{"id":"9e3be625-1cf6-46a5-a859-47a797084a4d","arxiv_id":"2606.03006","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Expected spectral measures of unimodular graph operators satisfy µ(I) = O(1/ln(1/|I|)) on intervals, extending Craig–Simon's log-Hölder regularity beyond Z^d to indicable groups, Anderson models, percolation, and quasi-transitive graphs.","lead":"This paper proves that the expected spectral measure of self-adjoint operators on symmetric infinite graphs obeys logarithmic Hölder regularity: the mass of every small interval shrinks at least like 1/log(1/|I|). It extends the classical 1983 Craig–Simon theorem beyond the Euclidean lattice to group algebras of indicable groups, Anderson-type random operators, percolation graphs, and Benjamini–Schramm limits.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 1's proof has an invalid linear-algebra step: F′ = F∩P (P = prodigy subspace) is shown trivial, but this does not imply F ⊆ B + Σ L_j; a 3-vertex path labelling gives F∩P = 0 yet an eigenvector has nonzero prodigy components. The main interval-regularity theorems inherit this gap.","rationale":"The paper is serious, well-structured, and the central goal — logarithmic Hölder regularity of expected spectral measures in the unimodular setting — is natural and potentially important. The (a,k)-indicable hypothesis is a genuine scope restriction, but the paper honestly flags its limits via Theorem 6 and Remark 2, so I do not treat that as the main problem. The most load-bearing concern is internal: the finite-dimensional counting lemma, Lemma 1, contains a demonstrable gap in its proof. The claim F ⊆ F′ + B + Σ L_j, which is used to reduce the dimension count to bad vertices and level-subgraph spectral subspaces, is false; the 3-vertex path example shows an eigenvector can have nonzero prodigy components while F∩P = 0. Since Theorem 4 explicitly 'essentially repeats' the proof of Lemma 1 in the von Neumann setting, and Theorem 5 is a block analogue, the interval regularity statements are not fully proven as written. The reader's verdict of CONDITIONAL is therefore appropriate and should not change: the concern warrants either a repaired proof of Lemma 1 or an independent verification of the lemma before the main theorems are accepted.","tokens_in":18985,"tokens_out":14677,"duration_ms":154099,"concrete_test":"Independently verify Lemma 1. First, check the disputed linear-algebra step on the 3-vertex path example above: it fails, so a replacement argument is needed that bounds dim F by |B| + Σ_j dim(F∩L_j) without the false containment. Second, exhaustively enumerate all weighted graphs on ≤4 vertices (loops allowed) and all labelings, compute L_G(I) and the right-hand side of Lemma 1 for intervals of length below the stated threshold. If any instance violates the inequality, the lemma is false; if none does, that supports the statement but still leaves the proof gap unresolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Lemma 1 (Section 3) rests on the assertion that 'F is contained in the sum of F′, B and the L_j′s', where F′ = F∩B^⊥∩∩_j L_j^⊥ = F∩P, with P = ⊕ P_j the prodigy subspace. This containment is not implied by the definitions and is false in general. Concretely, take the 3-vertex path 1–2–3 with labels η = 0,1,2 and unit edge weights. Vertex 1 is bad, vertices 2 and 3 are prodigy, and there are no level vertices. For the eigenvalue √2, the eigenvector f = (1,√2,1) lies in F, so F∩P = 0 (hence F′ = 0), yet f ∉ B. Thus F is not contained in B + Σ L_j. The subsequent induction only proves that no nonzero vector of F is supported entirely on prodigy vertices; it does not control the prodigy components of a general f ∈ F. Since Theorem 4 repeats the same step using von Neumann dimension, and Theorem 5 follows the same pattern, the interval estimates in Theorems 1–3 are not established as written. This is a proof gap, not a demonstrated contradiction — the lemma may be true — but the present derivation is incomplete.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a monotone labelling method to prove logarithmic Hölder regularity of expected spectral measures for self-adjoint local operators on unimodular random weighted graphs. The main technical results are Lemma 1 (finite graphs), Theorem 4 (unimodular graphs, vertex labelling), and Theorem 5 (block labelling). These are applied to elements of group algebras of indicable groups (Theorem 1), invariant random operators and Anderson-type models (Theorem 2), quasi-transitive operators (Theorem 3), and a construction for general indicable groups (Theorem 6). The claimed results would substantially extend Craig–Simon log-Hölder continuity and atomlessness of the density of states beyond Euclidean lattices, with explicit constants.","tokens_in":19194,"tokens_out":13337,"duration_ms":129240,"significance":"If the proofs were correct, the paper would be a significant advance in the spectral theory of operators on infinite graphs. The framework is elegant and unifies several models: group-algebra elements, invariant random operators, and Benjamini–Schramm limits. The constants α and β are explicit, and no parameter is fitted to data, which is a strength. The applications to Anderson-type potentials with arbitrary compactly supported distributions and to quasi-transitive graphs are natural and would be valuable. However, the central finite-dimensional lemma contains a false linear-algebra step, and this step is inherited by the infinite-dimensional theorems. Since the main results all rest on this lemma, the current manuscript does not establish its claims as written.","major_comments":[{"comment":"The assertion 'Since F is contained in the sum of F′, B and the L_j’s' is not justified and is false in general. Let W = B + Σ L_j. Then F′ = F ∩ W^⊥, and a general f ∈ F need not have its orthogonal projection onto W^⊥ again in F. Concretely, on the 3-vertex path 1–2–3 with labels η = 0,1,2 and unit edge weights, the eigenvalue √2 has eigenvector f = (1,√2,1). Then B = span(e_1), L_j = 0, P = span(e_2,e_3), and F′ = F ∩ B^⊥ ∩ ∩_j L_j^⊥ = 0, yet F is not contained in B. Thus the linear-algebra decomposition on which the entire proof rests is invalid. The subsequent induction only proves triviality of F∩P, i.e. of vectors in F supported entirely on prodigy vertices; it does not control the prodigy components of a general eigenvector. Since Theorem 4 and Theorem 5 repeat this step, the interval estimates in Theorems 1–3 and 6 are not established as written.","section":"Section 3, proof of Lemma 1"},{"comment":"In the infinite-dimensional proof, F is defined as 'the vector space spanned by eigenvectors of A_G with eigenvalues in I'. For a self-adjoint operator with continuous spectrum, the spectral subspace E(I)H is not spanned by eigenvectors. The equality dim(F) = E[µ_G^{δ_o}(I)], used immediately after the definition, does not follow for the continuous part of the spectral measure. The proof should take F to be the direct-integral spectral subspace E(I)H_ρ (or prove a limiting argument), not a space of eigenvectors. As written, Theorem 4 is not proved for the continuous part of µ_G^{δ_o}, which is essential to the claimed log-Hölder regularity.","section":"Theorem 4, proof"}],"minor_comments":[{"comment":"Typo: 'spectal measure' should be 'spectral measure'.","section":"Abstract"},{"comment":"The convention for a singleton interval I={λ} is stated, but the formula contains ln(β/|I|). For |I|=0 the expression is undefined; the intended limiting convention (right-hand side equals 0) should be stated explicitly.","section":"Section 1 / Theorem 1"},{"comment":"The remark that the constant 3 in α and κ is 'somewhat arbitrary' and that 'the proof gives a constant 2.09±0.01' is informal and not used; either remove it or make the purported better constant part of a theorem.","section":"Section 3, Lemma 1"},{"comment":"The name 'Grigorchuk and ˙Zuk' contains a typographical artifact; use 'Grigorchuk and Żuk'.","section":"Section 4.3 / Remark 2"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about Lemma 1 is valid and must be addressed before publication. The counterexample does not disprove the lemma, only the present proof step, so I am not recommending rejection. The central argument needs a substantive revision: either a correct decomposition or a different proof of the finite-dimensional eigenvalue-counting estimate, and the infinite-dimensional proof must handle the spectral subspace rather than the eigenvector span. If these can be repaired, the paper would likely be a strong contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a real paper by a serious author, and it deserves a careful referee, but the main theorems are not established as written. The key problem is in the proof of Lemma 1. The author claims that the eigenspace F is contained in the sum F′ + B + ΣL_j, where F′ = F ∩ B^⊥ ∩ ∩_j L_j^⊥. That containment is not valid. A concrete 3-vertex path with labels 0,1,2 gives B = {1}, L_j empty, F = span((1,√2,1)); then F′ = 0 and F is not contained in B. The induction that follows only proves F′ is trivial, not that dim F is bounded by |B| + Σ dim L_j. The same step is reused in Theorem 4 via von Neumann dimension and in Theorem 5, so the interval regularity bounds in Theorems 1–3 inherit the gap. This is a proof gap, not a demonstrated counterexample to the statements — the lemma may well be true — but the current derivation is incomplete.\n\nWhat the paper does well: it genuinely extends the monotone labelling method from singletons to intervals, from amenable groups and k=1 to arbitrary unimodular settings, and to block labellings with matrix-valued coefficients. The applications are substantial: Anderson-type models with arbitrary compactly supported potentials on Cayley graphs, anisotropic percolation, and quasi-transitive operators. Theorem 6 is a thoughtful fallback for indicable groups where the main geometric hypothesis fails, and Remark 2 honestly discusses the Grigorchuk–Żuk example. The paper is also careful about its own limitations, including the arbitrary constant 3.\n\nThe other soft spot is Theorem 4: its proof is delegated to [8] as an \"easy extension,\" but given the gap in finite dimension, that delegation is not acceptable without a written argument. A referee should ask for a corrected proof of Lemma 1 — likely by replacing the invalid containment with a projection argument that controls F ∩ (B + ΣL_j) directly — and then a self-contained proof of Theorem 4.\n\nWho is it for: researchers in spectral theory of random operators and L² invariants. The paper deserves peer review, not desk rejection, because the claims are important and the gap looks fixable. But I would not cite the main theorems until the proof is repaired.","headline":"Serious, ambitious extension of Craig–Simon to unimodular graphs, but the proof of the central Lemma 1 has a real gap that propagates to Theorems 4–5.","tokens_in":19854,"tokens_out":3334,"would_cite":false,"duration_ms":33920,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["47B80","46L10","05C80"],"pacs":[],"model":"deepseek-v4-flash","headline":"For a broad class of infinite graphs, the expected spectral measure satisfies µ(I) ≤ C / ln(1/|I|) on every interval I.","keywords":["spectral measure","logarithmic Hölder regularity","monotone labelling","unimodular random graph","group algebra","indicable group","Anderson model","Craig–Simon theorem"],"falsifier":"Numerically diagonalize large finite quotients of a surface-group Cayley graph with the Anderson model (or a group-algebra element) and measure the integrated density of states on intervals of length ε; if for some ε the mass of a length-ε interval decays slower than 1/ln(1/ε) as ε→0, the claimed logarithmic regularity is false. More directly, any explicit (a,k)-indicable group with a self-adjoint p having non-vanishing p_a and a non-zero atom in µ_p would falsify Theorem 1.","tokens_in":18681,"feed_emoji":"📉","tokens_out":10863,"duration_ms":100537,"temperature":0.7,"pith_summary":"This paper proves that the expected spectral measure of many self-adjoint operators on infinite graphs is logarithmically Hölder regular: the mass of any short interval I is at most a constant over ln(1/|I|), so in particular the measure has no atoms. The operators covered include group-algebra elements on Cayley graphs, Anderson-type tight-binding models, anisotropic percolation operators, and operators on quasi-transitive graphs. The engine is a strengthened monotone labelling method: an invariant labelling splits vertices into 'prodigy', 'level', and 'bad' classes, and an invertible transfer operator propagates eigenfunction estimates from level to level. The main hypothesis, the (a,k)-indicable condition, is that a homomorphism to Z sends one generator strictly above all others; when it holds, the paper improves earlier no-atoms theorems to a quantitative bound. This extends the classical Craig–Simon theorem from Z^d to many non-amenable groups and random settings.","feed_headline":"Spectral measures get 1/ln(1/|I|) bound on indicable Cayley graphs","feed_subtitle":"Craig–Simon's log-Hölder regularity now holds for Anderson models and group-algebra operators on a wide class of infinite graphs.","key_machinery":"Monotone labelling with transfer operator. Given an invariant labelling η:V→{0,...,n-1}, a vertex is a prodigy if it has a neighbour x̂ with η(x̂)<η(x) and all other neighbours of x̂ have label <η(x); level vertices have all neighbours of label ≤η(x); the rest are bad. The proof defines T_j from prodigy vertices of label j to their parents x̂; the key requirement is that T_j be invertible with inverse norm at most 1/p*. The recursive eigenfunction estimate (Lemma 1) propagates the size of the spectral projection from level to level, yielding the interval bound. In the infinite setting, the finite-dimensional spectral count is replaced by the von Neumann dimension of the associated tracial al","core_discovery":"The central result is a quantitative monotone-labelling estimate: in any unimodular weighted rooted graph with an invariant labelling and deterministic bounds ∥A_G∥≤K and inf|p(x,x̂)|≥p*>0, the expected spectral measure at the root obeys E[µ_G^{δ_o}(I)] ≤ P(o∈B) + Σ_j E[µ_{G_j}^{δ_o}(I)1_{o∈L_j}] for intervals of length at most κ α^{-n}. Specializing to Cayley graphs of (a,k)-indicable groups, where the invariant labelling comes from a homomorphism φ:Γ→Z and ω mod n, this gives the explicit regularity bound µ_p(I) ≤ 2k ln(α)/(ln(β/|I|))_+ for every interval I, with α=3∥p∥_op/|p_a|, β=2√2|p_a|, and consequently no atoms. The same inequality is proved for right-invariant random arrays (Theorem","pith_inferences":["The bound depends on only the coefficient p_a and the global norm, suggesting that the spectral regularity mechanism is essentially one-dimensional: the monotone labelling along a single homomorphism is the only structure doing the work. One could try to exploit several independent homomorphisms to obtain stronger regularity (e.g., power-law) for groups with higher rank structure.","The phenomenon seen in the lamplighter group, where breaking the one-sidedness of the generator produces purely atomic spectra, suggests that the (a,k)-indicable condition might be close to a necessary condition for any universal logarithmic bound on Cayley graphs of indicable groups. Testing this necessity on other groups with multiple generators of equal φ-value would sharpen the picture.","The method may extend to operators on periodic manifolds with a co-compact group action, but the finite-trace von Neumann algebra is absent; a limiting argument using finite approximants (along the lines of Benjamini–Schramm convergence) might still yield a logarithmic regularity statement for the integrated density of states."],"forward_implications":["For every self-adjoint element of the group algebra of an (a,1)-indicable group (free groups, even Artin groups, surface groups), the spectral measure has no atoms and satisfies the explicit log bound.","The density of states of Anderson-type models on Cayley graphs of indicable groups has logarithmic Hölder regularity for arbitrary compactly supported potential distributions, without independence or density assumptions.","For anisotropic percolation models in which the distinguished generator's edge is always present, the expected spectral measure is uniformly log-regular regardless of the other bond probabilities.","The block-labelling extension applies to quasi-transitive operators, yielding iterative bounds and, in examples, finitely many atoms contained in the spectrum of a finite matrix."],"fun_headline_variants":["Craig–Simon log-Hölder bound extends to Cayley graphs","Log-Hölder regularity for spectral measures on infinite graphs","No atoms: spectral measure regularity on indicable groups","Spectral measures get log-Hölder bound on Cayley graphs","Beyond Euclidean: Craig–Simon theorem broadened to infinite graphs"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The argument requires the graph's group to admit a homomorphism to Z under which some generator has a strictly larger image than all others (the (a,k)-indicable condition), and requires the coefficient of that generator in the operator to be bounded away from zero; if either fails, the main bound can be vacuous or false.","fun_headline_variants_meta":{"raw":{"variants":["Craig–Simon log-Hölder bound extends to Cayley graphs","Log-Hölder regularity for spectral measures on infinite graphs","No atoms: spectral measure regularity on indicable groups","Spectral measures get log-Hölder bound on Cayley graphs","Beyond Euclidean: Craig–Simon theorem broadened to infinite graphs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000208,"raw_usage":{"total_tokens":1254,"prompt_tokens":768,"completion_tokens":486,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":410}},"tokens_in":512,"tokens_out":486,"duration_ms":4986,"temperature":1.0,"reasoning_tokens":410,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T12:32:31.791267+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Numerically diagonalize large finite quotients of a surface-group Cayley graph with the Anderson model (or a group-algebra element) and measure the integrated density of states on intervals of length ε; if for some ε the mass of a length-ε interval decays slower than 1/ln(1/ε) as ε→0, the claimed logarithmic regularity is false. More directly, any explicit (a,k)-indicable group with a self-adjoint p having non-vanishing p_a and a non-zero atom in µ_p would falsify Theorem 1.","supporting_citations":[],"review_version":2}