Papex

Open problem · Mathematics · Open

Solving chess

With perfect play from both sides, does a game of chess end in a win for White, a win for Black or a draw?

Summary

Chess is a finite game, so one of the three outcomes is forced, but its game tree is far too large to search. Endgame tablebases give perfect play for all positions with up to seven pieces, and some smaller variants have been solved.

Why it is open

Chess has no element of chance, and each player sees every move made so far. Claude Shannon, citing John von Neumann and Oskar Morgenstern, noted in 1950 that these two facts make every position a forced win for White, a draw or a forced win for Black [1]. He called this "for practical purposes, of the nature of an existence theorem", because "No practical method is known for determining to which of the three categories a general position belongs" [1]. In principle the answer can be found by following every line of play to its end, which the rules guarantee after a finite number of moves, and working backwards [1].

"Solved" can mean three things. In Victor Allis's 1994 definitions, a game is ultra-weakly solved when the result of perfect play from the start is known; weakly solved when a strategy that achieves at least that result for both players has been found "under reasonable resources"; and strongly solved when such a strategy is known for every legal position [2]. Allis added that without the condition on resources, chess "could be weakly solved" simply by searching its full game tree [2]. Jonathan Schaeffer and colleagues used the same three levels when they solved checkers [3].

The obstacle is size. Shannon estimated about 30 legal moves in a typical position and, for a game of 40 moves, 10¹²⁰ lines of play; at one line per microsecond, the first move alone would take "over 10⁹⁰ years" [1]. No body sets a bar for a resolution. The precedent is checkers, settled by a computational proof that includes an explicit strategy [3]. Its authors, noting that checkers has roughly the square root of the number of positions in chess, wrote that "chess will remain unsolved for a long time, barring the invention of new technology" [3].

What is known

The size of the game (1950–2022). Shannon put the number of possible positions at "roughly 10⁴³" [1]. In 2022 John Tromp estimated the number of legal positions, those reachable from the starting position in a legal game, at (4.82 ± 0.03) × 10⁴⁴ with 95% confidence [4]. He did this by numbering about 8.7 × 10⁴⁵ candidate positions and testing random samples of them for legality, and gives an estimate rather than an exact count because deciding whether a chess position is legal "is exceedingly hard" [4].

Endgame tablebases (2012–2026). A tablebase is built backwards from checkmate by retrograde analysis and records the result of perfect play for every position with a given set of pieces [8]. In 2012 a team at Moscow State University generated a complete tablebase of seven-piece endings on the Lomonosov supercomputer and found the longest known checkmate, in 549 moves [5]. In August 2018 Bojun Guo and Ronald de Man completed the first free seven-piece tablebase, in the Syzygy format: 423,836,835,667,331 unique legal positions in 18.4 TB, against 100 TB for the Lomonosov tables [6]. In its longest ending the defender holds out for 1,034 half-moves before the stronger side can make a winning capture [6]. Chess engines consult such tables to decide which endgames to aim for [6].

In 2021 Marc Bourzutschky reported that, after generating about 15% of the eight-piece endings without pawns, the longest win he had found took 400 moves to reach checkmate or a capture [7, 8]. Kirill Kryukov's count gives about 3.8 × 10¹⁶ legal positions with eight pieces, two of them kings, almost 100 times the seven-piece total [8]. In February 2026 Lichess put online Bourzutschky's "op1" tables, which cover eight-piece positions with at least one pair of opposing pawns on the same file: about 5.0 × 10¹⁴ positions, 1.3% of all legal eight-piece positions, but about half of the eight-piece endings reached in practical games [8]. The announcement adds that "full 8-piece tables are not on the horizon" [8].

Smaller games (2007–2017). Checkers, with about 5 × 10²⁰ positions, was weakly solved in 2007 after computations that had run almost continuously since 1989: perfect play by both sides leads to a draw [3]. The proof combined a database of 3.9 × 10¹³ positions with ten or fewer pieces, built backwards, with a forward search, and consists of "an explicit strategy that never loses" [3]. In 2013 Mehdi Mhalla and Frédéric Prost weakly solved Gardner's minichess, played on a 5 × 5 board with about 9 × 10¹⁸ legal positions: it is a draw, and their proof can be checked by a human [9]. In a note published in 2017, Mark Watkins reported a completed computer proof that in the variant losing chess the first move 1. e3 wins for White [10]. In 2019 an error was found in some of the tablebases that proof used, and he published a revised proof [10].

What engines show (2018–2020). AlphaZero, which learned chess by playing against itself, "convincingly defeated a world champion program" [11]. Schaeffer and colleagues note that such programs, however strong, "are not perfect", and describe solving a game as "replacing the heuristics with perfection" [3]. In a 2020 preprint, Nenad Tomašev, Ulrich Paquet, Demis Hassabis and Vladimir Kramnik wrote that "It is widely hypothesised that classical chess is theoretically drawn" [12]. In their AlphaZero self-play, 8,820 of 10,000 games of classical chess at one second per move and 979 of 1,000 at one minute per move were drawn [12]. In every variant they tested, draws became more frequent with longer thinking time, which they read as a sign that the variants might be theoretically drawn, "like in Classical chess" [12]. These are results of imperfect play: they fit a draw but do not prove it [2, 12].

Main approaches

Proof search with databases. Search forward from the starting position towards positions whose values are already known from databases built backwards, as was done for checkers [3]. Strongest evidence: this method weakly solved checkers [3], and a far smaller proof was enough for 5 × 5 minichess [9]. Main difficulty: chess has about 10⁴⁴ legal positions [4], while checkers alone needed computations from 1989 to 2007, with more than 200 processors at the 1992 peak [3].

Larger tablebases. Extend exact backward analysis to positions with more pieces [8]. Strongest evidence: complete seven-piece tablebases exist [5, 6], and a practically important part of the eight-piece positions is online [8]. Main difficulty: eight-piece positions outnumber seven-piece ones almost 100 to 1, and complete tables must resolve every promotion, including promotions to pieces other than a queen, in a fixed order [8].

Self-play estimates. Use the strongest programs to estimate the value of the starting position from games against themselves [12]. Strongest evidence: AlphaZero's draw rate in classical chess rose with thinking time, to 979 of 1,000 games at one minute per move [12]. Main difficulty: the games are played below perfect play [12], and a weak solution requires a strategy that never does worse than the game's value, not a high draw rate [2].

Arguments without search. Prove facts about the value from the structure of the game. Strongest evidence: Shannon showed that if a player could pass instead of moving, White could not lose: with no winning move, White passes and Black faces the mirror image of the same position [1]; similar reasoning shows that hex is a first-player win without giving a winning strategy [2]. Main difficulty: Shannon's argument needs the rule change, and in chess a player must move [1].

Open sub-questions

  1. Saturation. Bourzutschky suggested that eight pieces may be at or close to a "saturation point", where more pieces no longer lengthen the longest wins [7]. Is the longest win in the op1 tables, in the same depth-to-conversion measure, longer than the longest pawnless eight-piece wins [7, 8]?
  2. Coverage beyond op1. About half of the eight-piece endings reached in games lie outside the op1 tables [8]. Is there another subset that, like op1, can be generated without the full tables and covers a large share of them [8]?
  3. Draws and thinking time. In AlphaZero self-play the draw rate rose from 88.2% at one second per move to 97.9% at one minute [12]. Does it keep rising towards 100% at longer times, and do White's wins fall faster than Black's?
  4. Checkable proofs. The losing-chess proof had to be revised after tablebase errors were found [10]. Can a proof that relies on tablebases emit certificates that an independent program can check, much smaller than the tables themselves [6, 10]?
  5. Larger boards. Gardner's 5 × 5 minichess was weakly solved with "surprisingly small computing power" [9]. Does the same method weakly solve a larger variant, and how does the proof size grow with the board?

Work on this in Papex

If you have a testable answer to one of the sub-questions above, you can draft it as a private hypothesis in Papex, or prepare the draft from Claude or ChatGPT with the Papex connector. You review the exact version, sign it and decide whether to publish it; a public record that names this problem is listed on this page. A Papex record shows who signed which version and when. It is not peer review and does not mean the hypothesis is correct.

Key references

  1. Shannon, C. E. (1950). Programming a computer for playing chess. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science (Series 7) 41(314), 256–275. 10.1080/14786445008521796 · Checked 15 Oct 2026
  2. Allis, L. V. (1994). Searching for solutions in games and artificial intelligence. PhD thesis, Rijksuniversiteit Limburg, Maastricht (defended 23 September 1994). 10.26481/dis.19940923la · Checked 15 Oct 2026
  3. Schaeffer, J., Burch, N., Björnsson, Y., Kishimoto, A., Müller, M., Lake, R., Lu, P. & Sutphen, S. (2007). Checkers is solved. Science 317(5844), 1518–1522. 10.1126/science.1144079 · Checked 15 Oct 2026
  4. Tromp, J. (2022). Chess Position Ranking. GitHub repository, README with the legal-position estimate, last revised for the estimate in May 2022. Source · Checked 15 Oct 2026
  5. Zakharov, V. B., Mal'kovskii, M. G. & Mostyaev, A. I. (2019). On solving the problem of 7-piece chess endgames. Programming and Computer Software 45(3), 96–98. 10.1134/S036176881903006X · Checked 15 Oct 2026
  6. Lichess (2018). 7-piece Syzygy tablebases are complete. Lichess blog, 19 August 2018. Source · Checked 15 Oct 2026
  7. Bourzutschky, M. (2021). 8-men tablebase: first explorations. E-mail reproduced by ARVES on an undated page; reference 8 dates the announcement to 2021. Source · Checked 15 Oct 2026
  8. Lichess (2026). Op1: partial 8-piece tablebase available. Lichess blog, 7 February 2026. Source · Checked 15 Oct 2026
  9. Mhalla, M. & Prost, F. (2013). Gardner's minichess variant is solved. ICGA Journal 36(4), 215–221. 10.3233/ICG-2013-36404 · Checked 15 Oct 2026
  10. Watkins, M. (2017). Losing chess: 1. e3 wins for White. ICGA Journal 39(2), 123–125. 10.3233/ICG-170017 · Checked 15 Oct 2026
  11. Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai, M., Guez, A. et al. (2018). A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play. Science 362(6419), 1140–1144. 10.1126/science.aar6404 · Checked 15 Oct 2026
  12. Tomašev, N., Paquet, U., Hassabis, D. & Kramnik, V. (2020). Assessing game balance with AlphaZero: exploring alternative rule sets in chess. Preprint, arXiv:2009.04374 (version 2, 15 September 2020), not peer reviewed. Source · Checked 15 Oct 2026

Page reviewed 15 Oct 2026

Source

Open, as listed by Wikipedia · Unsolved problems in mathematics. Checked 08 Oct 2026.

The status is the source’s. Papex does not decide whether a problem is solved.

Records on this problem

No public record addresses this problem yet.

More in Mathematics