Impartial games

Nim with heaps of 3, 5, 7. Heaps of counters; a move takes any number from one heap. The position is a loss for the player to move exactly when the binary digits of the heap sizes cancel in every column — the nim-sum — and that is the whole of the theory of Nim.

Nim, and the nim-sum

Three heaps of counters, take as many as you like from one of them, and the player who takes the last counter wins. The winning condition is not a search, not a table, and not a heuristic — it is the bitwise exclusive-or of the heap sizes, and it was found in 1901.

Every impartial position is a Nim heap. A heap in a subtraction game, its Grundy value, and the Nim heap it is equivalent to. The equivalence is exact: the two positions have the same options up to value, so they behave identically in any sum, which is the Sprague–Grundy theorem.

Every impartial game is a Nim heap

Sprague and Grundy proved, independently and four years apart, that any position in any impartial game is equivalent to a single heap of counters. Not similar to one — equal to one, interchangeable with it inside any larger game.

Grundy values for subtraction of 1, 2, 3. The Grundy value of every heap size for a take-away game, computed by the mex rule. A period, if the figure marks one, was found by searching the computed sequence rather than assumed — and where no period is marked, none was found in the range drawn, which is not the same as there being none.

Grundy sequences, and where they stop being predictable

Computing one Grundy value is a mex. Computing all of them produces a sequence, and the sequences do something nobody has fully explained — most of them eventually repeat, some of them take thousands of terms to start, and for a few nobody knows whether they ever do.

The octal game ·137, read out. An octal code is a rule table. The kth digit says what a player may do after taking k tokens from one heap: end that heap, leave one heap, or split the rest into two. Three bits, one digit, and the whole family of take-away games becomes something that can be listed and swept.

Naming a game with a number

An octal code is a rule table compressed into an integer. It turns "which game" into something that can be counted through, which is how the family was swept — and how the games nobody can solve were found.

Wythoff's game, and the line the losing squares lie on. A queen moves left, down, or diagonally down-left any distance, and whoever cannot move loses. Every square carries the Grundy value the mex rule gives it. The squares worth nothing — the ones a player wants to hand over — lie along two lines whose slopes are the golden ratio and its reciprocal, in a game with no geometry and no continuous quantity in its rules.

Wythoff's game, and the ratio nobody put there

Two heaps, three kinds of move, and losing positions that lie along a line of irrational slope. Nothing in the rules mentions a ratio, a length or a continuous quantity — and the golden ratio comes out anyway.

Poker Nim from 3, 5, 7, with reserves of 4 and 4. Nim with one extra kind of move: a player may put any number of counters back onto a heap from a private reserve. It looks as though a losing player could stall for ever. They cannot, and the winner is decided by exactly the same nim-sum as ordinary Nim — checked here over every position within a stated range rather than argued.

The move that gives counters back

Poker Nim adds one rule to Nim — a player may put counters back onto a heap from a private reserve. It looks as though a losing player could stall for ever. The winner is decided by exactly the same nim-sum, and the reason is the single most useful idea in the whole reduction apparatus.

Turning Turtles: a row of 12 coins. A row of coins, some heads and some tails. A move turns some of them over, and the rightmost coin turned must go from heads to tails — which is what makes the game end. The number under each place is what a lone head there is worth, and the row is worth the exclusive or of the places showing heads.

A row of coins is already a sum

Everywhere else on this site a sum is several positions side by side. In a coin-turning game it is one row — each coin showing heads is a game in its own right, and the row is worth the exclusive or of them. The decomposition is inside a single picture.

Grundy values for subtraction of 1, 3, 4. The Grundy value of every heap size for a take-away game, computed by the mex rule. A period, if the figure marks one, was found by searching the computed sequence rather than assumed — and where no period is marked, none was found in the range drawn, which is not the same as there being none.

Take one, three or four

A heap and a list of legal takes. It is the smallest interesting impartial game there is, and the only family in the subject where eventual periodicity is not observed, not conjectured, but guaranteed — with a bound on when it must appear.

Nim-multiplication below 16, and every field axiom checked. The nim-product, defined by taking the least value the product is not forced to be — the same manoeuvre as the mex rule, applied to a product rather than to a move. The result is that these values are not merely a group under nim-addition but a field: every axiom is checked over the whole table here, including an inverse for every non-zero value, and the sizes at which the axioms fail are reported rather than avoided.

The nimbers multiply

Nim-addition is exclusive-or and everybody meets it first. There is also a multiplication, defined by the same take-the-least-value-not-forced manoeuvre as the mex — and it makes the nimbers below sixteen a field, with every axiom checked here and an inverse for every non-zero value.

A grid of coins, and a multiplication table. The Grundy values of a two-dimensional coin-turning game, computed from its own move rules by a mex at every cell. Down the left and across the top are the one-coin values of the two one-dimensional games it is built from. Every cell is the nimber product of its two edge values — the multiplication defined for the nimber field on algebraic grounds — and beside the grid are the three combining rules a reader would try first, each killed on a named cell.

The tartan theorem

The nimbers are a field, with a multiplication defined by a mex-style rule that looks like an algebraist's amusement. Lay two coin-turning games on a grid and the Grundy value of each square is the nimber product of its two coordinates — which is the point at which the multiplication stops being a curiosity and starts computing answers.

Mock Turtles on 8 coins: every lost position. The rows a player to move has already lost, drawn in full. A filled disc is a coin showing heads. The set is closed under turning over every coin two of its members disagree about, which is what makes it a linear code, and the count of heads in the sparsest of them is the fewest coin turns that separate two lost positions.

The losing positions are a code

Turn over one, two or three coins, and the rows a player has already lost turn out to be closed under adding two of them together. That makes them a linear code — and on eight coins it is the extended Hamming code exactly, sixteen words with a weight enumerator of 1 + 14x⁴ + x⁸, produced by a move rule that knows nothing about codes.

Grundy values for subtraction of 2, 5, 7. The Grundy value of every heap size for a take-away game, computed by the mex rule. A period, if the figure marks one, was found by searching the computed sequence rather than assumed — and where no period is marked, none was found in the range drawn, which is not the same as there being none.

The period is small and the proof does not say so

Every subtraction game repeats eventually — that is a theorem, and its proof gives a bound of sixteen thousand for a three-move set. Over 112 sets the longest period measured is twenty-two. The proof and the fact are four orders of magnitude apart, and the rule of thumb that closes the gap is broken by one set in the sweep.

Moore's Nim with k = 2: the columns, divided by 3. The heap sizes in binary, with each column added as an ordinary sum rather than exclusive-or. In Moore's Nim a move may take from as many as k heaps at once, and the position is lost for the player to move exactly when every column sum is divisible by k + 1. Ordinary Nim is k = 1, where divisible by two means an even number of ones — the same picture with a different divisor.

Taking from several heaps at once

Moore's Nim lets a move take from as many as k heaps at a time, and the losing positions are still read off the binary columns — divisible by k + 1 rather than by two. The rule agrees with exhaustive search over 54,264 positions and never disagrees, and it decides every outcome while supplying no value at all: reading the same columns as a base-3 number gets the Grundy value right on 42 of 330 positions.

Cram

Domineering with one word of the rule changed: both players may place a domino either way up. That makes the game impartial, and the entire partizan apparatus collapses into a single Grundy value — on the 4 × 4 board, Domineering's canonical form runs to 114 characters of nested braces and Cram's answer is the one character 0.

Generalized Geography. A token on a directed graph. A move slides it along an edge to a vertex not yet visited, and a player who cannot move loses. That is the whole game, and deciding who wins it is as hard as anything decidable in polynomial space — which is the strongest hardness claim anybody makes about a combinatorial game.

A token on a graph

Geography is an impartial game whose position is a vertex and a history, so a ten-vertex graph has ten thousand states rather than ten. Take the arrows off and the same game is decided by a maximum matching — first player wins exactly when every maximum matching covers the start, verified on 41 vertices across eight graphs. One word in the rules separates a polynomial criterion from a PSPACE-complete problem.

Mock Turtles on 8 coins: finding the move is decoding. Every row of the game, sorted by what it takes to win from it. The lost rows are the codewords; a won row is a codeword with errors, and the winning move is the error pattern that turns them off. The distance column is a fact about the code and the coins column is a fact about the rules, and the two do not quite agree.

The code names the move

If the lost rows of a coin-turning game are a linear code, then a won row is a codeword with errors in it and the winning move is whatever turns the errors off. Over all 256 rows of Mock Turtles on eight coins: 16 codewords, 240 won rows, none more than two coins from a lost one — and 64 of them whose cheapest winning move has to turn three coins anyway.

Welter positions and what they are worth. Coins on a strip, with the Grundy value the recursion returns and the nim-sum the squares would have if they were independent heaps. The two columns are the essay: they hardly ever agree.

No two heaps alike

Welter's game is Nim with one extra clause — no two heaps may be the same size — and the clause is fatal to the nim-sum, which gives the right answer in none of the 120 three-coin positions. What replaces it is a function of pairs: ⟨a | b⟩ = (a ⊕ b) − 1, exact on all 55 two-coin positions, and nim-added over every pair it is exact on the whole board provided the number of coins is even.

How many different values a Grundy sequence has used. One curve per octal code: the number of distinct Grundy values among the first n heaps. A periodic game runs out of values and its curve levels off. The codes nobody has settled are still climbing at six thousand heaps.

The values that keep arriving

A Grundy sequence that repeats uses finitely many values and stops needing new ones. Six thousand heaps into ·007 the count of distinct values is 187 and still climbing, and the share of heaps carrying something outside the twenty-two commonest rises from 32% in the first thousand to 85% in the sixth. The rare values a periodicity argument needs to thin out are getting commoner.

Three take-and-break games, three kinds of answer. The Grundy sequences of Nim, Lasker's Nim and Kayles over the first heaps. Adding a move that removes nothing takes Nim's sequence from the identity to a four-line formula; bounding how much may be taken instead takes it somewhere with no formula at all.

Splitting is a move

Add to Nim a move that removes nothing — break a heap in two — and the Grundy sequence gets simpler, not harder. Lasker's Nim has a closed form with one clause per residue modulo four, exact on all 2,001 heaps checked: the identity with every fourth pair transposed. Kayles is the same kind of game with the taking bounded instead of the splitting, and it has no closed form at all, settling into a period of twelve only from heap 71 with fourteen values outside it for ever.

Every heap up to 40, won or lost. Heap sizes with the outcome for the player who moves first. The lost ones are shaded; they are exactly the Fibonacci numbers, which is a fact about a game with one heap, no board and no geometry in it anywhere.

The heap is not the position

Fibonacci Nim bounds a move by twice the previous move, which puts the state outside the board: a heap of six with a cap of two and a heap of six with a cap of five are different games. So there is nothing to add and no Grundy value to compute — and the game is completely solved anyway. The opener loses on exactly the nine Fibonacci numbers up to 120, and the smallest term of the Zeckendorf numeral is a winning move in all 110 winnable heaps.

A period with a constant added

An octal code says what a player may do when removing k counters, in three bits; a hexadecimal code adds a fourth — leave three heaps — and the digits run to fifteen. Over twenty-two codes swept to six hundred heaps, five hexadecimal ones repeat with a fixed amount added each time round and no octal one does. Their values climb for ever and never repeat, so a search that looks only for repetition reports them unsettled.

7 symmetries, and the one that is a strategy. 4 games and 7 candidate symmetries, each tested by playing the strategy out against every opponent line rather than by argument. A pairing strategy needs a map that fixes the start, is an involution, and carries one player's moves to the other's — and the last condition is where most of these fail.

The strategy that is a symmetry

A pairing strategy is a symmetry of the board that turns one player's moves into the other's, and it wins without computing anything. Tested by playing it out rather than argued, it wins one of seven candidate symmetries across four games — exactly the Cram boards with both sides even, which is exactly where no domino is its own image.

Looking for the symmetry

Answering every move with its mirror image wins Cram on a board with both sides even, which is the argument everybody meets. Asked of every connected shape of at most eight squares instead of of thirteen rectangles, it wins twelve — and accounts for a sixth of the second-player wins there are, because 852 of the 1,042 shapes have no symmetry to answer with in the first place.

Moore’s rule, reversed. Moore’s Nim under the misère convention at three values of k, with the normal-play rule and the same rule plus a clause about heaps of one. The patch is the one Nim takes, with the modulus the normal-play rule already carries, and it is right on every position swept.

The patch that generalised

Misère Nim takes a one-line patch: play the normal-play strategy until every heap holds a single counter, then invert. Moore's Nim, where a move may take from up to k heaps at once, takes exactly the same patch with exactly the same modulus — and the two rules disagree on six positions out of 923.

The Fibonacci numbers are one row of a table. The losing heaps of Fibonacci Nim with the factor two replaced by one, three, four and up to eight. Each factor gives a different integer sequence: the powers of two, the Fibonacci numbers, and four more with no common name.

The family the Fibonacci numbers belong to

Fibonacci Nim lets a player take at most twice what the last one took, and the heaps the opener loses are the Fibonacci numbers. Two is an arbitrary number. At one the losing heaps are the powers of two, at three and four and five they are four more sequences, each with a linear recurrence whose lag is twice one less than the factor — until the factor is six, where the pattern stops.

Parity decides it before the shape does. For each size, how many first-player wins can reach a position a half-turn pairs in a single move. Every odd size is nought and cannot be anything else, because a pairing needs an even number of squares and a move removes two.

The symmetry one move away

A pairing argument proves the second player wins and names no move to do it with. Asked of every shape of up to eight squares it settles twelve boards. Asked one move later — can the first player reach a position a half-turn pairs? — it settles 288, and which boards those are is decided by parity before anything about their outline is looked at.

The gap is the factor. The separation condition of each factor's greedy numeral system: the smallest gap between the indices of two terms. It is one at factor one, two at factor two — Zeckendorf's non-adjacency — and the factor itself at every factor swept.

What the numerals knew

Every factor in the Fibonacci Nim family gives a numeral system, and the rung below predicted its separation condition would be the lag of the recurrence the losing heaps satisfy. It is not. The gap is the factor — one at c = 1, Zeckendorf's two at c = 2, and c at every factor to eight — while the lag goes 1, 2, 4, 6, 8, 11, 14, 17 and leaves its own pattern at six. The numerals then solve every one of 194,480 states, cap and all.

Every saltus in the two-digit family. The constant added each time round, over all 255 two-digit hexadecimal codes. Forty-eight codes add one, thirteen add two, six add four and three add sixteen — and one code adds three.

A code that climbs by three

Five hexadecimal codes were known to repeat with a constant added, and every one of the five constants was a power of two — either a fact about exclusive-or or a coincidence over five cases. Sweeping all 255 two-digit codes settles it: seventy-one climb, seventy of them by 1, 2, 4 or 16, and one by three. The exception is ·3f, whose values are 3⌊n/6⌋ + (n mod 3) on every heap to twelve hundred.

The rule a smaller move breaks

Moore's Nim lets a player take from at most k heaps, and its winning condition is the binary columns summed modulo k + 1. Cap the amount as well and the obvious repair — reduce each heap modulo the cap plus one, then read the columns — is exact at every cap when k is one and wrong at every cap when k is two or three. The reason is stronger than a broken rule: at k ≥ 2 the residues do not determine the outcome at all, so nothing of that shape can work.

Four second parts, and none of them enough. The residues paired with each of four further counts, at three caps, with a move taking from two heaps. Every pairing leaves classes containing both a win and a loss.

The wider move is the easier game

An earlier essay ruled out every rule that reduces the heaps and reads the residues, and asked for a two-part statistic: the residues plus one more count. Four second parts are tested here and none of them decides. What turns up instead contradicts the premise the request was made under — a move that may reach three heaps is more predictable than one that may reach two, on every cap, every candidate rule, and after the change in the base rate is taken out.

Climbing is the ordinary case. The two-digit and three-digit hexadecimal families, each swept for exact and arithmetic periodicity. Seven in ten of the settled three-digit codes repeat with a constant added.

The third digit

The rung below found 71 of the 255 two-digit hexadecimal codes repeating with a constant added rather than exactly, and asked whether the same share holds one digit wider. It rises. Of the 4,095 three-digit codes, 1,433 climb and 617 repeat exactly — seven in ten of the settled ones — so a saltus is the ordinary way a hexadecimal game settles and the exact repetition the octal survey was built to find is the special case.

A check in front of a search

The rung below found a pairing one move away on 288 of the 767 even first-player shapes, and asked what a solver that tested for one before recursing would save on a real game. On an even Cram board it saves nearly the whole search — a 4 × 5 board takes 17,348 node expansions without the check and one with it — and the depth profile shows why that number flatters: the check settles every winning position at the opening and at the last two moves, and about one in ten in between.

Sorted by how many heaps are odd. The 2,002 positions of the census grouped by how many of their heaps hold an odd number of counters. Four of the six groups are entirely lost or entirely won.

The count of odd heaps

The rung below refused a family of two-part rules for bounded Moore's Nim and asked what the 364 losing positions have in common as a set. They have an invariant, and it is a statistic of the whole position rather than of a heap: how many heaps hold an odd number. Every all-even position is lost, at every width of move, by a restoring strategy — and the count settles every position at one heap a move and at four, and a little over half at two.

The only way to split into three

Nineteen three-digit hexadecimal codes climb by three, and the rung below asked whether they share a form and what digits they have in common. The digits are exact: on eighteen of them the only way to split a heap into three is by taking exactly three counters, and taking three counters can do nothing else. The form is not shared — the eighteen carry four distinct sequences, and exactly one of the four counts in base three.

The check fires on positions that lose. How often the pairing check accepts a position, and how often the position is a loss. On every even board in the sweep it is wrong between an eighth and a fifth of the time.

The check that was not a check

The rung below asked for a depth-conditioned solver and named the board to measure it on. Building it found two things first. The pairing check is unsound at interior positions — on a four by five Cram board it fires on 8,613 positions and 1,026 of them are losses — and the board it named has twenty-five squares, so the check can never fire there at all. Repaired, the check is right everywhere, and the policy that pays is the root alone.

Thirty-two words, four of them lost. Every position of five heaps grouped by the parities of its heaps in decreasing order of size. Each word is uniform, and four of the thirty-two are losing.

The parities, in size order

The rung below settled four of six parity classes in bounded Moore's Nim and asked whether the sizes pick out the losing positions in the two it could not. They do — but only through the order they put the parities in. Sort the heaps largest first, read off their parities, and that five-bit word settles the whole game at every width of move, with the losing words forming a subspace.

Two symmetries, the same two clauses. The half-turn pairing and the reflection pairing written side by side, with the fixed squares and self-paired dominoes each has to exclude.

A pairing, and the pairing

The rung below repaired the half-turn check and asked whether a reflection would fire where it does not. It does — forty positions of 58,830 on the largest board — and it is sound, and it is worth one node in a thousand to a solver. It can never fire on an empty rectangle at all, which is why the ladder's whole subject is the half turn.

Two counters, not one. The four periods of the odd-saltus class against the two base-three counters, with which each follows.

Two counters, and one displaced term

The rung below found four Grundy sequences in the odd-saltus class and asked which term each displaces and whether the digits predict it. They do — but there are two base-three counters and not one, chosen by whether a heap of one can be taken away. And there are three sequences rather than four: the fourth is the third with three isolated values, and was counted separately because its period had not settled.

Four conditions. The four linear conditions whose kernels are the losing sets, with the cases each covers.

The parameter was the difference

The losing words of bounded Moore's Nim form a linear subspace and no map was known whose kernel they are. The equations exist, four conditions cover all thirteen cases at three to six heaps, and they are indexed not by the heap count but by the heaps less the width of a move — which turns the failure at six heaps into a prediction about seven.

The clause that was free. Five requirements on a pairing strategy, with which of them each map meets.

A symmetry that is not a pairing

The quarter turn was the last symmetry a Cram pairing argument had not tried, and the one a square board seemed to offer. It fires on the empty four by four and it settles nothing the half turn misses — and the reason is a clause four rungs of this anchor never had to write down, because every map tried so far was its own inverse.

Not rare at all. How many hexadecimal codes whose sequence settles have a stretch of heaps before the pattern begins.

A pattern that has not started yet

A pre-period was supposed to be rarer in this family than a defect. Two hexadecimal codes in five have one, 321 have a pre-period longer than their own period, and the code the rung below found slow takes fifty-four heaps to settle rather than two blocks — which is also the account of three defects the rung below recorded and could not explain.

Five of six. The six predictions made for seven heaps by the difference reading, each scored against the sweep that was declined at the time.

The family with two witnesses

Six predictions about seven heaps were written down and deliberately not run. Five of them held. The one that broke is the condition that had been checked against two cases when it was proposed — the fewest of the four — and at seven heaps it does not merely give the wrong answer, it asks a question the parity word has stopped being able to answer.

A pairing no motion of the square gives. The smallest Cram shape carrying a pairing that is not a rigid motion, with its three pairs drawn as lines between the squares they join.

A pairing that is not a symmetry

Every pairing strategy this ladder has found is a rigid motion of the square, and the requirement mentions no geometry at all. Searching all 8.8 million fixed-point-free involutions instead of the eight maps more than doubles what a pairing explains — and the smallest new one turns out to be a reflection with its two fixed squares swapped.

One of the two quantities is inert. The candidate predictors of a pre-period's length, each scored by correlation against the measured length.

The quantity that carried nothing

The rung below proposed predicting a pre-period's length from the saltus and the period. The saltus correlates with it at −0.03, which is nothing; the period correlates at 0.77 with a coefficient of one, so a pre-period is about one period long. The digit that predicts whether there is one predicts nothing at all about how long.

The dual is the value table's span. The dual code against the span of the bit-planes of the one-coin Grundy values, on every game measured.

The dual was the value table

A coin-turning game's losing rows form a linear code, and a code has a dual that nothing in the game appeared to read. It reads it constantly: the dual is spanned by the bit-planes of the Grundy values — the parity checks are the value table stood on end — and on Mock Turtles over eight coins the losing rows are exactly the span of the table that decides them.

Two statements, two routes. The two measured identities the rung below left unproved, with the argument each was expected to need.

One proof, and one wrong lemma

Two measured identities were left for a proof: the move rule by induction, the gap condition from the reply bound. The induction is exact on 31,731 heaps at eight factors. The reply bound holds at c = 2 and on one index pair in twenty-seven at c = 3 — and the inequality that does the work is a third one nobody proposed.

Four readings, one game. The closed form, the strict mating, the cancelling matching and the parity term, with how much of the game each accounts for.

The pairing the formula hides

Welter's closed form sums a function over every pair of coins and needs an extra term when the count is odd, which the rung below called a surprise. Read as a matching it is not: an odd number of coins cannot be paired, the left-over coin contributes its own square, and some matching gives the value on every position measured.

What the theorem replaces. Every grid a brute-force solve can reach, valued both ways, with the grid the theorem is normally drawn at underneath. Twelve squares is four thousand arrangements against twelve products; sixty-four squares is eighteen quintillion against sixty-four.

Four hundred and seventy steps

The tartan theorem replaces a search with a multiplication. Measured on every grid a brute-force solve can reach, the two agree on all of them — and the ratio doubles with every square added. On the 8 × 8 grid the theorem is normally drawn at, the search would have to value eighteen quintillion arrangements; the theorem needs twenty-six different nimber products, and computing all of them by the rule that defines them looks at four hundred and seventy pairs.

Lasker's Nim in sixteen cells. A four-by-four table. Each row and column is a residue mod 4 of one part of a split heap, with the residue of that part's Grundy value beside it; each cell is the residue mod 4 of the split's value, the nim-sum of the two parts. Every split of every heap to four hundred lands in the cell its residues name.

The proof is sixteen cells

Lasker's Nim has a four-clause formula that was checked on two thousand heaps and never proved. The proof fits in a four-by-four table: the last two bits of a split's value are fixed by the last two bits of its parts, so no split can land in its own heap's class — except at 3 mod 4, where it lands exactly on the one value the takes leave missing and pushes the answer up by one.

One split is enough

A heap of n in Lasker's Nim offers ⌊n/2⌋ ways to split, and the values use at most one of them. Allow only the split that takes a single counter off and every heap to six hundred keeps its value; of all sixty-three sets of split sizes up to six, a set keeps the formula exactly when it contains 1 or 2. Equal halves alone give back plain Nim, because a split into equal parts is a move to nought.

The formula as the limit of periodic games. Lasker's Nim above eight versions of it with the number of counters that may be taken bounded at one to eight. Each bounded game is periodic and agrees with Lasker's formula on its first few heaps; the region of agreement grows with the bound.

The formula is a limit

Cap the take in Lasker's Nim at k counters and the game is a finite rule table, 4.33…3, whose Grundy sequence repeats with period k + 1 rounded up to even and follows Lasker's formula until the cap bites. The formula is what those periods converge to. And the same column of codes, with a free split in front, holds Kayles itself: the rule 4.4 on a heap of n + 1 is Kayles on a row of n.