In this paper
AbstractIntroductionThe field, the graph, and the local groupLarge idempotents with ordered annihilationsThe annihilated column spaceMany nonsingular diagonal pairingsRealizing the projections in the group algebraFrom opposite annihilations to an ordered zero productGluing the finite groups without collapsing themAn injectivity criterionApplication to the modular constructionRealizing the local ranks in a common central sectorAttaching the finite groupsThe matrix blocks of the central sectorTurning dimensions into rectangular identitiesCompression with a visible nonzero defectAn injective finite-group encodingThe cellular automaton on all configurationsReferencesWorksMathematics
A Counterexample to Kaplansky's Direct-Finiteness Conjecture in Odd Characteristic
Abstract
We construct a counterexample to Kaplansky's direct-finiteness conjecture in odd characteristic. For one specified odd prime , we obtain a field of order , a finitely generated group containing torsion, and finite sums with but . The same elements define a cellular automaton on that is injective but not surjective.
Introduction
A group algebra consists of finite sums , with , multiplied by the group law and bilinearity. A unital ring is directly finite when implies ; it is stably finite when this implication holds in every finite square matrix ring. Kaplansky’s direct-finiteness conjecture asks whether is directly finite for every field and every group , including groups with torsion. The result here is a negative answer in one specified odd characteristic.
Theorem 1.1. Put
Then is an odd prime. There are a specified field of order , a specified finitely generated group containing torsion, and specified finite sums satisfying and .
The word “specified” refers to finite formulas and choices in ordered finite sets. The field is a quadratic tower, the group is built from finitely many finite groups and explicit identifications, and the sums are defined without solving the word problem of that group. The construction is not claimed to have been evaluated as a practical computation. It concerns the displayed prime, not every odd prime.
The finiteness problem. For finite , the algebra and all its finite matrix rings are finite-dimensional over . A one-sided inverse between their left multiplication operators is therefore two-sided. This argument works even when the characteristic divides . The difficulty for an infinite group is to obtain a corresponding conclusion from finite support alone.
Kaplansky proved stable finiteness in characteristic zero using finite von Neumann algebras, and asked whether the one-sided-inverse property persists in positive characteristic [19 pp. 122–123]. Ara, O’Meara, and Perera proved stable finiteness for free-by-amenable groups over arbitrary division rings [2]. Elek and Szabó [10 Corollary 4.7] established the same conclusion for all sofic groups. Such groups admit finite permutation models that approximate multiplication while keeping distinct elements separated. These theorems are constraints on the example, not hypotheses about its group. In particular, Elek–Szabó’s theorem implies, after the scalar construction is completed, that cannot be sofic.
A separate characteristic-two construction uses incidence structures, polynomial pairings, a planar group-embedding argument, and character packing [20 Theorem 1.1]. The construction here has a different local mechanism. A high-degree subspace of a modular translation representation is annihilated on one side; a weighted pairing gives a compatible annihilation on the other. The characteristic-two theorem is not an input to this odd-prime argument, and the two methods retain distinct coefficient hypotheses.
The two quantities that must be compared. A matrix fits a simple graph when and for distinct nonadjacent vertices. We use a triangle-free graph on vertices and a specified factorization
This is an explicit threshold-Kneser instance of the polynomial constructions of Haviv and Golovnev–Haviv [15 Proposition 13, 11 Theorem 3.2 and Lemma 3.5], whose polynomial ancestry includes Alon–Babai–Suzuki [1]. Low-rank fitting matrices also give storage-code recovery equations. The triangle-free constructions of Barg–Zémor, Huang–Xiang, and Barg–Schwartz–Yohanov provide related context [4, 16, 3]. Their results are not used to supply a group-algebra idempotent.
At every vertex we use a copy of a finite group and an idempotent . Write and , the rank of left multiplication by on the regular representation. The second quantity is
The idempotents satisfy a directed condition: after identifying certain cyclic subgroups at adjacent vertices, for every edge with . There is no assertion about the reverse product. The asymmetry is what allows a triangular calculation to combine the two strict bounds.
The characteristic-dependent construction appears in Section 3. High augmentation degree gives the annihilated columns, in the elementary abelian setting of Jennings’ theory [17]. A diagonal weighted Gram determinant and Schwartz’s grid bound [23 Lemma 1 and Corollary 1], with related polynomial-identity work of Zippel [25], supply enough nondegenerate pairings. Fourier character orbits turn these projections into actual elements of . The full algebra need not be semisimple; only its prime-to- normal abelian subgroup is diagonalized, in the character-orbit setting of Clifford [8].
From the comparison to scalar elements. The proof has three transitions that numerical rank inequalities alone cannot justify. First, the quotient identifying the cyclic subgroups must retain the finite vertex groups. Its injectivity proof uses the – graphical-complex mechanism described by Prytula [22 Section 8]: local short-word restrictions and triangle-freeness rule out a spherical map with exactly one nontrivial face. Ordinary graph-of-groups normal form does not by itself establish this first quotient’s injectivity. Section 4 proves it locally, including the reductions for loops and bridges.
Second, regular ranks must become identities over the same ring. A further group enlargement has a nonzero central idempotent in . In the ring , whose identity is , the elements admit rectangular matrices
The finite-group blocks in this step have elementary abelian -group stabilizers, possibly nontrivial and nonnormal. Their group algebras have nilpotent augmentation ideals, and a finite change of basis proves exact integer rank identities before coordinates are paired. Section 5 gives the full construction of these matrices, with their entries in explicitly attached finite subgroups.
Third, a matrix counterexample does not yet give two scalar elements of the same group algebra. In Section 6, the fit and the ordered zero products produce a split injection
whose composite is the identity. We derive square matrices with a one-sided inverse and certify the reverse defect by a nonzero kernel vector. Section 7 then implements the finite-factor reduction of Dykema–Juschenko [9 Lemma 2.1 and Theorem 2.2]. A finite Heisenberg group gives explicit matrix units, their linear independence makes the encoding injective, and a complementary idempotent restores the full scalar identity. The shift and modulation model is classical [14 Section 1.2]; the composite cyclic order and finite coefficient field used here are checked directly.
The dynamical consequence. A configuration over a finite alphabet is any function . A cellular automaton has one finite memory set and one rule applied to at every . A group is surjunctive when every injective such automaton, over every finite alphabet, is surjective. Gottschalk proposed this for all groups [12]; Gromov and Weiss proved it for sofic groups [13][24]. The group-ring connection is recorded by Elek–Szabó and developed in the linear theory of Ceccherini-Silberstein–Coornaert [10][5]. Phung proved stable finiteness over every field for surjunctive groups [21 Theorem B]; subsequent proofs and extensions include [6][7].
Section 8 gives the elementary transfer with its actual multiplication convention. The coefficients of define an injective automaton, those of give a left inverse, and a nonzero coefficient of specifies a missing point indicator. The argument excludes preimages of arbitrary support. It is a consequence of the scalar construction, not a further existence theorem.
All matrices below act on columns of free right modules. Matrix entries therefore multiply coordinates on the left. Transpose is ordinary transpose over , with no involution. Cardinalities and dimensions in rank comparisons are integers, not elements of the coefficient field. For a positive integer , write .
The field, the graph, and the local group
We first construct the finite data. This fixes every coefficient field and denominator before the modular argument begins.
Lemma 2.1. For in Theorem 1.1, is an odd prime and is a square in . If is a nonsquare in , then
are quadratic field extensions, so . Proof. A least divisor greater than one of an integer is prime: any nontrivial proper divisor of a composite least divisor would be smaller. If , then divides both and , which is impossible. Hence , and the defining divisibility gives in . There are nonsquares because each nonzero square has exactly two square roots. The first quotient is therefore a quadratic field. The conjugate of is , so its norm is . Since is square, is nonsquare. A square in would have square norm, proving that is nonsquare and that the second quotient is a field.
Choose the least nonsquare using representatives ; order by its coefficients in . Put and choose the least element of order . This element exists: if is the least common multiple of all orders in , then , while the root bound for gives . For each prime-power factor of , an appropriate power of one element has that exact order. Multiplying the resulting commuting elements of coprime orders gives order . In particular, and all its powers are invertible in .
Use natural orders on integer indices, increasing-list lexicographic orders on subsets, and lexicographic orders on tuples and finite function tables. Order characters of by their exponent lists on the ordered coordinate generators. These conventions specify the finite choices below.
Let be the -subsets of in order. Join distinct in exactly when . Put
Proposition 2.2. The graph is triangle-free and has a specified fitting factorization over with .
Proof. Three pairwise adjacent sets would have union of size at least , exceeding 1200. To construct the fit, let when , and zero otherwise. Define as the coefficient of obtained by expanding
and replacing every positive exponent in each monomial by one. All denominators are in , hence invertible. This replacement preserves evaluations on zero-one vectors and does not increase degree. Thus
It equals one for . For distinct nonneighbors the intersection size lies between 200 and 599, so one factor is zero. This is the threshold-Kneser polynomial specialization of [15 Proposition 13] and [11 Theorem 3.2].
The binomial coefficients increase up to index 400. Consequently
In the second inequality the first hundred factors are at most and the rest are at most one. The last inequality follows from and .
For the local algebra define the additive group , the function group , and
Its order is . Let have value one at zero and zero elsewhere, and put
The coordinate formula gives , because when . We next construct an idempotent with incoming annihilation by and outgoing annihilation by .
Large idempotents with ordered annihilations
The graph has supplied the ratio . Our present objective is an idempotent in the finite group algebra whose regular rank fraction exceeds . It must also have two prescribed annihilations, on opposite sides. Here
The regular rank of means , the dimension of the image of left multiplication. Although this map multiplies on the left, its image is a right ideal.
Proposition 3.1. There is a specified idempotent satisfying, for every ,
Its regular rank satisfies . Its coefficients are obtained by finite operations in the specified field .
The annihilated column space
On , let be the standard basis and . For the three standard generators of set . The commute, and characteristic gives . The vectors
are a basis. Indeed, every translation of is a product of powers of applied to , so the displayed family spans ; its cardinality is the dimension .
This is the elementary abelian augmentation filtration in the setting studied by Jennings [17]. Here the exact degree argument can be given directly. Let have as columns the displayed basis vectors with , in order. For every , the matrix is a polynomial in the with zero constant term. Multiplication by its th power raises every column’s total degree to at least . Such a monomial has an exponent at least , so vanishes. Applying the same assertion to and transposing yields
The transpose identity uses ; no scalar conjugation is present.
The columns are independent. Their number can be counted by setting : the new nonnegative indices have sum at most . Their individual upper bounds are automatic because . Stars and bars gives
For the strict inequality, the numerator minus is for .
The same subspace is killed on the two sides in Equation (3.1), but an arbitrary projection onto it need not give the outgoing conjugate annihilation. We therefore choose the row factor by a diagonal weighted pairing.
Many nonsingular diagonal pairings
For a finite function group , its characters are all exponent lists of powers of . Put
The denominator is invertible. Reindexing the sum gives . A nontrivial character has sum zero, since translating its sum by an element on which it is nontrivial multiplies it by a scalar different from one. It follows that
For the last identity, sum coefficients over the independent character exponents: every nonzero coordinate of a nonidentity element is detected by a power of . Each idempotent is nonzero, there are of them, and orthogonality makes them independent. They form a basis of .
Specialize to . Its character action is . Define
Call admissible if its stabilizer in is trivial and is invertible. This condition is a finite test.
Lemma 3.2. More than half of the characters are admissible.
Proof. For each nonzero , translation partitions into cycles of length . A character fixed by has constant exponent on each cycle, so exactly characters are fixed. A union bound therefore bounds the fraction with nontrivial stabilizer by
Let be the coordinate generator at . Since , the diagonal entries of range independently over as ranges over all characters. For independent variables , Cauchy–Binet gives
Here consists of the indicated rows in their fixed order. Some maximal minor is nonzero because has full column rank. The monomials for distinct are distinct, so is nonzero of total degree , even though the unweighted Gram matrix could be singular.
We use the polynomial grid estimate of Schwartz [23 Lemma 1 and Corollary 1]; see also the related identity-testing method of Zippel [25]. A nonzero polynomial of total degree at most vanishes on at most an fraction of a product grid whose coordinate sets each have size . To prove this, induct on the number of variables. If the degree in the last variable is , its leading coefficient has degree at most . It vanishes at at most an fraction of the earlier choices. For each remaining choice, the last-variable polynomial has at most roots, contributing at most . The sum proves the bound; the constant case starts the induction.
Applying the estimate on to (3), the singular-pairing fraction is at most . Both excluded fractions are strictly smaller than :
The estimates use ; for example, . Their union therefore excludes fewer than half of the characters.
Select the first admissible character from each -orbit containing one, and call the set of selected characters . Each selected orbit is free, of size . These orbits cover all admissible characters, so
This argument does not need admissibility to be constant on an orbit.
Realizing the projections in the group algebra
Fix . For , define
Conjugation gives . Fourier orthogonality and freeness of the orbit therefore imply
Indeed the Fourier product can survive only when , which is equivalent to . Products between different orbits vanish. The sum is central: it commutes with and its summands are permuted by . The basis of its ideal is indexed equally by through . Thus these units prove
This is an explicit character-orbit block, in the normal-subgroup representation setting of Clifford [8]. It does not assert semisimplicity of the entire modular algebra.
On this block, the elements , , and act respectively by
For example, left multiplication by shifts row to row , whereas multiplies row by .
Define a matrix and then a finite group-algebra element by
The inverse exists by the selection criterion. Since , we have and . Its image is exactly , of dimension . Orthogonality of the blocks proves ; on the complementary ideal its component is zero.
We now compute the regular rank, rather than the smaller rank of a single representation matrix. In , left multiplication by has image the matrices whose columns all lie in . Its image dimension is . Summing the selected blocks gives
The incoming annihilation follows from . For the outgoing annihilation, keep the factors in their actual order:
Equation (3.1) proves both assertions on every selected block. They also hold on the complementary ideal, where is zero. This proves Proposition 3.1, including its finite prescription and the strict rank estimate.
From opposite annihilations to an ordered zero product
Take copies of . For each edge , identify in with in , where bars denote residues in . Let be the resulting free-product quotient, and let denote the image of from . The injectivity of these copies is proved in Section 4. The following ordered identity already follows from the local calculations and the quotient relations.
Corollary 3.3. For every edge of , in .
Proof. Let be its generator, initially of order on either local side, and put . In the local cyclic group algebra , the annihilator of right multiplication by is precisely the ideal . Choose representatives of the left cosets in ; they give a free right -basis of . Thus the outgoing identity , applied coefficientwise in this basis before taking images in the quotient, gives for some local . On the incoming side . Since ,
No factor is commuted across , and no reverse ordered product is asserted.
We have constructed the large regular rank and the edge vanishings. To compare them with the graph factorization, the remaining structural step must put exact rectangular realizations of all those ranks inside one nonzero ring.
Gluing the finite groups without collapsing them
The quotient of Section 3 identifies a cyclic subgroup across each edge of . It must retain every vertex group: otherwise a later nonzero element of a finite group algebra could disappear. We first prove a general injectivity criterion and then verify it for the groups used here. The local injectivity argument belongs to the C(4)–T(4) theory of graphical complexes of groups [22 Section 8]. We give a direct proof using only labelled spherical maps and Euler’s formula.
An injectivity criterion
Lemma 4.1 (Triangle-free gluing). Let be a finite simple triangle-free graph. For each vertex , let be a finite group. For each edge , let be a nontrivial finite cyclic group, with injective homomorphisms and . Assume the following condition at every vertex . If , the edges are incident to , and , then
whenever consecutive edge labels are distinct cyclically. For the distinctness condition is omitted; for it includes . Then every natural homomorphism is injective, where
Proof. We first handle the subgroups generated by the incident cyclic groups. The diagram argument will rule out a relation that is nontrivial at one vertex but follows from the relations at all the vertices. At the end, a permutation action extends the conclusion to the full groups .
Put
Choose a generator of each , and let have one generator for each edge. Its defining relators are all words in the incident to a single vertex that evaluate to in that vertex group. There is a well-defined map : two words representing the same element of differ by one of these relators.
A failed injection produces a spherical map. Suppose a nonidentity element of maps to in , and choose a word representing it. There is a finite sequence of words from to the empty word, each step inserting or deleting a defining relator, its inverse, or a free inverse pair. Indeed, membership in the normal closure of the relators expresses in the free group as a finite product of conjugates of relators; free insertions and cancellations reach such a product, and the conjugated relators can then be erased.
We describe the spherical maps needed to encode this sequence. A map is a finite connected graph drawn tamely on the sphere, allowing loops and parallel edges. An edge is either neutral, with value , or is labelled by an edge of . In the latter case its two oriented traversals, called darts, have mutually inverse values in . Every face has a type incident to every label in its boundary walk. Its boundary word is evaluated in , using the specified embeddings. A face is ordinary when this evaluation is the identity. Call the map exceptional if precisely one face has nonidentity evaluation and all other faces have identity evaluation. Changing the starting point conjugates a boundary product, and reversing the walk inverts it, so identity versus nonidentity is well-defined.
Draw the successive words in horizontal levels of a rectangle, with each letter an oriented edge. Represent an empty level by a single neutral edge. In every intervening strip join corresponding vertices along the unchanged prefix and suffix by noncrossing neutral edges. At an insertion or deletion, join the single insertion vertex on the shorter path to the two endpoints of the inserted block on the longer path. The strip faces then read either a letter followed by its inverse, or the inserted relator, with neutral edges interspersed. If a level is empty, the adjacent nonempty word is the entire inserted or deleted block; one face between these two levels suffices. Assign the relator’s vertex type to its face, and either endpoint type of a letter’s label to an unchanged-letter face. All these faces evaluate to . Include the neutral edges along the two vertical sides of the rectangle. The exterior face has word or , apart from neutral edges; give it type . This connected drawing, viewed on the sphere, is exceptional.
We shall prove that no exceptional map exists. Choose one with the smallest possible number of edges. This number is positive: the only face of an isolated vertex has the empty word, which evaluates to . The reductions below will make every vertex degree and every ordinary face length at least four away from a possible loop. Euler’s formula will then rule out both the loop-free case and the remaining one-loop case.
Boundary walks and elementary reductions. Fix an orientation of the sphere. At each vertex let be the cyclic successor of outgoing darts and let reverse a dart. Boundary walks are the cycles of : after traversing a dart, take the successor of its reverse. For a disconnected drawing we use each component’s own boundary walks, counting one empty walk for an isolated vertex. If a drawing has vertices, edges, such walks and components, then
For completeness, in a connected spherical drawing a nonbridge edge lies on a simple circuit. Its two sides are therefore on distinct boundary walks; deleting it merges those walks and preserves connectedness. Repeated deletion leaves a tree, which has one boundary walk and . This proves the formula for a connected drawing; summing proves the displayed formula. The same walk rules follow by taking small vertex discs and thin edge bands.
A neutral nonloop edge can be contracted, with its neutral darts omitted from the boundary words. This preserves every face evaluation and contradicts minimality. For a neutral loop based at , retain the closed side containing the exceptional face, then remove the loop. The retained graph is connected: each side of the loop attaches to the other only at . The other dart slots on the retained side form an interval between the loop slots; closing that interval gives precisely the old walks on that side with the neutral loop dart removed. Thus their evaluations are unchanged. If only remains, its walk is empty and could not be exceptional. In every possible case a smaller exceptional map results. Hence every edge is nonneutral.
There are no bridges. To see this, delete a bridge with value , leaving two components. Its incident face walk splits into words , one in each component, possibly empty. Both receive the old face type. With suitable starting point and orientation, the old word was
If it was exceptional, at least one has nonidentity evaluation; retain that component. If it was ordinary, the evaluations of and are either both identity or both nonidentity. In the first case retain the component with the original exceptional face. In the second retain the component without that face: its new boundary walk is now its unique exceptional face. All choices give a smaller exceptional map.
Every edge consequently separates two distinct faces. Their types must also be distinct. Otherwise delete the edge and merge the faces, using their common type. If their words, starting at its opposite darts, are and , the merged word is , and
Thus two ordinary faces merge to an ordinary face, whereas one ordinary and one exceptional face merge to an exceptional face. Again this would reduce the number of edges. The types across an edge labelled are therefore exactly the two endpoints of .
There is at most one loop, and if it exists it bounds the exceptional face as a monogon. In fact, given a loop, keep its closed side without the exceptional face, including the loop and its base vertex, and erase the graph on the other side. All retained old faces are ordinary. The erased side becomes a monogon with the loop’s nonidentity value; either endpoint of its label supplies its type. This is another exceptional map. Minimality forces the erased side to have contained no other edges in the original drawing. That side was consequently already the exceptional monogon. This applies to every loop, and the one exceptional face cannot be a monogon for two different loops.
Consecutive equal labels can be smoothed. We have reduced to nonneutral, nonbridge edges with different face types on their two sides; the sole possible loop bounds the exceptional monogon. It remains to rule out consecutive nonloop darts with equal labels and then compare degrees with Euler’s formula. The remaining reduction ensures that the short-word hypothesis applies to face boundaries. Suppose a face has consecutive nonloop darts
with a common label and respective values . The two darts belong to distinct edges: otherwise the edge would occur on both sides of and would be a bridge. Let be the faces containing , respectively. They are distinct from and have a common type , namely the other endpoint of from the type of . They may coincide with each other.
Remove the two edges and insert an edge along their arcs, smoothed through the empty corner of at . Give it value and label , or make it neutral when that product is 1. If , this produces a loop. Keep all vertices, including any that become isolated. At the new darts occupy the old slots of ; at delete the consecutive slots . These instructions specify the spherical embedding even when vertices or other boundary segments coincide. The walk of replaces by and has unchanged evaluation.
When , write
The walks avoid the removed edges, so the new graph remains connected. The two opposite faces merge to the walk . All other walks, apart from the specified change in , are unchanged. Its evaluation in is
Thus the merged face is ordinary when both old faces were ordinary, and exceptional when exactly one was exceptional. This yields a smaller exceptional map.
The case requires keeping track of components. Write
The slot rules replace this single walk by the two walks
One can read this directly from : the transition from is the old transition from , and removal of the two slots at joins the end of to its beginning. If is empty, the old transitions were and . Equivalently, and , so had degree two and now becomes an isolated vertex with its stipulated empty face. If is empty, then and is the monogon of the new loop. Thus the description includes all empty cases. Figure 1 illustrates the local replacement and the two boundary walks; all other incidences are suppressed.

Figure 1. Smoothing two consecutive nonloop darts of one label. All other incidences are omitted. The opposite faces may be the same face; in that case its boundary splits as shown below the local picture. The resulting drawing has two components, as the boundary-walk calculation shows. If is empty, its component is the isolated vertex .
Across the components, the total number of boundary walks increases by one, the edge count decreases by one, and the vertex count is unchanged. Formula (4.2) gives exactly two components. The new edge places in one of them. Every component must contain an endpoint of a removed edge, because the old graph was connected; hence is in the other component. In particular, the two walks in (8) belong to different components.
Give both walks type , and denote their evaluations by
The old word has evaluation . Moving the first factor to the end is a cyclic change of starting point, so it preserves identity versus nonidentity. The resulting evaluation is
If was exceptional, at least one of is nonidentity; retain its component. If was ordinary, and are either both identity or both nonidentity. In the former case keep the component with the original exceptional face, whose evaluation is unchanged even if it is . In the latter keep the component without that face; its new boundary is its sole exception. This again gives a smaller exceptional map. We have proved that no face of the minimal map has consecutive nonloop darts with equal labels, including the last and first darts of its cyclic boundary walk.
Euler’s formula rules out the minimal map. Consider a vertex that is not the base of a loop. Reading the adjacent face types around it gives a closed walk in . Each incident edge crosses between its two different endpoint types, and successive edge labels are distinct by the smoothing reduction. Thus the walk has no immediate reversal. Its length is positive since the connected map has edges. A simple triangle-free graph has no such closed walk of length less than four, so this vertex has degree at least four.
An ordinary face not incident to a loop has a nonempty boundary walk. Indeed every boundary walk of a connected graph with at least one edge is a nonempty dart cycle; an empty walk occurs only for an isolated vertex. Its consecutive labels are distinct, and the short-word condition (7) therefore forces its length to be at least four. Here and below a loop counts twice toward its vertex degree, and boundary lengths count dart occurrences.
For our connected spherical map, Euler’s formula and the two degree sums give
Without a loop, only the exceptional face can contribute positively, and its nonempty walk contributes at most 3. With a loop, its base vertex contributes at most 2, the exceptional monogon contributes 3, and the face on the other side of the loop contributes at most 2: that face is ordinary and cannot have length one, since its sole letter would be the loop’s nonidentity value. There are no other loops, so every remaining vertex and face contributes at most zero. The total is at most , again contradicting (9). Consequently no exceptional map exists, and every is injective.
Extending to the full vertex groups. Let , with acting by left multiplication on the first coordinate. The restriction to each embedded is free and has countably infinitely many orbits. The same is true of the left -action on . Pair their orbits and choose one point in each pair; this determines a -equivariant bijection . Transport the free left -action along this bijection. Its restriction to is exactly the original -action. Make this choice separately for each . The resulting actions agree on every edge group because both copies of its chosen generator act as . They therefore define an action of on , faithful on every . This proves the required injections.
Application to the modular construction
Recall from Section 2 the group , with and . Its subgroups and satisfy . The quotient identifies, for every edge , the elements
where denotes the residue of in . Both elements have order , so the corresponding maps from are embeddings. We now verify the remaining local condition.
Proposition 4.2. Every natural map is injective. In particular, the maps induced on group algebras are injective.
Proof. We verify the short-word hypothesis of Lemma 4.1. At a fixed vertex, every incident edge subgroup lies either in or in . Within either subgroup, the labels supplied by distinct neighbors correspond to the vectors ; a nonidentity syllable corresponds to a nonzero scalar multiple of its vector. Any three with distinct residues are independent, because their determinant is, up to sign,
The residues are distinct since . The analogous first two coordinates show independence for two vectors, and a single vector is nonzero. Thus a word of at most three nonidentity syllables with distinct labels, all in or all in , cannot evaluate to .
For a mixed word of length two or three, one of supplies exactly one syllable. After a cyclic permutation, a putative relation would equate that nonidentity syllable to a product of syllables from the other subgroup. This contradicts . For lengths two and three, cyclically distinct consecutive labels are precisely distinct labels, so all required words have been covered. The graph is simple and triangle-free, and all groups in question are finite. The lemma proves the injections. The group-algebra assertion follows because distinct elements of an embedded subgroup remain distinct basis elements of .
Realizing the local ranks in a common central sector
An idempotent in a finite group algebra has a well-defined rank as a linear operator. We now turn each of these numerical ranks into an isomorphism between modules over one common ring. The enlargement of the group is needed because the different simple or modular summands of the original group algebra need not have the same rank fraction.
Here are the precise inputs to this step. Let be a field of characteristic , let be a positive integer prime to , and assume that contains an element of order . In the application, and . Write and put
where acts on the functions in by translation. Suppose that a countable group contains specified copies of . For each , let be an idempotent, and assume that the positive integer
is independent of . All matrices below act on column vectors in free right modules, so their entries multiply coordinates on the left.
Proposition 5.1. Under these assumptions, there is a group containing and a central element of order with the following property. Set
Then is a nonzero central idempotent. There are explicitly specified matrices
satisfying
The entries of and lie in the group algebra of a finite subgroup of , multiplied by . Their construction uses only finite group operations and linear algebra over .
We prove the proposition in three stages. First we attach the finite groups while preserving all the given embeddings. Next we describe their -parts as matrix algebras over group algebras of elementary abelian -groups. Finally, nilpotence of the corresponding augmentation ideals makes the ranks exact and produces the two rectangular matrices.
Attaching the finite groups
Let initially generate an independent copy of , and write . For each , define
The groups are written additively. Let be the constant function of value . It has order and is fixed by , so it is central in . Moreover, . Thus embeds in both and , with sent to in the latter. Define to be the quotient of the free product
that identifies these two copies of for every .
For completeness, the following action proves that all the factors embed in this quotient. Let be a countable disjoint union of copies of the left regular -set. Its restriction to each is free and has countably infinitely many orbits. A countable disjoint union of regular -sets has the same property when restricted to . Pair their -orbits and choose one point in each paired orbit; extending those point pairings equivariantly gives a -equivariant bijection . Transport the regular -action to using this bijection. The resulting actions of agree on every prescribed , hence define an action of on . Each factor acts faithfully, which proves its injectivity in . This argument uses the assumed embeddings in defining the subgroups .
The element is central in every generating factor of and therefore in ; its order is still because embeds. In , direct multiplication gives
The coefficient of the identity in is , since the elements are distinct. Thus is a nonzero central idempotent, and is a ring with identity . Each is an idempotent in . Also, any identity already valid in remains valid after multiplication by in .
The matrix blocks of the central sector
Fix for the moment. A character of means a homomorphism . If denotes the coordinate function with value at and elsewhere, then every character is uniquely specified by exponents through . Define
Apply the Fourier calculation of (2) to the finite index set . The denominator is invertible because . The elements are a basis of pairwise orthogonal idempotents whose sum is , and .
Use the action
Then . Since is fixed by , the set
is -invariant, and the Fourier formulas give
For each orbit in , choose one representative and put .
Lemma 5.2. For each such , the subgroup is elementary abelian of exponent and rank at most , with the trivial subgroup allowed.
Proof. Write . Since , the condition on is
Suppose that a nonidentity has order . The exponent of divides , so divides . Left multiplication by partitions into cycles of length exactly . Invariance of says that the exponents are constant on these cycles. Their sum consequently lies in . It cannot equal 1: reduction modulo would give . This contradiction proves . The quotient homomorphism is therefore injective on . Every subgroup of the additive group is an -linear subspace, giving the assertion.
Eliminating the prime-to- part of each stabilizer is what makes its augmentation ideal nilpotent, and will permit the later change of basis. We next give the block isomorphism, including its inverse. This is the character-orbit construction already used in Section 3, now allowing nontrivial stabilizers; compare the Clifford-theoretic orbit picture in [8]. It will identify the underlying -linear rank of in every block with the same original integer . For one orbit representative , abbreviate
Choose representatives of the left cosets . The space is a free right -module with this ordered basis. Let act on it by left multiplication , and let act diagonally by
Both actions commute with the right action of . For the diagonal action this follows from for . The semidirect-product relation is also respected, because
Thus we have an algebra homomorphism . The element acts as , so .
For and , consider
It belongs to by (11). In the -representation it sends to for every and vanishes on every other basis coset. To check the last assertion, its left group factor sends the coset to ; the Fourier projection keeps this coset precisely when it equals , which is equivalent to . In a representation associated to another orbit, the Fourier projection vanishes identically. Consequently (5.3) maps to the matrix with entry in position of the -block and zero elsewhere.
Taking the product of these representations gives a surjective map
It is injective as well. The elements , for and , form a basis of its source. An orbit of size contributes basis elements, equal to the dimension of the corresponding target block. Hence the source and target have equal finite -dimension. Formula (5.3), extended linearly in and summed over the positions and blocks, is an explicit inverse of .
Turning dimensions into rectangular identities
Continue in one block and let . As a -linear operator on , is exactly left multiplication by , since acts as the identity. Therefore
This is the link between the original rank calculation and the block decomposition. We will now obtain a basis in which is a coordinate projection over , not just over . The argument is a constructive nilpotent-ideal instance of the freeness of projective modules over local rings, studied by Kaplansky [18 Theorem 2]; no general projective-module theorem is needed here.
Let be augmentation, and set . By Lemma 5.2, we may choose generators of with . The elements commute, satisfy , and generate . Expanding and counting dimensions gives
It follows that
Indeed, a monomial of total degree greater than has an exponent at least . This also covers , when and . Matrices with entries in have the same nilpotence bound under multiplication, because each entry of a product of such matrices lies in .
Put , applying augmentation entrywise, and let . As is idempotent, its image and the image of form complementary subspaces of . Scan the columns of in order and keep each column that increases their span; then do the same for . Build by taking the corresponding columns of , followed by the corresponding columns of . Its augmentation is invertible over . Regard as a scalar matrix over , and set . Its entries belong to , so . Since ,
Each of the first columns of is fixed by , and every remaining column is killed by . Thus, for ,
Conjugation by preserves -dimension of images. Comparing with (5.5) gives the integer equalities
In particular, these are equalities of ordinary dimensions, not equalities after reduction to .
We can now match coordinates exactly. Order pairs lexicographically and choose the increasing bijection between
the sets have equal size by (5.7). Let be the -by- block matrix with -by- entries that sends coordinate to its paired coordinate . Thus is the sum of standard matrix units over all such pairings. Let be the -by- block matrix reversing each pairing, with the corresponding sum of . It sends all unpaired coordinates with to zero. Directly on these coordinates,
Define block matrices
They have outer shapes and , respectively, with entries in . Their products are
Finally define the entry of by applying to the tuple , and similarly define entrywise from the . The isomorphism (12) carries to the tuple of identity matrices and to the tuple of matrices . The last two identities therefore pull back to (5.1). This proves Proposition 5.1.
All choices in the matrix construction are over finite sets: character orbits, coset representatives, augmentation matrices, and coordinate pairings. Taking the first element or the first independent columns in fixed orders makes the prescription definite. The inverse (5.6) is a finite sum, and the inverse (5.3) is a finite group-algebra formula. Thus the matrices furnished by the proposition are finite sums in the explicitly attached finite groups.
We now apply Proposition 5.1 to the embedded groups of Proposition 4.2 and the identical idempotents of Proposition 3.1. The group is countable because it is finitely generated. All the ranks equal the positive integer from Equation (6). Hence the group , the nonzero central idempotent , and the matrices are now fixed by the constructions above. Corollary 3.3 also gives on every edge with . These are precisely the inputs to the next section.
Compression with a visible nonzero defect
We now use only the graph factorization and the precise common-sector output. The first lemma is independent of the group construction. Its proof explains why ordered edge annihilations, rather than symmetric orthogonality, suffice.
Lemma 6.1 (Triangular compression). Let be a nonzero unital algebra over a field , and let be positive integers. Suppose are idempotents and satisfy
Suppose there are and with , . Then explicit matrices
satisfy .
Proof. Define of shape and of shape by
Their product has entries . Therefore , where and is strictly lower triangular. In particular . Every off-diagonal entry begins with and ends with , so idempotence gives
This is all the commutation information needed. The finite sum
satisfies by cancellation of consecutive powers. Put ; then . For the sum is empty and the same identity holds.
Take block diagonals and of copies of and . Use ordered coordinates on and on . Their product is
Define matrices and by
The assumed rectangular identities give and . Thus
have the asserted shapes and satisfy
Only associativity and the stated identities have been used; the are not assumed to commute.
Apply the lemma to , whose identity is . Distinct nonedges have by Proposition 2.2; edges with have . The rectangular identities are exactly those of Proposition 5.1. Consequently
The strict, ordinary rational inequalities are
This is more than an abstract rank paradox: it gives a nonzero vector that will certify the reverse-product defect.
Lemma 6.2 (A square inverse and its certificate). Let be a unital ring, a central idempotent, and . Suppose and matrices , satisfy . Let include the first coordinates and retain them, with entries and zero. Then
satisfy and . More precisely, if has last coordinate and all others zero, then
Proof. Every entry of belongs to and is annihilated on both sides by . Since ,
The vector is nonzero because . The equation proves . But makes , so
Multiplying by and subtracting gives the final displayed identity and proves the strict reversal defect.
Using , in Lemma 6.2 gives square matrices over the entire algebra , not only the -summand. The nonzero certificate is rooted in the exact order of , which proves , and then in . It does not depend on identifying a coefficient by deciding equality of group words. The remaining step is to retain this defect when replacing matrices by scalar elements.
An injective finite-group encoding
We have a strict matrix one-sided inverse over . The finite-factor reduction of Dykema and Juschenko [9] explains the appropriate next step: put a full matrix algebra inside a finite group algebra and fill the complementary summand with the identity. We prove the actual matrix block needed here, including its injectivity. Its shift and diagonal operators are the finite Heisenberg model; the comparison with [14 Section 1.2] is methodological, since our cyclic order can be composite and our coefficient field is finite.
Proposition 7.1. Let be a field, let be invertible in , and let have order . For an integer put and . There is a finite group of order and a nonzero central idempotent for which explicitly specified elements , , form a basis of and obey
Thus by its standard matrix units.
Proof. Use the set with
All coordinates and the dot product use residues modulo . Associativity follows from the bilinear identity
The zero tuple is the identity, and the inverse is . No primality of is involved. The element is central of exact order . Define
Reindexing gives , and convolution gives . The identity coefficient is nonzero, so .
Write and , and put
The denominator is invertible because . Character orthogonality here means the direct identity
When , choose a nonzero coordinate . Translation of the sum by the th coordinate vector multiplies it by . In the field , this forces the sum to vanish. The argument still works when is a zero divisor modulo the composite integer .
Convolution in the abelian group of gives ; indeed the coefficient of in the product is
Summing (14) over also gives . The group law gives , and hence
Set
Then
and the diagonal sum is .
The multiplication identities alone do not yet prove that this block has no kernel. On define
The scalar in a composite is , so Equation (7.1) proves that is a representation. It sends to the identity and to the coordinate projection at . Therefore
which proves independence of all elements . On the other hand, the elements span , because . Their group-basis supports have distinct coordinates, so they too are independent. The sector has dimension exactly , and the matrix units form a basis. When , the empty tuple set has one element: , , and the sole matrix unit is . All formulas remain valid.
The elementary identity-completion step is worth recording separately. Its unit need not be central in the target ring.
Lemma 7.2. Let be an injective ring homomorphism between unital rings, not necessarily preserving their identities, and put . If and , then
satisfy and .
Proof. The element is idempotent and is an identity on the image of on both sides. Thus annihilates that image on both sides. Expanding the two products gives
Only the last inequality uses injectivity; no global centrality of is required.
Corollary 7.3. Under the field hypotheses of Proposition 7.1, suppose for a group and an integer , with and . Finite formulas give a finite group and with and . In particular, over any fixed finite field with at least three elements, a failure of stable finiteness transfers to a scalar failure after a finite direct factor, without changing the field.
Proof. Choose the least with . This search is finite, since . Pad the matrices by identities:
The padding is empty if . Then and . Use Proposition 7.1 and the group-basis identification . For an -square matrix over , put
The matrix-unit products prove multiplicativity, with . Independence of the gives injectivity: extend their coordinate dual functionals to and apply each in the second tensor factor to recover . This works even when has zero divisors.
Lemma 7.2 now yields the promised finite sums
Their full scalar product is one, and their reverse defect is . For a finite field of order , take and a generator of its multiplicative group. Then is invertible in the field, so the preceding construction applies.
Proof of Theorem 1.1. Lemma 2.1 supplies the specified odd prime and field of order . Propositions 2.2 and 3.1 give the graph fit and the strict regular-rank comparison. Proposition 4.2 embeds the local groups in , and Proposition 5.1 places their ranks in the nonzero ring with the required rectangular identities. Corollary 3.3 supplies every ordered edge product. Lemmas 6.1 and 6.2 produce , with their nonzero certificate. Corollary 7.3, using and , gives and the sums in (16). Their product is one and their reverse-product defect is nonzero by the injectivity of the matrix encoding.
The group is generated by the finite groups . The construction of adds only and the finitely many finite groups , and is finite. Thus is finitely generated. The embedded copy of any supplies nontrivial torsion. All field, group and coefficient choices were given by finite formulas and fixed orders, as required.
Finite prescriptions and scope. Every choice of character, basis, field element, or finite inverse occurs in an explicitly finite set. Products of the resulting group words define finite group-algebra sums even when different words happen to coincide. No word comparisons are needed to define the elements, and the kernel certificate and injective matrix block certify the nonzero reverse defect. The countable permutation actions used to certify group embeddings are not coefficient-search inputs.
Containing torsion does not say that every element has finite order. The group must be infinite, since finite-dimensional left multiplication proves direct finiteness for every finite group algebra, in modular characteristic as well as otherwise. The coefficient field throughout remains the specified of order .
More generally, the supports of any two finite sums in a group algebra generate a finitely generated subgroup. Its group algebra embeds by the group bases, preserving both a one-sided product and its nonzero reverse defect. Thus a universal direct-finiteness theorem for finitely generated groups would already imply the assertion for all groups.
The cellular automaton on all configurations
The final transfer is short and characteristic-independent. We give it locally to fix the multiplication order and the meaning of a missing configuration. Its group-ring ancestry appears in [10 Section 1] and the linear theory of [5].
For any field , group , and finite sum , define on all functions
Only the coefficient sum is finite; may have infinite support. Expanding the two finite sums gives
Thus and . If is the point indicator at , then , which also proves faithfulness.
Proposition 8.1. Let be a finite field, a group, and with . Then is an injective cellular automaton, with
For every with , the configuration is absent from its image.
Proof. The support of is a finite memory set with rule . Equation (8.1) gives , proving injectivity. If , then . Conversely, a fixed point of has , hence lies in the image. Finally,
so is not a fixed point and cannot be in the image.
Apply the proposition to the scalar sums in Equation (16). Their nonzero defect has a nonzero coefficient in the group basis, so the resulting automaton is injective and nonsurjective. The missing point indicator has no preimage even among infinitely supported configurations. If the finite formula for displays a group element more than once, its terms simply add in the same local rule; specifying the rule requires no word-problem oracle. Thus the constructed group also gives a negative answer to Gottschalk’s surjunctivity conjecture.
References
- [1]Noga Alon, László Babai, and Hiroshi Suzuki. Multilinear polynomials and Frankl–Ray-Chaudhuri–Wilson type intersection theorems. Journal of Combinatorial Theory. Series A, 58(2):165–180, 1991.DOI
- [2]Pere Ara, Kevin C. O’Meara, and Francesc Perera. Stable finiteness of group rings in arbitrary characteristic. Advances in Mathematics, 170(2):224–238, 2002.DOI
- [3]Alexander Barg, Moshe Schwartz, and Lev Yohananov. Storage codes on coset graphs with asymptotically unit rate. Combinatorica, 44(6):1193–1209, 2024.arxiv.org/abs/2212.12117
- [4]Alexander Barg and Gilles Zémor. High-rate storage codes on triangle-free graphs. IEEE Transactions on Information Theory, 68(12):7787–7797, 2022.DOI
- [5]Tullio Ceccherini-Silberstein and Michel Coornaert. Linear cellular automata: Garden of Eden Theorem, L-surjunctivity and group rings. Algebra and Discrete Mathematics, pages 22–35, 2006. no. 2.
- [6]Tullio Ceccherini-Silberstein, Michel Coornaert, and Xuan Kien Phung. First-order model theory and Kaplansky’s stable finiteness conjecture for surjunctive groups. Groups, Geometry, and Dynamics, 19:495–503, 2025.DOI
- [7]Tullio Ceccherini-Silberstein, Michel Coornaert, and Xuan Kien Phung. Stable finiteness of monoid algebras and surjunctivity. Theoretical Computer Science, 1042:115228, 2025.DOI
- [8]A. H. Clifford. Representations induced in an invariant subgroup. Proceedings of the National Academy of Sciences of the United States of America, 23(2):89–90, 1937.DOI
- [9]Ken Dykema and Kate Juschenko. On stable finiteness of group rings. Algebra and Discrete Mathematics, 19(1):44–47, 2015. https://admjournal.luguniv.edu.ua/index.php/adm/article/download/1174/663.admjournal.luguniv.edu.ua/index.php/adm/article/download/1174/663
- [10]Gábor Elek and Endre Szabó. Sofic groups and direct finiteness. Journal of Algebra, 280(2):426–434, 2004.DOI
- [11]Alexander Golovnev and Ishay Haviv. The (generalized) orthogonality dimension of (generalized) Kneser graphs: Bounds and applications. Theory of Computing, 18(22):1–22, 2022.arxiv.org/abs/2002.08580
- [12]Walter Gottschalk. Some general dynamical notions. In Anatole Beck, editor, Recent Advances in Topological Dynamics, volume 318 of Lecture Notes in Mathematics, pages 120–125. Springer, Berlin, Heidelberg, 1973.DOI
- [13]Mikhail Gromov. Endomorphisms of symbolic algebraic varieties. Journal of the European Mathematical Society, 1(2):109–197, 1999.DOI
- [14]Shamgar Gurevich and Ronny Hadani. On the diagonalization of the discrete Fourier transform. Applied and Computational Harmonic Analysis, 27:87–99, 2009.arxiv.org/abs/0808.3281
- [15]Ishay Haviv. On minrank and the Lovász theta function. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques (APPROX/RANDOM 2018), volume 116 of Leibniz International Proceedings in Informatics, pages 13:1–13:15. Schloss Dagstuhl – Leibniz-Zentrum für Informatik, 2018.arxiv.org/abs/1802.03920
- [16]Hexiang Huang and Qing Xiang. Construction of storage codes of rate approaching one on triangle-free graphs. Designs, Codes and Cryptography, 91:3901–3913, 2023.DOI
- [17]S. A. Jennings. The structure of the group ring of a p-group over a modular field. Transactions of the American Mathematical Society, 50(1):175–185, 1941.DOI
- [18]Irving Kaplansky. Projective modules. Annals of Mathematics, 68(2):372–377, 1958.DOI
- [19]Irving Kaplansky. Fields and Rings. Chicago Lectures in Mathematics. University of Chicago Press, Chicago, second edition, 1972.
- [20]OpenAI. A Counterexample to Kaplansky’s Direct-Finiteness Conjecture in Characteristic Two. OpenAI Math Release preprint OAI:A-Counterexample-to-Kaplanskys-Direct-Finiteness-Conjecture-in-Characteristic-Two-September-23-2026, 2026.github.com/openai/math/blob/main/preprints/A-Counterexample-to-Kaplanskys-Direct-Finiteness-Conjecture-in-Characteristic-Two-September-23-2026/paper.pdf
- [21]Xuan Kien Phung. A geometric generalization of Kaplansky’s direct finiteness conjecture. Proceedings of the American Mathematical Society, 151(7):2863–2871, 2023.arxiv.org/abs/2111.07930
- [22]Tomasz Prytuła. Graphical complexes of groups. Journal of Group Theory, 27(2):223–254, 2024.DOI
- [23]J. T. Schwartz. Fast probabilistic algorithms for verification of polynomial identities. Journal of the ACM, 27(4):701–717, 1980.DOI
- [24]Benjamin Weiss. Sofic groups and dynamical systems. Sankhyā: The Indian Journal of Statistics, Series A, 62(3):350–359, 2000.
- [25]Richard Zippel. Probabilistic algorithms for sparse polynomials. In Edward W. Ng, editor, Symbolic and Algebraic Computation, volume 72 of Lecture Notes in Computer Science, pages 216–226. Springer, Berlin, Heidelberg, 1979.DOI