Press. . ca. volume 3. International Computer Chess Association Journal 10(3) (September 1987). Supercomputing Research Center. John Horton Conway. 17 January 1991. Dekker et al. 1990. Bird. Institute of Mathematical Statistics. 125{138. Cheron 1952] Andre Cheron. Carrera 1617] Pietro Carrera. and Geraint Jones. and Richard K. Mathematics of Computation 19 (1965).

American Chess Journal (2) 1 (September 1993). 's-Gravenhage. Basic Chess Endings. 1992] John A. IEEE Signal Processing Magazine 9 (1992). \Matrix representations of the multidimensional overlap and add technique". 1343{1382. Ellis Horwood. pp. German translation: Das Endspiel. Beal). Frobenius 1896] Ferdinand Georg Frobenius. Gasser 1996] Ralph Gasser. 1940. Graduate Texts in Mathematics 129. Archive for History of Exact Sciences 7 (1971). F. \Gray codes and paths on the -cube". David McKay. \The origins of the theory of group characters". Gilbert 1958] E. 1991] John A. 1364{1366. ??{?? in this volume. \Computation of the mean residence time of water in the hydration shells of biomolecules". 48{52. 1957. Hamburg. Granata. and Richard Tolimieri. L. 289{90. F. Hawkins 1971] Thomas William Hawkins. 1976. Bell Systems Technical Journal 37 (1958). Elkies 1993] Noam David Elkies. van Goor Zonen. Das Schach-Archiv. IEEE Transactions on Circuits and Systems 38 (1991). 1624. Golombek 1976] Harry Golombek. Granata and Richard Tolimieri. B. Fine 1941] Reuben Fine. pp. 52. England. Elkies 1996] Noam David Elkies. \A tensor product factorization of the linear convolution matrix". New York. Garc a and Lewis Benjamin Stiller. pp. G. Rattman. Fur Haus und Familie (supplement to Duna-Zeitung ) 70 (28 March{10 April 1909). New York. Chess: A History. 1991. Granata et al. Michael Conner. 815{826. Het Eindspel The endgame ]. New York. ??{?? in this volume. \Applying retrograde analysis to Nine Men's Morris". Springer. Representation Theory: A First Course. Michael Conner. \On numbers and endgames: Combinatorial game theory in chess endgames". Gasser 1991] Ralph Gasser. L. 161{173 in Heuristic Programming in Arti cial Intelligence: The Second Computer Olympiad (edited by D. N. P. 76{78. 1991. IEEE Transactions on Circuits and Systems for Video Technology 1 (1991). G. N. 142{170. Fulton and Harris 1991] William Fulton and Joe Harris. Granata. n . Trattato del nobilissimo et militare essercitio de scacchi nel qvale si contengono molti bellissimi tratti et la vera scienza di esso gioco Treatise on the very noble and military exercise of chess: the science of that game ]. Levy and D. \The tensor product: a mathematical programming language for FFTs and other fast DSP operations". \Chess art in the computer age". 40{48. Gilbert. Greco 1624] Gioacchino Greco. Putnam's Sons. Duna-Zeitung 1909b] \Friedrich Ludwig Amelung". Euwe 1940] Machgielis Euwe. Sitzungsberichte der Koniglich Preu ischen Akademie der Wissenschaften zu Berlin (1896). Garc a and Stiller 1993] Angel E. Granata et al. Chichester. \Uber die Primfactoren der Gruppendeterminante" On prime factors of group determinants]. 1941. 1396{1406. and Richard Tolimieri.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 35 Duna-Zeitung 1909a] Announcement of Amelung's death in Fur Haus und Familie (supplement to Duna-Zeitung ) (10{23 March 1909). Granata and Tolimieri 1991] John A. \Solving Nine Men's Morris". Journal of Computational Chemistry 14 (1993).

243{287. Jansen 1992c] Peter Jozef Jansen. \A data base on data bases". Also published as Tech. International Computer Chess Association Journal 9(1) (March 1986). \Embedding meshes in Boolean cubes by graph decomposition". \Computer checks on human analyses of the KRKB endgame". 26{31. 325{339. Stanford University Department of Computer Science. Hawkins 1974] Thomas William Hawkins. \International Computer Chess Association Journal 11(1) (March 1988). Hendriks. S. Hillis 1985] William Daniel Hillis. \Exploiting symmetry in high-dimensional nite-di erence calculations". \KQKR: Awareness of a fallible opponent". Press.36 LEWIS STILLER Hawkins 1972] Thomas William Hawkins. Cambridge. and J. Computer Architecture and Parallel Processing. van den Herik and Herschberg 1986] H. \Hypercomplex numbers. Jansen 1992b] Peter Jozef Jansen. 1968. I. September 1992. 1992. International Computer Chess Association Journal 15(3) (September 1992). Herschberg. Technical Report CS 106. McGraw-Hill. Journal of Parallel and Distributed Computing 8 (1990). 217{ 143. Herschberg. Archive for History of Exact Sciences 12 (1974). van den Herik and Herschberg 1985b] H. Jaap van den Herik. Hwang and Briggs 1985] Kai Hwang and Faye Alaye Briggs. 66{87. thesis. \New light on Frobenius' creation of the theory of group characters". 1988] H. van den Herik et al. The Connection Machine. Jaap van den Herik and I. Lie groups.D. Jaap van den Herik and I. Report CMU-CS92-192. Herschberg. New York. S. Ho and Johnsson 1990] Ching-Tien Ho and S. McGraw-Hill Series in Computer Organization and Architecture. Whyld. International Computer Chess Association Journal 15(4) (December 1992). International Computer Chess Association Journal 8(2) (June 1985). S. Oxford Univ. T. \The construction of an omniscient endgame database". Hillis and Taylor 1990] William Daniel Hillis and Washington Taylor IV. Jansen 1992a] Peter Jozef Jansen. Huberman 1968] Barbara Jane Huberman. Lennart Johnsson. van den Herik and Herschberg 1985a] H. Journal of Parallel and Distributed Computing 8 (1990). \KQKR: Assessing the utility of heuristics". 77{79. a conjecture refuted". and the creation of group representation theory". \Elementary theory improved. Ph. Stanford Arti cial Intelligence Project Memo AI-65. Herschberg. \Using Knowledge about the Opponent in GameTree Search". 1985. Jaap van den Herik and I. P. Wit. . Pittsburgh. PA. K. 111{ 131. The Oxford Companion to Chess (second edition). MA. 1985. \A program to play chess end games". S. International Computer Chess Association Journal 8(3) (September 1985). Carnegie Mellon University Department of Computer Science. 179{ 191. 141{149. R. Archive for History of Exact Sciences 8 (1972). Hooper. ACM Distinguished Dissertations. 29{34. MIT Press. Hooper and Whyld 1992] D.

. and Jeremey R. ACM Press. king and rook (KRBKR)". 1994] Robert Lake. New York. Domingo Rodriguez. modifying and implementing Fourier transform algorithms on various architectures". pp. Lenz 1970] Wilhem Lenz. Komissarchik and Futer 1974] E. Circuits. Jonathan Schae er. Skeet. \A methodology for designing. Johnson. Lake et al. Problemy Kibernetiki 29 (1974). 1992. 1988). Johnson 1989] Robert W. Theoretical Computer Science 4 (1977). \Computing the fast Fourier transform on a vector computer". C. A. 977{992. Mathematics of Computation 33 (1979). Kopec et al. Johnson. 1970. and Richard Tolimieri. Fedor duardoviq Molin 1861{ 1941 Fyodor Eduardovich Molin 1861{1941]. \The network architecture of the CM-5". and Chris Cook. Kling and Horwitz 1851] Josef Kling and Bernhard Horwitz. University of Limburg. Kaufmann. 1994. \Chess Studies.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 37 Johnson et al. rook and bishop vs. Lambiotte. \Algebraic structures for transitive closure". \Fast Fourier transforms on nite nonAbelian groups". pp. . Komissarchik and Aaron L. 1990] Jeremy R. and Signal Processing 9 (1990). manuscript. Futer. 272{285 in Proceedings of the 4th Annual ACM Symposium on Parallel Algorithms and Architectures (San Diego.). Korn and Jules J. 1988] Danny Kopec. 1992). England. and Paul Lu. or Endings of Games. \Ob analize ferzevogo ndxpil pri pomowi VM Analysis of a queen endgame using an IBM computer]. 60{61 in Proceedings: 1988 Spring Symposium Series: Computer Game Playing (Stanford. Deutschbaltisches biographisches Lexikon 1710{1960 German-Baltic biographical dictionary 1710{1960 ]. pp. Netherlands. 1992] Charles Eric Leiserson et al. Kanunov 1983] Nikolai Fedorovich Kanunov. Containing Upwards of Two Hundred Scienti c Examples of Chess Strategy. 1993] Shivnandan D. Korn and Lambiotte 1979] David G. and Chua-Huang Huang. Chua-Huang Huang. Charing Cross. M. \The endgame king. van den Herik et al. 1{26. San Mateo. Bohlau. J. editor. Introduction to Parallel Algorithms and Architectures: Arrays. 1992. Leiserson et al. Maastricht. Johnson. Leighton 1992] Frank Thomson Leighton. Karpovsky 1977] Mark Girshevich Karpovsky. 189{217. Hypercubes. 59{76. 24 April 1989. Kopnin 1983] Aleksey Grigoryevich Kopnin. Kaushik et al. Robert W. \Solving large retrograde analysis problems using a network of workstations". Nauka. 1983. Johnson et al. \Some special features of the endgame struggle Rook and Knight against 2 Knights (GBR class 0107)". J. Brent Libby. 1851. Koln. Sanjay Sharma. 211{220. Journal of Information Science and Engineering 9 (1993). Jr. American Association for Arti cial Intelligence. Amsterdam) 5(#70) (January 1983). Johnson. 89{92. Journal of Supercomputing 5 (1991). \Multilinear algebra and parallel programming". 1028{1030. Systems. EG (ARVES. 1991] Robert W. Moscow. Johnson. Lehmann 1977] Daniel J. Kaushik. CA. Lehmann. Trees. 135{162 in Advances in Computer Chess 7 (edited by H. \Automatic implementation of tensor products". IEEE Transactions on Computers C-26 (1977). \An algebraic theory for modeling multistage interconnection networks". 449{500.

International Computer Chess Association Journal 10(1) (March 1987). 211{224. Merris 1981] Russell Merris. American Mathematical Monthly 88 (March 1981). ca. 83{156. 1985. \Pattern inventories associated with symmetry classes of tensors". . Repeticion de Amores y Arte de Ajedrez Treatise on love and the game of chess ]. Marcus 1973] Marvin Marcus. Merris and Watkins 1983] Russell Merris and William Watkins.38 LEWIS STILLER Lolli 1763] Giovanni Battista Lolli. 534{547. 208{209 in \Sammlung baltischer Schachprobleme und Endspiele aus den Jahren 1890 bis 1897". Baltische Schachblatter 6 (1898). errata in 42 (1893). Salamanca. 1973. Sohn]. 1953. pp. Sitzungsberichte der ber Naturforscher-Gesellschaft Dorpat (Yurev. Molien 1901] Theodor Molien. Beograd. \Ideas on knowledge synthesis stemming from the KBBKN endgame". \Baltische Schachblatter 7 (1900). New York. Merris 1992] Russell Merris. Molien 1898b] Theodor Molien. Part 1. Merris 1980] Russell Merris. Facsimile reprint Coleccion Joyas Bibliogra cas. April 1897. 1561. Partie 73 score of game against W. Bologna. 1763. \Uber die Anzahl der Variabeln einer irreductibelen Substitutionsgruppen" On the number of variables of irreducible substitution groups]. Berlin (1898). Alcala. Molien 1898a] Theodor Molien. Sahovski informator. Madrid. Estonia) 11 (1897). Linear Algebra and Applications 29 (1980). \Uber die Invarianten der linearen Substitutionsgruppen" On the invariants of linear substitution groups]. \Ueber Systeme hoherer complexer Zahlen" On systems of higher complex numbers]. Finite Dimensional Multilinear Algebra. 1152{1156. 179{ 212. Matanovic 1985] Aleksandar Matanovic. Estonia) 11 (1897). Michie and Bratko 1987] Donald Michie and Ivan Bratko. 3{10. 179{185. 1497. Cited in the index of Baltische Schacblatter 8 (1901). Sitzungberichte Naturforscher-Gesellschaft Dorpat (Yurev. Molien 1897a] Theodor Molien. Molien 1900] Theodor Molien. Mathematische Annalen. \Tensors and graphs". \Applications of multilinear algebra". 346. \Polya's counting theorem via tensors". Molien 1897b] Theodor Molien. Enciklopedija Sahovskih Zavrsnica Encyclopedia of Chess Endings ]. Lucena 1497?] Luis Ramirez de Lucena. Lopez 1561] Ruy Lopez de Sigura. Sitzungsberichte Akademie der Wissenschaft. Linear and Multilinear Algebra 32 (1992). \Zi ermassig genaue Ausrechnung aller 12 millionen Gewinne und Remisen im Endspiel "Thurm gegen Laufer" Numerical exact computation of all 12 million wins and draws in the endgame \rook against bishop"]. \Eine Bemerkung zur Theorie der homogenen Substitutionsgruppen" A remark on the theory of homogeneous substitution groups]. 308{312. 277{288. Marcel Dekker. Libro de la invencion liberal y arte del juego del axedrez The book of liberal invention and the art of the game of chess ]. 225{230. 41 (1893). SIAM Journal on Algebraic and Discrete Methods 4 (1983). 259{274. Molien 1893] Theodor Molien. page 72. Osservazione teorico-pratiche sopra il giuoco degli scacchi Theoretical-practical observations on the game of chess ].

21{25. Computer and Information Science Department.). Journal of the Association for Computing Machinery 15 (1968). 1929. \Qubic: 4 4 4 tic-tac-toe". Rasmussen 1988] Lars Rasmussen. Nunn 1992] John Nunn. Journal of the Association for Computing Machinery 29 (1982). Tioga. Technical Report COINS 88-84. 463{482 in Machine Learning: An Arti cial Intelligence Approach (edited by Ryszard Stanislaw Michalski et al. New York. Rinck 1950] Henri Rinck. \Ultimates in KQKR and KQKN". Philidor 1749?] Francois-Andre Danican Philidor. Holt. EG (ARVES. H. London. Mag. Alfonso the Wise ]. \An adaptation of the fast Fourier transform for parallel processing". Syracuse University. Madrid. Roycroft 1967] Arthur John Roycroft. Edinburgh University Press. Oxford University Press. December 1988. Wiley. Amsterdam) 1 (April 1967). Bruxelles. New York. pp. Palo Alto. 1950. 53(4) (September 1980). CASE Center. 1992. 642{667. Quinlan 1983] John Ross Quinlan. . University of Massachusetts at Amherst. London. 1944. 197{198. Nefkens. 1913. El Ajedrez de D. L'analyze des echecs: contenant une nouvelle methode pour apprendre en peu temps a se perfectionner dans ce noble jeu The analysis of chess: containing a new method for perfecting oneself in a short time in this noble game ].MULTILINEAR ALGEBRA AND CHESS ENDGAMES 39 Mullin 1988] Lenore M. La Academica. 217{224. Patashnik 1980] Oren Patashnik. 1994. Restifo Mullin. Theory of Games and Economic Behavior. H. \A note on 2S's v R + B". International Computer Chess Association Journal 8(4) (December 1985). \Parallel permutation and sorting algorithms and a new generalized connection network". A History of Chess. pp. L'Echiquier. 1979. \A mathematics of arrays". Nunn 1994] John Nunn. Perez 1929] Juan Bautista Sanchez Perez. New York. Rinck and Malpas 1947] Henri Rinck and Louis Malpas. von Neumann and Morgenstern 1944] John von Neumann and Oskar Morgenstern. Nefkens 1985] Harry J. Dame contre tour et cavalier Queen against rook and knight ]. Rosenberg 1988] Arnold Leonard Rosenberg. Quinlan 1979] John Ross Quinlan. 11 October 1988. 252{264. Technical Report 8814. Batsford Chess Library. Holt. \Constructing data bases to t a microcomputer". Alfonso el Sabio The chess of D. Murray 1913] Harold James Ruthven Murray. \Discovering rules by induction from large collections of examples". 1749?]. Secrets of Pawnless Endings. Nassimi and Sahni 1982] David Nassimi and Sartaj Kumar Sahni. International Computer Chess Association Journal 11(1) (March 1988). 1983. Math. Secrets of Rook Endings. 1947. Pease 1968] Marshall Carleton Pease. CA. 202{216. Batsford Chess Library. \Learning e cient classi cation procedures and their application to chess end games". Barcelona. 1414 Fins de partie 1414 endgames ]. Franco-Espa~ nola. 168{201 in Expert systems in the micro-electronic age (edited by Donald Michie). \Shu e-oriented interconnection networks".

Berlin. Stamma 1737] Philip Stamma. 1993. 1972. 1993. England. Los Alamitos. EG (ARVES. Wiley. Singleton. 110{115 in Proceedings of the 1993 Workshop on Programming Models for Massively Parallel Computers. Los Angeles. Communications of the ACM 10 (1967). ACM Press. 647{654. pp. \Group graphs and computational symmetry on massively parallel architecture". \Structuring data parallelism using categorical data types". pp. and Duane Szafron. \Versuch eines mathematischen Schachprinzips" An attempt toward a mathematical principle for chess]. Journal of Supercomputing 5 (1991). Test Tube Chess: A Comprehensive Introduction to the Chess Endgame Study. Roycroft 1984] Arthur John Roycroft. Schae er et al. IEEE Computer Society Press. \Bit-serial SIMD on the CM-2 and the Cray-2". London. Paul Lu. \Solving the Game of Checkers". Sims 1994] Charles Co n Sims. 164{170. \A World Championship caliber checkers program". Stiller 1989] Lewis Benjamin Stiller. 1634. Trattato dell'inventione et art liberale del gioco di scacchi Treatise of the invention and liberal art of the game of chess ]. Skillicorn 1993] David Benson Skillicorn. Deutsche Schachzeitung 53(11) (November 1925). 3: Almost Miniatures: 555 Studies With Eight Chessmen. April 1984). Arti cial Intelligence 53 (1992). Essai sur le jeu des echecs Essay on the game of chess ]. Sims 1971] Charles Co n Sims. Schae er and Lake 1996] Jonathan Schae er and Robert Lake. \Parallel analysis of certain endgames". Computation With Finitely Presented Groups. New York 1971. Norman Treloar. Is Man a Robot?. Simons 1986] Geo Leslie Simons. Chichester.40 LEWIS STILLER Roycroft 1972] Arthur John Roycroft. 99{117. \A proposed revision of the `50-move rule': Article 12. University of Limburg. England. CA. \On computing the fast Fourier transform". \Two bishops against knight". \American Mathematical Monthly 74 (March 1967). 1992] Jonathan Schae er. International Computer Chess Association Journal 7(3) (September 1984). 1994. Journal of Parallel and Distributed Computing 11 (1991). \The group of automorphisms of the game of 3-dimensional Ticktacktoe". Cambridge University Press.4 of the Laws of Chess". vol. 1737. \Computations with permutation groups". 273{290. Brent Knight. 135{145. International Computer Chess Association Journal 12(2) (June 1989). 55{64. Selected Endgame Studies. Singleton 1967] Richard C. Paris. Rusinek 1994] Jan Rusinek. H. Joseph Culberson. Encyclopedia of Mathematics and its Applications 48. Cambridge. 23{28 in Proceedings of the Second Symposium on Symbolic and Algebraic Manipulation. Napoli. pp. 247{254. Stiller 1991a] Lewis Benjamin Stiller. 1971 (edited by Stanley Roy Petrick). Roycroft 1984] Arthur John Roycroft. Smitley 1991] David Smitley and Kent Iobst. 1986. Schwarz 1925] K. 321{324. Schwarz. Faber and Faber. . Salvio 1634] Alessandro Salvio. Amsterdam) 5 (#75. 1994. ??{?? in this volume. 249. Silver 1967] Roland Silver.

\Is there a future for functional languages in parallel programming?" pp. Tolimieri et al. and Chao Lu. Szymanski. and James E. \Karpov and Kasparov: the end is perfection". Szymanski et al. and Robert K. Vivek Sarkar. Thompson 1983] Kenneth Lane Thompson]. IEEE Computer Society Press. Byrnes and Jennifer L. S. Thompson 1990] Kenneth Lane Thompson. 198{ 201. Ph. Luke L. 1993. \International Computer Chess Association Journal 14(4) (December 1991). New York. Myoung An. Dordrecht and Boston. Los Alamos National Laboratory. Thompson 1986] Kenneth Lane Thompson. thesis. Stone 1993] Harold Stuart Stone. Reading. Tarjan 1981b] Robert Endre Tarjan. \ -body simulations on massively parallel architectures". Scienti c American 1915] \Torres and his remarkable automatic devices. 577{593. pp. 1989. 1995. and Chao Lu. Stiller 1994] Lewis Benjamin Stiller. thesis. Springer. Department of Computer Science. EG (ARVES.D. February 1970. Myoung An. Tarjan 1981a] Robert Endre Tarjan. \An algebraic paradigm for the design of e cient parallel programs". 1994. 299{304 in Proceedings of the 1994 International Conference on Computer Languages. Toulouse.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 41 Stiller 1991b] Lewis Benjamin Stiller. 1993] Richard Tolimieri. Algorithms for Discrete Fourier Transform and Convolution. James Hicks. Gubernatis. 1989] Richard Tolimieri. Tolimieri et al. MA. Yates. \Exploiting Symmetry on Parallel Architectures". R. Baltimore. CA. \A uni ed approach to path problems". Journal of the Association for Computing Machinery 28 (1981). 53 (November 1915). 237{250 of Recent advances in Fourier analysis and its applications (edited by J. Journal of Computational Physics 115 (1994). Journal of the Association for Computing Machinery 28 (1981). Jagannathan. Springer. November 1992. \Untersuchungen uber Kombinatorische Spiele" Investigations of combinatorial games]. AddisonWesley. Stiller 1995] Lewis Benjamin Stiller. Technical Report LA-UR-92-5211. n . personal communication. Amsterdam) 5(74) (November 1983). Department of Computer Science. Los Alamitos. Stiller 1992b] Lewis Benjamin Stiller. Stiller 1992a] Lewis Benjamin Stiller. Fakultat fur Allgemeine Wissenschaften der Technischen Hochshule Munchen. Daemen. Kluwer. 131{139. Baltimore. Tolimieri and An 1990] Richard Tolimieri and Myoung An. New York. Technical Report JHU-92/26. \Computations in Xray crystallography". \An algebraic foundation for Fortran 90 communication intrinsics". 550{552. High-Performance Computer Architecture. The Johns Hopkins University. August 1992. David Benson Skillicorn. 296{298. \Retrograde analysis of certain endgames".D. \Fast algorithms for solving path problems". International Computer Chess Association Journal 9 (1986). Byrnes) NATO ASI Series C: Mathematical and Physical Sciences 315. The Johns Hopkins University. 594{614. \Scienti c American Supplement. April 1990. 1994] Boleslaw K. 1993. Strohlein 1970] Thomas Strohlein. Ph. 1990. Mathematics of Multidimensional Fourier Transform Algorithms.

and Ralph Brickner. \Dva kon protiv pexek (teoretiqeski$ i oqerk)" Two knights against pawn (theoretical essay)]. New York. Berlin/New York. Moscow. \Multilinear enumerative techniques. Groups. Publications du CNRS. ::: Lewis Stiller Department of Computer Science The Johns Hopkins University Baltimore. pp. Tony Warnock. van. 1934. MA. Elsevier.com . Cambridge. Cambridge. \Uber eine Anwendung der Mengenlehre auf die Theorie des Schachspiels" On an application of set theory to the theory of playing chess].xo. Texts in Applied Mathematics 16. England. 1984. MIT Press. MD 21218 stiller@symmetry. Colloques internationaux du Centre National de la Recherche Scienti que 37. Applied Numerical Mathematics 12 (1993). 1994.42 LEWIS STILLER Torres-Quevedo 1951] Gonzalo Torres-Quevedo. Die Genesis des abstrakten Gruppenbegri es: ein Beitrag zur Entstehungsgeschichte der abstrakten Gruppentheorie. 501{504 in Proceedings of the Fifth International Congress of Mathematicians. pp. 1913. White 1975] Dennis Edward White. Valiant. Graphs. 945{971 in Handbook of Theoretical Computer Science: Algorithms and Complexity. Troitski 1934] Alekseii Alekseevich Troitski . 1953. Berlin. English translation by Abe Shenitzer: The Genesis of the Abstract Group Concept: A Contribution to the History of the Origin of Abstract Group Theory. pp. Concurrent Scienti c Computing. Fizkul'tura i turizm. Wendro et al. 1912. 1984. \General purpose parallel architectures". pp. 341{352. Dean Mayer. Paris. 1951. 1985. 285{295. Cambridge University Press. Communications of the ACM 33(8) (1990). White 1984] Arthur Thomas White. Amsterdam. 2 (edited by Ernest William Hobson and Augustus Edward Hough Love). van der Waerden 1985] Bartel Leendert van der Waerden. and Surfaces. Valiant 1990a] Leslie G. Springer. van de Velde 1994] Eric F. volume A (edited by Jan van Leeuwen). Valiant. A History of Algebra: From Al-Khwarizm to Emmy Noether. \A bridging model for parallel computation". \Presentation des appareils de Leonardo Torres-Quevedo" Presentation of the machines of Leonardo TorresQuevedo]. North-Holland/ Elsevier. \Linear and Multilinear Algebra 2 (1975). van de Velde. 1969. vol. 103{111. 1993] Burton Wendro . \Bits and pieces: constructing chess endgame databases on parallel and vector architectures". Valiant 1990b] Leslie G. 248{288 in Sbornik xahmatnyh t dov Collection of chess studies ]. von ] See under last name. 1990. Springer. Deutscher Verlag der Wissenschaften. Zermelo 1913] Ernst Zermelo. 383{406 in Les machines a calculer et la pensee humaine. Wussing 1969] Hans Wussing. Paris. Lewis Benjamin Stiller. New York.

Sign up to vote on this title
UsefulNot useful