Unavoidable Configurations for Torsion-Free Bredon Homology

When mathematicians compute the K-theory of the algebra attached to a symmetry group, they often use a spectral sequence: a staged bookkeeping device that assembles a hard global invariant from simpler homology groups. The computation becomes easy when that homology is "concentrated", meaning it sits in only one or two degrees and has no torsion (no finite cyclic pieces). In that case every later correction term vanishes and the answer can be read off directly. For compact three-dimensional hyperbolic reflection groups, Lafont, Ortiz, Rahm and Sánchez-García got exact integral formulas this way. They proved torsion-freeness by finding, by hand, coordinates in which the boundary matrices are totally unimodular, meaning every square minor is −1, 0 or 1. That proof covers an infinite family of groups, but each new configuration needs another manual search for good coordinates. The same search is what stops the method in higher dimensions.

This note proposes to replace the manual search with a minimal-counterexample argument of the kind used for the four-colour theorem. The idea rests on the fact that a compact Coxeter polyhedron in dimension three has a simple planar graph as its boundary combinatorics. Suppose some polyhedron in the family has torsion in its Bredon homology, which is homology with coefficients in the representation rings of the cell stabilizers. Take the smallest such polyhedron. We would show two things. First, it cannot contain any configuration from a finite list of local patterns, each certified "reducible": the pattern can be removed while provably preserving torsion-freeness. Second, Euler's formula forces every polyhedron to contain one of these patterns. The contradiction proves torsion-freeness for the whole family. It also turns a family-wide statement, which naively needs an unbounded check over infinitely many members, into a finite computation. That shift in logical complexity is the contribution that computability degree theory frames.

Core Claim

Torsion-freeness of low-degree Bredon homology across a combinatorially defined infinite family of reflection groups can be certified by three pieces together. The first is a finite unavoidable set of local configurations: small patterns of faces, edges and vertex stabilizers. The second is a reducibility certificate for each pattern. This is a totally unimodular splitting of the connecting map in a Mayer–Vietoris sequence, the exact sequence relating the chain complex of a polyhedron to that of a smaller one obtained by removing the pattern. The third is a minimal-counterexample argument that combines the first two.

Seen through computability theory, the statement "every member of a computably enumerated family has torsion-free $H_0$" is a $\Pi^0_1$ property. It can be refuted by exhibiting one bad member but never confirmed by finite search. A finite unavoidable reducible set makes it decidable, and in fact checkable by bounded local inspection with no unbounded search. That is the "punctual" (primitive recursive) regime studied by Kalociński, San Mauro and Wrocławski.

The note is also explicit about where the approach must fail. Their paper shows that functional structures such as trees with long branches resist punctual presentation. By analogy we expect families whose torsion comes from a global cycle with no local trace. For those families no finite unavoidable set exists, and identifying one is itself a result.

What is new is the transfer of reducibility-and-discharging technology from structural graph theory into the problem of making spectral sequences collapse, together with an honest classification of which families admit it. Here "discharging" is the counting argument, based on Euler's formula, that forces a small configuration to exist. The method is not a generic solver.

Where the Ingredients Come From

Several papers motivate the target by showing that degree concentration is the mechanism that makes these computations possible. Sánchez-García's computation for SL(3,Z) finds equivariant K-homology concentrated in degree zero. The Coxeter-group computation reduces right-angled and even Coxeter groups to a combinatorial sum over the Davis complex once relative Bredon homology concentrates. Lafont et al. show that the remaining obstruction in three dimensions is torsion in $H_0$. They kill it with their "vertex block" criterion, a total-unimodularity condition, and they name both the manual basis search and the growth in stabilizer complexity above dimension three as the barriers. Our method attacks exactly those barriers.

Paolini and Salvetti supply the second methodological template. They use precise matchings in discrete Morse theory, which pair cells so that the spectral sequence collapses at the $E^2$ page and twisted Artin-group homology reduces to ranks of incidence matrices. They do this family by family ($A_n$, $B_n$, $\tilde A_n$, $\tilde C_n$), and they leave $D_n$ and the exceptional types open. We reuse their matching idea as an alternative reducibility certificate. Their uniform family proofs are hand-crafted; we aim to produce such proofs from a finite verified catalogue.

The minimal-counterexample machinery comes from graph colouring. Lan, Shen, Wang and Zhou prove the Borodin–Kostochka bound for odd-hole-free graphs with $\Delta\ge 7$. They take a vertex-minimal counterexample, force it to be critical, and exclude local configurations by recolouring lemmas. Chudnovsky et al. do the same for edge colouring with reducible configurations. We borrow the logical shape but not the content. In our setting "recolouring" becomes "a unimodular change of basis on a Mayer–Vietoris connecting map," and "criticality" becomes "torsion is not inherited from any smaller polyhedron."

Kalociński et al. act as a constraint rather than a method. Their contrast is that relational digraphs are punctually presentable, while functional trees can force unbounded search. This warns us that locality may fail, and it gives us the vocabulary to state the negative half of the program precisely.

Research Program

The first stage builds the reduction step. We would take a compact Coxeter polyhedron $P$ and a pattern $C$ in it, such as a triangular, quadrilateral or pentagonal face with its adjacent dihedral angles and the induced vertex stabilizers ($D_m$, $S_4$, $A_5$ and the other spherical triangle groups). We would define a smaller polyhedron $P'$ by a combinatorial surgery that stays inside the class of Coxeter polyhedra, for example contracting a face or removing an edge cycle. We then write the Mayer–Vietoris sequence of Bredon chain complexes relating $P$, $P'$ and the patch, with connecting map $\partial_C$.

Call $C$ reducible if a change of basis makes the relevant block of $\partial_C$ totally unimodular, independently of the rest of $P$. The key lemma, which still has to be proved, is the following. If $C$ is reducible and $P'$ has torsion-free $H_0$ and $H_1$, then so does $P$. This holds because a totally unimodular block has Smith normal form invariant factors in $\{0,1\}$, so it introduces no new torsion. That is exactly the mechanism Lafont et al. use globally. A natural complication is that the surgery may leave the space of hyperbolic polyhedra, because Andreev's angle conditions can fail. Part of the work is to find surgeries that stay inside the class, or to enlarge the class to all Bredon complexes with the same local data. Any proof done in the enlarged class still applies to the original groups.

The second stage generates certificates. For each candidate pattern up to a size bound, we would search for a unimodular basis change with an integer program, encoding the requirement that all minors of the block lie in $\{-1,0,1\}$. Total unimodularity is decidable in polynomial time by Seymour's decomposition, so the integer program only has to search over basis changes, and every answer is then checked exactly. Every certificate is a small integer matrix that anyone can verify by hand or with a short script.

The third stage is discharging. Simple 3-polytopes satisfy Euler's formula, which forces faces with at most five sides. Hyperbolicity rules out triangular faces with large angle sums, and Andreev's theorem limits which angle patterns can occur. We would assign charges to faces and vertices and redistribute them by local rules. The aim is to show that every admissible polyhedron contains a pattern from the certified list. If the discharging closes, the theorem follows by contradiction: a minimal torsion polyhedron would contain a reducible pattern, so the smaller polyhedron $P'$ would also have torsion. Using the Lafont et al. formulas $K_0(C^*_r(\Gamma))\cong\mathbb Z^{cf(\Gamma)}$ and $K_1(C^*_r(\Gamma))\cong\mathbb Z^{cf(\Gamma)-\chi(\mathcal C)}$, where $cf(\Gamma)$ counts conjugacy classes of elements of finite order and $\chi(\mathcal C)$ is the Euler characteristic of the associated complex, re-deriving their theorem this way serves as a calibration.

The fourth stage tests whether locality survives in new families. The first is compact four-dimensional hyperbolic reflection groups, where the boundary combinatorics is no longer planar. Discharging there needs Euler-type relations for simple 4-polytopes, which are weaker, so success is uncertain. The second is the twisted homology of type-$D_n$ Artin groups. Here the "patterns" are local pieces of the Coxeter diagram, and the reducibility certificates are Paolini–Salvetti precise matchings instead of unimodular blocks.

The fifth stage covers the negative direction. If the search finds a family member with torsion, or a pattern that is irreducible yet unavoidable, we would try to show that torsion-freeness in that family depends on global features, for example via a family of polytopes whose torsion is detected only by a long cycle of faces. Then no finite local certificate exists, which is the analogue of the non-punctual functional trees.

Baselines

The do-nothing reference point is per-instance exact computation. We enumerate admissible Coxeter polyhedra up to a face-count bound $N$, build each Bredon chain complex, and compute integral homology by Smith normal form. This is exact for every instance it reaches, and it serves as our oracle: any member with torsion immediately refutes the family claim. What it cannot do is prove anything about members beyond $N$. Our method is compared with it on the same enumerated set, with the same machine and wall-clock budget. Our method counts as beating it if it certifies the full infinite family with a finite catalogue, or a family defined by a stated restriction, while agreeing with the oracle on every enumerated instance. No reported result was found on Lacuna for running time or coverage of such an enumeration.

The strongest existing method is the manual vertex-block criterion of Lafont et al.. It proves torsion-freeness for all compact three-dimensional hyperbolic reflection groups, through a hand-built global choice of basis. Its published result is a theorem, not a benchmark score, so there is no number to beat. The comparison is qualitative but checkable. In dimension three, matching it means automatically re-deriving the same integral formulas from a finite verified catalogue. Beating it means certifying a family it does not cover: a natural subclass of compact four-dimensional reflection polytopes, or groups with stabilizers beyond those in its character tables. For Artin groups the corresponding strongest baseline is Paolini and Salvetti, with closed formulas for $A_n$, $B_n$, $\tilde A_n$ and $\tilde C_n$. Beating it means a certified uniform statement for $D_n$, which it leaves open.

The ablation that isolates the key ingredient keeps the integer-programming certificate search but drops the minimal-counterexample structure. It looks for one global unimodular basis per instance, polyhedron by polyhedron. If this per-instance global search reaches the same instances at similar cost, the local surgery adds nothing computationally. In that case the contribution reduces to the family-level proof, and we would report it that way. A second ablation varies the allowed pattern size and records the smallest radius at which the discharging closes. A closing radius that keeps growing with $N$ is our empirical sign of non-locality.

Why not just…?

Why not just compute Smith normal form for each group? For any single group this is exact and fast, and we use it as the oracle. But the theorems at stake are about infinite families, and no finite enumeration settles a $\Pi^0_1$ statement. Lafont et al. needed a uniform argument for the same reason. The real threat is weaker than this question suggests: per-instance computation might find torsion early, which would make the family claim false. We would run the oracle first, up to the largest feasible $N$, before investing in discharging.

Why not just extend the vertex-block search by hand to dimension four? That is the route the literature already takes, and Lafont et al. report that growing stabilizer complexity is what blocks it. A manual global basis also has to be rebuilt whenever the combinatorics changes. A local catalogue is reusable, and a machine can check it. If the four-dimensional catalogue turns out to be huge, hand extension is no worse, and we would say so.

Why not just reduce to finitely many primes and check torsion modulo $p$? If torsion could provably occur only at primes dividing stabilizer orders (2, 3 and 5 for the three-dimensional stabilizers), comparing ranks modulo $p$ against rational ranks would detect torsion per instance cheaply. This is a real simplification of the oracle, and we would adopt it if the prime bound can be proved. It does not remove the need for a family-level argument, because it is still a check on one instance at a time. Combined with discharging, it could shrink each reducibility certificate to a check modulo $p$, which would be a welcome outcome.

Why not ask an automated prover or a language model to produce the family proof? The evidence is discouraging for constructive tasks in particular. On AMBER, pass@16 is 6.00% for DeepSeek-V3.2-Thinking and 4.00% for Kimina Prover-72B, and 47.3% of errors come from hallucinated library identifiers. Our certificates are small integer matrices and finite discharging rules, so they can be verified exactly without trusting a prover. Formalizing the final argument in Lean is a later step, not a replacement.

Proposed Experiments

The calibration experiment re-derives the three-dimensional theorem. We enumerate compact Coxeter polyhedra satisfying Andreev's conditions up to a face-count bound set by compute, and compute the oracle homology for all of them. We then run certificate search on patterns up to the smallest size at which discharging closes. Success means a finite catalogue, independently verified, that reproduces the Lafont et al. formulas with zero disagreements against the oracle. The reported outcomes are catalogue size, the closing radius, and total verification time.

The extension experiments target compact four-dimensional reflection polytopes and type-$D_n$ Artin groups, using the same oracle-then-certify protocol. The decisive outcomes are clear-cut. Either a finite certified catalogue closes, which yields a new uniform theorem. Or the oracle finds a member with torsion, which refutes the family claim and is worth publishing. Or the closing radius grows with $N$ with no plateau, which is evidence of non-locality. In the last case we would try to make it rigorous by building a family whose torsion is invisible to every bounded-radius pattern.

Because the graph-colouring benchmarks in the file (Layout, COLOR, the citation networks) measure a different task, they serve here only as a reminder that local methods fail in a provable way. Li et al. show that fixed-depth local message passing is not optimal on random regular trees. The fourth-stage analysis is the topological counterpart of that result: it asks where local certificates stop being enough.