This action might not be possible to undo. Are you sure you want to continue?

**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

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

in which he investigated non-commutative multiplication and obtained important general results of which the properties of the group determinant are special cases. the classic text Wussing 1969] barely mentions him." . I have bene ted from Hawkins' translation in supplying mine. 1897a. brie y red). has independently of me considered the group determinant. Bande der Mathematischen Annalen eine sehr schone. Theodor Molien in Dorpat. before Molien moved to Tomsk.2 (This letter of February 24. and despite Frobenius' support. He was president of the Dorpat chess club. at least in one case. With the publication of Hawkins' series of articles 1971. da sich ein jungerer Mathematiker Theodor Molien in Dorpat unabhangig von mir mit der Gruppendeterminante beschaftigt hat. worin er die nicht commutative Multiplication untersucht hat und wichtige allgemeine Resultate erhalten hat. because the committee considered his work too theoretical and without practical applications Kanunov 1983. May 6. for example. pp. Molien was rejected from a number of Russian academic positions. aber schwer zu lesende Arbeit `Ueber Systeme hoherer complexer Zahlen' vero entlicht. Molien's work was unknown or underestimated in the West for a long while. Frobenius wrote to Dedekind: You will have noticed that a young mathematician. 1897b. 206{209. 1898]. and more recent historical appraisals gives him due credit van der Waerden 1985. von denen die Eigenschaften der Gruppendeterminant specielle Falle sind. 1972. were obscure and di cult to understand. They anticipated Frobenius' classic paper on the determinant of a group-circulant matrix Frobenius 1896]. it is quoted here with the permission of Springer-Verlag. Not mentioned in Kanunov's biography is that. despite their importance. 35{36]. 1898).) Despite these results. pp. 237{238]. in a transcription made by Walter Kaufmann-Buhler. a fact that Frobenius readily admitted Hawkins 1974]. He has published : : : a very beautiful but hard-to-read work \On systems of higher complex numbers". 1898 was kindly supplied by Thomas Hawkins. His remaining mathematical work had little in uence and he spent most of his time teaching. partly because of the Czarist politics of the time (according to Kanunov) and. After studying medieval mathematical manuscripts at the Vatican Library in 1899. Er hat im 41.6 LEWIS STILLER Molien's early papers on group representation theory 1893. Referring to Molien 1893]. and several of his games were published in 2 \Sie werden bemerkt haben. 1995). he accepted a post at the Tomsk Technological Institute in Siberia. he was one of the strongest chess players in Dorpat and was particularly known for his blindfold play (Ken Whyld. 1974] on the history of group representation theory. personal communication. As a consequence. although he had tremendous di culty understanding Molien's work (letter to Alfred Knezer. where he was cut o from the mathematical mainstream and became embroiled in obscure administrative struggles (he was. the signi cance of his work became better-known. in fact.

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

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

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

1990b. One common form of interconnection network is the hypercube Leighton 1992]. First. Valiant 1990a. many trade-o s that can be made in the architecture of a computer. 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 @ . (3. The winning procedure proved to be quite di cult for humans to understand Michie and Bratko 1987].. It is also possible to perform arbitary permutations on a hypercube. Two processors connected by an edge can communicate directly in a single timestep. Place each processor at a distinct vertex of the unit cube in Euclidean n-space. of course. in fact. C A v8 . each with some local memory. such as the type of interconnection network. The hypercube can compute the e ect of the end-o shift matrix E8 . 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. Stone 1993].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]. and imagine each edge of the cube as a wire that directly connects the processors at its endpoints. 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]. Applied to an array. A @ . In practice general . E8 shifts each element of the array down. Each processor in an n-dimensional hypercube can be viewed as having a length n binary address.. 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. 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.10 LEWIS STILLER twofold. many more moves were required to win certain endgames than had previously been thought. Higher-dimensional shift patterns are also easy to compute on a hypercube Ho and Johnsson 1990]. Blelloch 1990]. 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. 3. There are. v7 . the granularity of the processors. In particular. won. see Figure 1. and whether each processor executes the same instruction (SIMD) or di erent instructions (MIMD) Hillis 1985. Consider a parallel computer with 2n processors. Second.

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

The e ect of Pln on a vector is to stride through the vector. taking m steps of size l. A . l = 2. where v 2 Vm and w 2 Vl . we have n = B A: Pln (A B )Pm This theorem. and n = ml.2 that some evaluation orders naturally correspond to vectorization. taking m = 3.12 LEWIS STILLER n0 nn0 If A 2 Mn m and B 2 Mm0 . C 0 1 0 1 2 1 1 ( 1) (4.... the matrix of the tensor product A B 2 Mmm0 is given by 0 1 BA B A B A B=B B . 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. For example.. 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 .. Anm B C A (In 0 0v 11 B B B AC @ .... 1991]: An1 B An2 B . The n-point stride l permutation matrix Pln is the n n matrix de ned by Pln (v w) = w v . and n = 6. the Commutation Theorem will be the method by which one type of execution is traded o for another.1. . B 2 Ml .... C B = B) B C B C C v m? @ .. . for B 2 Mm Johnson et al. A1m B A2m B C C . We shall see in Section 4. C B . 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..1) Suppose n = ml. @ . C B C B v C BB @ . 1989]: Theorem 4. . allows the order of evaluation in a tensor product to be varied. v nm? C B C B 1 0 B v n? m C C B B A @B B AC @ . If A 2 Mm . A B C B . which is easy to prove even when the entries are from a semiring.

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

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

This generalization has been used for expressing graph algorithms Backhouse and Carre 1975. in the natural way. each of length l. We write + and for _ and ^.2). Bird et al. the Fm Il term performs an m-point FFT on vectors of size l. and matrix sum remain essentially unchanged. Imagine. and the vectorized Korn{Lambiotte FFT Korn and Lambiotte 1979] can be derived by unrolling (4. 1989]. The commutation theorem can be used to derive a parallel form n (Il Fm ) P n Tl (Im Fl ) P n . Pm performs m l-point FFTs in parallel on each segment. . many different FFT algorithms have been derived. 1. and the relation between and parallelization still holds. Given an input n v forms a list of m segments. 1981b]. These ideas could. Granata et al. matrix product. with di erent tradeo s between parallelization and vectorization Chamberlain 1988. Application to Chess This section describes the chess endgame algorithm in a generalization of the tensor product formalism described in Section 4. be presented categorically using the approach of Skillicorn 1993. Fn =Pm m l (4. the commutation theorem. as does. Averbuch et al. the notion of code(M ).3). : : : . By using the commutation theorem and varying the factorization. or using the mathematics-of-arrays formalism of Mullin 1988]. Abdali and Saunders 1985. Tl just multiplies each element by a twiddle factor. of course. Johnson 1989.2) (4. Auslander and Tolimieri 1985]. !. _. In particular. therefore. Mn m . Tarjan 1981a. 1990. that a matrix over GF2 is actually a Boolean matrix whose entries are taken from the Boolean algebra f0. 1991. Finally. ^g. The notion of linear transformations then changes. Lehmann 1977.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 15 where Tl is a diagonal matrix encoding the twiddle factors: Tl = m ?1? M j =0 diag(1. 5. The Im Fl term vector v. The de nitions of . for the moment.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. !l?1 ) j : This can be interpreted as a mixed parallel/vector algorithm.

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. De nitions. 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). The ideas of Section 4. Any element of C is a sum of distinct basis elements. B.s is the function that. each of which is formed from pieces of S . Each basis element of C is of the form ei ej for 1 i. and thereby becomes an element of M64k . captures.1. j 8. and therefore corresponds to a set of squares White 1975]. is associated with a unique element of B : the sum of the basis elements corresponding to each of the positions from the set. k Let B C be the hyperboard corresponding to S . Q. We write S = hS1 . Xp.s in terms of primitive operators. we need not worry about this distinction. We denote by Vn the space of length-n Boolean vectors. denotes a unique square on the 8 8 chessboard. K.s can be extended to a linear function from elements of B to itself. For example. For simplicity of exposition. : : : . and j V the j -th tensor power of V. Any set of positions. Any such basis element. C V V: N Let fei gi N be the standard basis for V . Each basis element of B is of the form c1 c2 ck . the unmove operators are only quasilinear. R. each basis element of B can be thought of as a sequence of k squares of the chessboard.2 may then be used to derive e cient parallel code from this factorization. This correspondence between sets of chess positions and elements of B forms the link between the chess algorithms and the tensor product formulation. Each position that is formed from the pieces of S is thereby associated with a unique basis element of B. castling. and thus B is not a module.16 LEWIS STILLER 5.) The core of the chess endgame algorithm is the e cient computation of the Xp. It can be thought of as k . R. the unmove operator Xp. q. where each cs is some basis element of C. 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. Let S be an ordered set of k chess pieces. pawns. ri: An S -position is a chess position that contains exactly the k pieces in S .s . In the following. (Technically. if k = 6 one could choose S = hk. 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. therefore. stalemates. Ng is a piece. If p 2 fK. Since each cs is a square on the chessboard. Q. However. and the fty-move rule will be disregarded unless otherwise stated. S2 . We now describe a factorization of Xp. since the Boolean algebra is not a ring.

D4 . Section 6.1 gives the fundamental factorization.1. R G x= X g2G gx: 6. We introduce a few group actions Fulton and Harris 1991]. It is a xed point of the G action: g Gx = RGx for all g 2 G. for s 2 Sk and cs 2 C. Merris and Watkins 1983].2 describes the modi cation of the equations of Table 1 to exploit symmetry. 1981. p.3 describes the control structure of the algorithm.MULTILINEAR ALGEBRA AND CHESS ENDGAMES 17 5. s=1 s s=1 ss The dihedral group of order 8. r rotates the chessboard counterclockwise 90 and f ips the chessboard about the horizontal bisector. We let Z to The notation G x is intended to represent the group average of x withR respect G Fulton and Harris 1991. We will use the group-theoretic terminology both to give concise descriptions of certain move operators and to describe the exploitation of symmetry. 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. 6. Factoring the unmove operator. Recall from (3. Group actions. is the group of symmetries of the square. Section 6.2. A group G acting on Vn and Vm acts on g(M g?1(v)). combinatorial enumeration. 6]. 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. and group actions which motivates much of this section Merris 1980. 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.1) that E8 was de ned to be the unit one-dimensional 8 8 end-o shift matrix. 1992. . The symmetric on Sk acts on B by permuting the order of N group Nkk elements the factors: s k c = c . f(ei ej ) = ei e8?j+1: Thus. There is a close correspondence between multilinear algebra. Endgame Algorithm We now present the endgame algorithm using the notation developed in Section 5.

Other factorizations appropriate for combined vector and parallel architectures.s = L Z XQ.s = Z LUs(I64k + LUs )6 Us (r(Us2 )) LUs (I64k + LUs rUs )6 Us + Us rUs Z D4 D4 XK.1 . For example. s0 for s=1 s which cs = cs0 . By modifying the factorizations. These equations are vectorizable as well Smitley 1991]. can also be derived Kaushik et al. By using the relations Us = (1 s)U1 and Xp. Sk i1 < <ik . if the interconnection architecture is a two-dimensional grid. The vectorized implementation of Table 1 by Burton Wendro et al.s C4 Table 1.s + XB.2) and (4.1 in Table 1. Table 1 de nes the piece-unmove operators. and de ne L 2 M64k diag Z X ! ci1 c ik Certain basis elements of B do to legal S -positions. conjugation by r of the operation of moving the R right corresponds to moving the R up: if one rotates the chessboard clockwise 90 . 1993]. such as a parallel network of vector processors. 1993]. and then rotates the chessboard counterclockwise 90 . has supported this claim Wendro et al. where (1 s) 2 Sk interchanges 1 and s. These Nnot ccorrespond \holes" are elements of the form k such that there exist distinct s. only Us for s = 1 can be directly computed.s = L XB. one can derive code suitable for di erent architectures.18 XR.s = (1 s)Xp.s = XR. If v 2 B then Lv is the projection of v onto the subspace of B generated by basis elements that are not holes. As in the case of fast Fourier transforms (4. This corresponds to moving the R to the right. moves the R right.3). by varying the factorization. code suitable for execution on a wide range of high-performance architectures can be derived. Fix a basis fci g64 i=1 of C. the result will be the same as if the R had been moved up to begin with. The average over C4 means that the R must be moved in 4 directions. Figure 2 illustrates the computation of the integrand in the expression for XR. equations appropriate for a grid architecture can be derived.s = LEWIS STILLER Z C4 XN. These equations de ne the core of a portable endgame algorithm. For example.

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

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

Black arrows correspond to r. The Cayley graph for D4 . The communication complexity of the routing can be reduced by exploiting the Cayley graph structure Stiller 1991a]. The position shown arose during a game between Anatoly Karpov and Gary Kasparov in Tilburg. . and ip the board horizontally. 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. An edge labeled h connects node g to node g0 if hg = g0 . October 1991.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. thus. and whose edges are labeled by one of the generators r or f. the chess value of each of these nodes is the same. and rotate the board counterclockwise 90 . This is a graph whose elements are the eight transformations in D4 . Gray (diagonal) arrows correspond to f.

Sims 1971. Lake et al. and so on. or move them backward. : : : . We let XWhite = XBlack = X fs : Ss fs : Ss X is White g XSs . that are suitable for the denotation of the algebraic structures used Butler 1992. 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. of the number of moves required to win that position Strohlein 1970. v 2 .22 LEWIS STILLER communication pattern used is that of a group action graph. 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. 1990.3.e. pp. is Black g XSs . 472{475]. which looks like a number of disjoint copies of the Cayley graph. Control structure. over all positions from which White can win. There are e cient algorithms. 1992. Although their terminology is di erent from our terminology. as well as languages. plus some cycles White 1984]. The method we use turns out to be similar to the orbital exchange method. Then v = v i is the set of positions from which White can win. 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. i White moves and i ? 1 Black moves). MAGMA.s : The backup rule used is vi+1 = XWhite(XBlack (vi )). 1992.s . . such as GAP. The method for computing vi from vi?1 is called the backup rule. where Symj is the j -th symmetric power of C Fulton and Harris 1991. possibly in conjunction with more traditional move generators. Thus. Several backup rules have been used in various domains Schae er et al. Sims 1994]. however. which is used to compute the FFT of a data set invariant under a crystallographic group An et al. in general.. The groups encountered here are so small. until some i is reached for which v i = v i+1 . and AXIOM. The overall structure of the algorithm is to iteratively compute the sets v1 . They are all characterized by the use of an unmove generator to \unmove" pieces. their general ideas are similar Hillis and Taylor 1990]. for performing the purely algebraic operations required. v1 is the vector of positions from which White can checkmate Black on the next move. 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. where a vertical bar denotes complement. that computer-assisted group-theoretic computation was not required. Tolimieri and An 1990]. and i is the maximin value of the set S : the maximum. 1994]. 6. Thompson 1986].

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

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

417. Cheron 1952. Unlike the \maximin" positions such as the one in Table 3. mutual zugzwangs can sometimes be understood by humans. p. maximum D of distance-to-win over all positions that are wins for White. 1982]. particularly when no pawns are involved. chess is a very hot game in the sense of Winning Ways Berlekamp et al. 50{53. seems to reserve judgment. but Black to move loses. con guration was \known to be a draw". pp. Berger 1890. he called only \controversial or unknown". The 92-move win in KQRKQR is somewhat surprising too. 167{ 169. Vol. this reduces the state space roughly by a factor of four. Fine 1941. 521]. while KBBKN. p. The fty-move rule would a ect the value of each endgame listed with max-towin of fty or more. Such positions seem amusing because. whose analysis is fairly impenetrable. 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. pp. Bishops linked with a bar are constrained to lie on squares of opposite colors. from 6:19 109 to 1:67 109 nodes. however. and number Z of mutual zugzwangs. number and percentage of positions that are wins for White.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. . Most of the standard works concurred with the opinion that KRBKNN was not a general win Euwe 1940. Statistics gathered for six-piece endgames: description of endgame. Cheron. 5. which was considered a draw by most players.

Ke4 Kc7 (If Ke7? 6. 4. it would be captured. Kf5 Kd8 7. shows that this general position in KBNNKR is drawn. : : : Ke5?. and then the Black K captures the other White N. but now 2. Kg6 Ke8 8. However. Qf4+ Kg1 4. Kg5 Qe3+ forces either perpetual check or a queen trade. Mutual zugzwang: White to play draws. any Black move loses. f8/Q (interpolating further checks does not help) when 2. Any other move of the B on move 1 allows Black to win the race. Ke4 Ke7! draws. Note also that neither N can move. : : : Kh2 2. If Kb1 then Qc2 mates. #546] is shown in Figure 7. discovered by the program. and the resulting subgame of KBNK would be winning for White. This composition Elkies 1993. 1. f8/Q Qh3+ 4. since g8 is guarded by the B on a2. : : : b1/Q loses to 3. as the R would immediately capture the other N. 1. For example. Bf7 Kc3 2. the Black R can sacri ce itself for a White N. Kg5 b1/Q with Kh1 and Be4 draws. left. Be4+ and mate. If the Qf6 moves then Qb2 or Qf1 will mate. On the other hand. 1. Ba2 Kc2. right. Now consider Figure 7. Kc2 Kc5 4. For example. f8/Q. : : : Kc3 2. Thus. 2 : : : Qb5+ . Qe5+ Kxg2 3. Kb1 Kd4 3. Ng6+ wins) 6. 1. Figure 6 shows an example. since other moves by White on the second move allow the Rh8 to escape via g7. If the Black K reaches g7 or e8 rst. Bf3. Qd6+? Kxg2 2. Thus. We quote from Elkies' analysis: \1.26 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. Ng6+ wins the R) 4 Kd3 Kd6 5. Rusinek 1994. con rmed by the program. : : : Qh3+ 3. Black tries for perpetual check. Qg7+ Not 1. but Black to play loses. discovered by the computer. Kg7 and White wins. This mutual zugzwang. Kb1 Kd4 3. The R is trapped on h8. Bb1+ Kc3 forces 2. If the Black R were to capture a N. Kc2 Kc5 (If 3. drawing. is somewhat inelegant in that includes promoted pieces. The position seems to be a race between Kings to see who will reach the upper right corner area rst. one that follows accepted aesthetic practice by avoiding the use of promoted pieces in the original position and by appearing more natural. Noam Elkies used it as the basis for a much more elegant composition. and not with 2 : : : Qd1+? 3. Kd3 Kd6 5. if the White K reaches g7 rst. It is not di cult to see that Black to play loses: White gets in rst. If Black moves the Qa7 then Qhg1 or Qa2 mates. Chess theory. White to move from the position in Figure 6 must move the B. it simply captures the Black Rh8. If 2. and the N's guard each other. if White moves rst then Black can force the draw. On the other hand. leaving the drawn endgame KBK.

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

but are not shown. f5-f4). Occasional local variations are possible (for instance. 15. . An optimal line of play for the position of Figure 8.28 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.

s or XB. . Nefkens 1985.s is slower than either XR. Implementation Notes and Run Times The implementation was on a 64K processor CM-2/200 with 8 GBytes RAM. XQ. 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 .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. van den Herik et al. 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. Thompson 1986. Exact timings depend on S (for instance. 1988] based on the ve-piece timings of the code reported here. The processors were interconnected in a hypercube and clocked at 7MHz (10 MHz for the CM-200).s ) as well as run-time settable factorization choices and load on the front end. as is clear from Table 1.

This would permit a more informative timing comparison. Molien. However. Thus. 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 .cs. Amelung. The CM-200 source code implementing Table 1 is currently available from ftp://ftp. 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. From a computational point of view. In addition to the work of Babbage. and Quevedo.edu/pub/stiller/snark. 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. 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. 1993]. argued for a postional evaluation function similar to the squares-attacked heuristic used in some full-chess programs Schwarz 1925]. 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. Schwarz. with larger and faster serial machines. In fact.jhu.2.30 LEWIS STILLER Unfortunately. of course. Bellman 1965. 10. in a little-known 1925 article in Deutsche Schachzeitung. 6] refers to private papers held by Molien's daughter. currently we are trying to locate these papers in the hope that they might shed light on the questions raised in Section 2. six-piece endgames are in fact coming within reach of classical solution techniques. Zermelo. p. and to locate and check his analysis of KRKB. we remark that K. 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. not currently possible since six-piece endgames could not have been solved in a practicable amount of time using classical techniques on previous architectures. and with enough spare cycles. direct comparison of six-piece timing against other work is. although this implementation currently solves only a single four-piece endgame. Amelung himself is also a gure about whom little is known. in my opinion a more promising approach would be to follow up on the suggestions rst made by Bellman Bellman 1961. and the remarks here would seem to suggest that a detailed reassessment of his contribution to the endgame study would be desirable.

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

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

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

Lecture Notes in Computer Science 559. \An algorithm for the machine calculation of complex Fourier series". Fundamental Algorithms For Permutation Groups. 22 February 1988. Cheron 1952] Andre Cheron. Technical Report SRC-TR-91-035. Draper 1991] Richard Noel Draper. Dekker. Journal of the American Mathematical Society 3 (1990). 1952. H. Guy. Fast Fourier Transforms. Clausen and Baum 1993] Michael Clausen and Ulrich Baum. \The rook and bishop against rook". \A fast distributed routing algorithm for supertoroidal networks". 1780]. \An overview of supertoroidal networks". with illustrations ]. Campbell. and hypercubes". Dekker et al. \Complexity starts at ve". 225{233. Science of Computer Programming. Campbell 1988] Murray S. Academic Press. Berlin/New York. Diaconis and Rockmore 1990] Persi Diaconis and Daniel Nahum Rockmore. John Horton Conway. France. 1982. Maryland. Lecture Notes|Monograph Series 11. Nouveau traite complet d'echecs: La n de partie New complete treatise of chess: The endgame ]. Maryland. Blelloch. 1993. Institute of Mathematical Statistics. Herschberg. fast Fourier transforms. \Bibliographisches Institut Wissenschaftsverlag. MIT Press. Winning Ways for Your Mathematical Plays I. Supercomputing Research Center. Diaconis 1988] Persi Diaconis. \Using chunking to solve chess pawn endgames. Springer. Supercomputing Research Center. Arti cial Intelligence 23 (1984). \Gray codes. \Chunking as an abstraction mechanism". Carnegie-Mellon University. Blelloch 1990] Guy E. Mathematics of Computation 19 (1965). volume 3. 97{120. Technical report. 93{104. 17 January 1991. 125{138. Butler 1992] Gregory Butler. Vector Models for Data-Parallel Computing. M. 1990. Bird. 297{332. Del gioco de gli scacchi T he game of chess ]. Jeremy Gibbons. Parallel Computing 6 (1988). Chamberlain. Campbell. 1989] Richard S. July 1990. 1617. Berliner and Campbell 1984] Hans Jack Berliner and Murray S. Jaap van den Herik. Bird et al. Group Representations in Probability and Statistics. and Geraint Jones. \E cient computation of the Fourier transform on nite groups". Technical Report CMU-CS-88-116.34 LEWIS STILLER Berlekamp et al. Crosskill 1864] Alfred Crosskill]. MA. de Rossi ?]. S. CA. Lille. London. 12 (1989). Carrera 1617] Pietro Carrera. Yves Demailly. 305{311. avec gures Analytical essay on chess. 1982] Elwyn Ralph Berlekamp. Arti cial Intelligence series]. 1988. Draper 1990] Richard Noel Draper. Chapais 1780] Chapais. Chamberlain 1988] R. 1987] Sito T. and Richard K. Cooley and Tukey 1965] James William Cooley and John Wilder Tukey. 1992. \Formal derivation of a pattern matching algorithm. Bowie. Essais analytiques sur les echecs. and I. Hayward. International Computer Chess Association Journal 10(3) (September 1987). Trento ?]. The ChessPlayer's Magazine 2 (1864). Bowie. . ca. Mannheim. Cambridge. 297{301.

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

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

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

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

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

Journal of Parallel and Distributed Computing 11 (1991). \The group of automorphisms of the game of 3-dimensional Ticktacktoe". New York 1971. Singleton 1967] Richard C. Salvio 1634] Alessandro Salvio. Paris. Stamma 1737] Philip Stamma. 249. 1994. \Computations with permutation groups". 273{290. Sims 1994] Charles Co n Sims. \On computing the fast Fourier transform". 1971 (edited by Stanley Roy Petrick). pp. \Structuring data parallelism using categorical data types". 3: Almost Miniatures: 555 Studies With Eight Chessmen. Rusinek 1994] Jan Rusinek. Is Man a Robot?. ACM Press. 1972. \A proposed revision of the `50-move rule': Article 12. Schwarz 1925] K. 99{117.4 of the Laws of Chess". International Computer Chess Association Journal 7(3) (September 1984). 55{64. Trattato dell'inventione et art liberale del gioco di scacchi Treatise of the invention and liberal art of the game of chess ]. EG (ARVES. Encyclopedia of Mathematics and its Applications 48. H. Essai sur le jeu des echecs Essay on the game of chess ]. 247{254. and Duane Szafron. International Computer Chess Association Journal 12(2) (June 1989). Stiller 1991a] Lewis Benjamin Stiller. 1986. 321{324. \American Mathematical Monthly 74 (March 1967). Selected Endgame Studies. London.40 LEWIS STILLER Roycroft 1972] Arthur John Roycroft. Los Angeles. 1737. Journal of Supercomputing 5 (1991). Silver 1967] Roland Silver. England. 164{170. CA. Chichester. \Solving the Game of Checkers". Amsterdam) 5 (#75. pp. Schae er et al. April 1984). \Versuch eines mathematischen Schachprinzips" An attempt toward a mathematical principle for chess]. Napoli. Joseph Culberson. Cambridge. 1993. Roycroft 1984] Arthur John Roycroft. \Bit-serial SIMD on the CM-2 and the Cray-2". Wiley. Communications of the ACM 10 (1967). Singleton. Los Alamitos. Brent Knight. Norman Treloar. . Skillicorn 1993] David Benson Skillicorn. Computation With Finitely Presented Groups. ??{?? in this volume. IEEE Computer Society Press. 135{145. Berlin. Smitley 1991] David Smitley and Kent Iobst. 1994. Faber and Faber. Roycroft 1984] Arthur John Roycroft. 23{28 in Proceedings of the Second Symposium on Symbolic and Algebraic Manipulation. Cambridge University Press. \A World Championship caliber checkers program". \Group graphs and computational symmetry on massively parallel architecture". \Parallel analysis of certain endgames". Sims 1971] Charles Co n Sims. Deutsche Schachzeitung 53(11) (November 1925). vol. Schae er and Lake 1996] Jonathan Schae er and Robert Lake. 110{115 in Proceedings of the 1993 Workshop on Programming Models for Massively Parallel Computers. Arti cial Intelligence 53 (1992). Paul Lu. Schwarz. \Two bishops against knight". pp. 647{654. University of Limburg. 1993. Test Tube Chess: A Comprehensive Introduction to the Chess Endgame Study. England. 1634. Stiller 1989] Lewis Benjamin Stiller. Simons 1986] Geo Leslie Simons. 1992] Jonathan Schae er.

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

Colloques internationaux du Centre National de la Recherche Scienti que 37. and Surfaces. Cambridge. England. Wendro et al. Paris. 103{111. vol. Dean Mayer. Moscow. Berlin/New York. \Dva kon protiv pexek (teoretiqeski$ i oqerk)" Two knights against pawn (theoretical essay)]. A History of Algebra: From Al-Khwarizm to Emmy Noether. and Ralph Brickner. New York. 341{352. pp. \A bridging model for parallel computation". 383{406 in Les machines a calculer et la pensee humaine. 1984. 945{971 in Handbook of Theoretical Computer Science: Algorithms and Complexity. van de Velde 1994] Eric F. English translation by Abe Shenitzer: The Genesis of the Abstract Group Concept: A Contribution to the History of the Origin of Abstract Group Theory. Springer. ::: Lewis Stiller Department of Computer Science The Johns Hopkins University Baltimore. Springer. Valiant 1990a] Leslie G. 1953. Applied Numerical Mathematics 12 (1993). 1985. Valiant. 1994. \Uber eine Anwendung der Mengenlehre auf die Theorie des Schachspiels" On an application of set theory to the theory of playing chess]. 1993] Burton Wendro . White 1984] Arthur Thomas White. 501{504 in Proceedings of the Fifth International Congress of Mathematicians. Valiant.42 LEWIS STILLER Torres-Quevedo 1951] Gonzalo Torres-Quevedo. Tony Warnock. 2 (edited by Ernest William Hobson and Augustus Edward Hough Love). Wussing 1969] Hans Wussing. pp. 1990. van de Velde. Deutscher Verlag der Wissenschaften. 285{295. van. Zermelo 1913] Ernst Zermelo. 1934. \Presentation des appareils de Leonardo Torres-Quevedo" Presentation of the machines of Leonardo TorresQuevedo]. Elsevier. Communications of the ACM 33(8) (1990). Valiant 1990b] Leslie G. Concurrent Scienti c Computing. MIT Press. North-Holland/ Elsevier. 1913. Fizkul'tura i turizm. Graphs. \General purpose parallel architectures". von ] See under last name. volume A (edited by Jan van Leeuwen). Troitski 1934] Alekseii Alekseevich Troitski . MA. Amsterdam.com . Cambridge University Press. Texts in Applied Mathematics 16. 1912. Paris. 1984. Cambridge. Publications du CNRS. MD 21218 stiller@symmetry. Lewis Benjamin Stiller. Die Genesis des abstrakten Gruppenbegri es: ein Beitrag zur Entstehungsgeschichte der abstrakten Gruppentheorie.xo. van der Waerden 1985] Bartel Leendert van der Waerden. White 1975] Dennis Edward White. 1951. \Multilinear enumerative techniques. pp. \Linear and Multilinear Algebra 2 (1975). New York. Groups. Berlin. pp. 248{288 in Sbornik xahmatnyh t dov Collection of chess studies ]. \Bits and pieces: constructing chess endgame databases on parallel and vector architectures". 1969.

Sign up to vote on this title

UsefulNot usefulHow to calculate chess endgame tables

How to calculate chess endgame tables

- Readme
- E-MiniMax
- 291.minimax
- Ocena pozicije
- Chess Mantis
- Bruce Pandolfini-Pandolfini's Chess Complete
- Computer Chess
- Chess for Rookies - Craig Pritchett
- Carol Us
- play chess with god
- chESS reP
- Chess Cryptosystem
- aigames
- Winning Chess Strategies (2nd Edition) [2005 Vasser Seirawan with Jeremy Silman].pdf
- Chess - Encyclopedia of Chess Wisdom Excerpt
- Chess Bites
- Model Templates for PROCESS for SPSS and SAS Part 2
- 56497190-Middle-Game-in-Chess.pdf
- E-MiniMax
- שחמט
- Davies 43
- Positional Chess Sacrifices
- Shannon Chess
- 137-161
- Chris 15
- A - Chess Improvement-A !!!!!
- American Chess Cod 00 Brit
- ReadMe
- Berliner Hans-Artículos Breves-Kasparov vs Deep Blue-1998
- AMIGA - Art of Chess, The Manual
- Endgame Maths