{"id":"2fdb9600-3784-4de4-953b-cb0cfaf1022e","arxiv_id":"2609.28397","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For any IID query distribution and any initial list order, the transposition rule achieves expected query cost at most OPT plus 1250 after only n^29 queries.","lead":"This paper proves that the transposition rule for self-organizing lists reaches within an additive constant of optimal expected cost after a polynomial number of queries, resolving a conjecture of Coester. It matters because it gives a finite-time guarantee that holds for every IID query distribution and every starting list order, without learning the distribution.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Constant slip in Lemma 4.1 is real but absorbed by slack; mixing-time concern does not land after checking the arithmetic.","rationale":"The reader's conditional verdict is appropriate, but for a different reason than their weakest_assumption. The arithmetic in Lemma 3.2 is actually fine: the quoted Levin–Peres bound does yield t_mix(n^{-2}) ≤ n^4 for the relevant parameter range θ ≥ 1/(2n) and n ≥ 1024, so the reader's specific worry about the n^4 bound does not land. Instead, the more substantive issue is the false constant in Lemma 4.1's Cauchy–Schwarz step, which inflates the n^5 coefficient by about 1.77 but is absorbed by the slack in C = 1250. This is a genuine correctness defect in the exposition, though not in the theorem's conclusion. The proof is otherwise coherent: the adaptive buffer, coupling to the biased exclusion chain, and entropy-decay argument all check out. The final Markov-property step is unnecessary and slightly off, but (4.4) already gives the bound for all larger t. Therefore the paper should be conditionally accepted pending a corrected constant in Lemma 4.1.","tokens_in":11423,"tokens_out":25653,"duration_ms":205829,"concrete_test":"Re-derive Lemma 4.1's constant by evaluating the Cauchy–Schwarz step with the correct inequality ν((1−K)J) ≤ (1/2)√2 n^2 sqrt(EK) and then simulate the final bound with the corrected coefficient (1+√e) in place of 3/2, for n = 1024 and t = n^29, to confirm the total stays below 1250. Independently, plug θ = 1/(2n), N = n into the Levin–Peres expression 9N/θ^2 (4 log n + θN/2) for n = 1024 and verify it is less than n^4, thereby checking the claimed mixing bound.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most concrete defect is in the Cauchy–Schwarz step of Lemma 4.1: the displayed inequality claiming the second factor is bounded by 3/2 n^5 sqrt(4 EP) is false. Re-deriving, the correct upper bound is 2(1+√e) n^5 sqrt(EP) ≈ 5.436 n^5, and after using 4 EP(√f,√f) ≤ EP(f,log f) the Lemma 4.1 constant becomes 1+√e ≈ 2.65 instead of 3/2. This inflates the n^5 coefficient in (4.4) by about 1.77, but the slack in C=1250 absorbs it: 3C0 ≈ 1197, the entropy-decay term is ≈ 2.7, and the coupling error e^{−n} is negligible, so the final conclusion still holds. The proof as written, however, contains a false intermediate inequality and must be corrected.\n\nThe reader's main worry about Lemma 3.2 does not appear to hold. For the worst case θ = 1/(2n), N = n, the quoted Levin–Peres bound gives 9N/θ^2 (4 log n + θN/2) ≈ 144 n^3 log n, which is < n^3 · n = n^4 for n ≥ 1024 (e.g., 144·6.93 ≈ 998 < 1024). For larger θ the product decreases because 1/θ^2 shrinks faster than the bracket grows, so the n^4 mixing bound is consistent with the stated formula. Thus the dependence on the external theorem is a reproducibility concern, but not an internal inconsistency.\n\nThe final step extending t = T to all t ≥ T via the Markov property is unnecessary and slightly misstated, since the bound (4.4) applies directly to every t ≥ T with an even smaller entropy term. No fatal mathematical gap is apparent.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper proves a uniform polynomial-time additive-optimality guarantee for the transposition self-organizing list under IID queries. Theorem 1.2 states that for every n≥1, every probability vector p, every initial permutation σ0, and every integer t≥n^29, the expected access cost E[Cost(σ_t)] ≤ OPT + 1250. The proof decomposes excess cost as F(σ)=∫ e^x I_x(σ) dx (identity (1.1)), controls the integral with an adaptive buffer θ(x) (Lemma 2.1, Lemma 2.2), compares surprising inversions to a biased exclusion chain (Prop 3.4 via Levin–Peres mixing and canonical paths), converts cost to a Dirichlet form (Lemma 4.1), controls the distance from the initial law through a discrete-time entropy-decay bound (Lemma 4.2), and handles zero/small probabilities by an exponentially small perturbation with coupling (Lemma 4.3). The final arithmetic, 3C0 ≈ 1197.3 plus an O(1) entropy term and an exponentially small coupling term, is below 1250 with a healthy margin.","tokens_in":13652,"tokens_out":2007,"duration_ms":216845,"significance":"If correct, the result resolves Coester's conjecture and is a genuine advance: it gives a worst-case, distribution-independent, arbitrary-start polynomial-time OPT+O(1) guarantee for transposition even though the underlying chain can mix arbitrarily slowly. The paper's structure is a strength: the adaptive-buffer decomposition is clean and hand-checkable, the comparison with a biased exclusion process plus canonical paths is explicit, and the entropy-decay step gives a transparent finite-time error in the spirit of the locally stationary framework of Liu et al. [7]. The final constants and exponents are fully displayed, so the claim is checkable in principle; the paper is also commendably explicit about its AI-assisted preparation. The main mathematical substance — the cost-to-Dirichlet conversion and the discrete-time entropy decay — is novel in this application and is not borrowed from the cited literature.","major_comments":[{"comment":"The final displayed inequality of the proof is false. Re-deriving the line \"(1+√e)/2 · √2 n² · √(2n⁶ E_P(√f,√f)) ∫ e^{x/2} dx ≤ (3/2)n⁵√(4E_P(√f,√f))\": the left side equals 2(1+√e) n⁵ √(E_P(√f,√f)) ≈ 5.436 n⁵ √(E_P(√f,√f)), which is strictly larger than 3 n⁵ √(E_P(√f,√f)). The correct bound after the reduction 4E_P(√f,√f) ≤ E_P(f,log f) is (1+√e) n⁵ √(E_P(f,log f)) ≈ 2.65 n⁵, not (3/2)n⁵. This inflates the n⁵ coefficient inherited by (4.4) by a factor of about 1.77. I verified the numerical conclusion survives: 3C0 ≈ 1197.3, the entropy term in (4.4) is O(1) (on the order of 2.7 before and roughly 5 after the correction), and the coupling term (T+1)n²e^{−n} is exponentially small, so the total stays below 1250. The theorem is therefore unaffected, but Lemma 4.1's stated bound is wrong and must be corrected, with the corrected constant propagated through (4.4) and through the final display of the proof of Theorem 1.2.","section":"§4.1, Lemma 4.1"},{"comment":"After fixing Lemma 4.1, the coefficient 9/2 in (4.4) must consistently become 3(1+√e) ≈ 7.95, and the displayed bound \"3C0 + 3 < C\" should be replaced by a display with the actual corrected constants. Since 3C0 ≈ 1197.3 and the corrected entropy term computed at t = T = n^29 with H_q ≤ n³/2 evaluates to an O(1) quantity well under 50, the slack in C = 1250 comfortably absorbs the correction; however, as printed the chain of inequalities (3/2 → 2.65) is not numerically transparent, and the paper should show the single-line evaluation 3C0 ≈ 1197.3 < 1250 − (corrected error terms) explicitly.","section":"§4.2, (4.4) and proof of Theorem 1.2"}],"minor_comments":[{"comment":"The text refers to \"Theorem 3.1\", \"Theorem 3.2\", \"Theorem 3.4\", \"Theorem 2.1\", \"Theorem 2.2\", \"Theorem 4.1\", and \"Theorem 4.2\" where the corresponding objects are Definitions 3.1/3.3, Lemmas 2.1/2.2/3.2/4.1/4.2, and Proposition 3.4. The numbering conventions should be unified so that a reader can retrieve the exact statement invoked at each displayed step of Lemma 4.1.","section":"Throughout, §3 and §4"},{"comment":"The extension from t = T to all t ≥ T via the Markov property is unnecessary and, as written, slightly imprecise: one cannot conclude sup_σ (P^T g)(σ) ≤ sup_σ (P^T F)(σ) merely from g ≤ F. Since bound (4.4) applies directly to every t ≥ T with an even smaller entropy term, the Markov-property sentence should be deleted or replaced by the direct claim that (4.4) holds for all t ≥ T.","section":"Proof of Theorem 1.2, final step"},{"comment":"The decisive quantitative scale t_mix(n^{−2}) ≤ 9N/θ²(4 log n + θN/2) ≤ n^4 is imported verbatim from [6, Theorem 10] and not derived. My arithmetic check confirms the displayed expression is consistent with ≤ n^4 under the stated hypotheses: for θ = 1/(2n), N = n the bound is ≈ 144 n³ log n, which is below n^4 for n ≥ 1024, and the quantity decreases in θ beyond that point. This is therefore not an internal inconsistency, but the paper is entirely opaque about the prefactor's provenance; the authors should quote the exact Levin–Peres theorem (including its scope conditions) or include a short derivation, since this estimate is the only source of the polynomial time-scale n^29.","section":"Lemma 3.2 / Proposition 3.4"},{"comment":"The line \"use F_p ≤ n < C\" should state explicitly that OPT ≥ 1 (every query costs at least 1), so F_p ≤ n − 1; the inequality as written is harmless but slightly imprecise for the reader checking tightness.","section":"Proof of Theorem 1.2, n < 1024 case"},{"comment":"The definition of C0 in (4.1) and the final claim \"C = 1250 suffices\" are never numerically evaluated. Since the whole proof's credibility rests on the gap between 3C0 and 1250, the authors should include the one-line evaluation 3C0 ≈ 1197.3 (or an upper bound) in the proof of Theorem 1.2.","section":"§4.1, definition of C0"}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Clay,\n\nBottom line: this is the first polynomial-time, distribution-uniform OPT+O(1) bound for the transposition rule under IID accesses, and it confirms Coester's conjecture. The machinery is new: an adaptive buffer around each log-probability scale, a coupling to a biased exclusion chain, and a discrete-time entropy-decay lemma in the spirit of Liu et al. The inversion decomposition of the excess cost, the buffer integrability (Lemma 2.1), the entropy decay (Lemma 4.2), and the closing arithmetic all check out.\n\nTwo soft spots, in proportion. First, the reader is right about Lemma 4.1: as displayed it is false. The Cauchy–Schwarz step ends at (3/2)n^5·sqrt(4E_P), but the preceding lines give 2(1+sqrt(e))·n^5·sqrt(E_P) — about 1.81x larger. After 4E_P(sqrt f,sqrt f) ≤ E_P(f,log f), the correct Lemma 4.1 constant is (1+sqrt(e)) ≈ 2.65, not 3/2. I traced the corrected version through (4.4): 3C0 ≈ 1197.3, entropy term ≈ 2.3, coupling error n^31 e^{-n} negligible, total ≈ 1199.6 < 1250. So the theorem survives with roughly 50 units of slack, but the displayed inequality is wrong and must be corrected.\n\nSecond, the reader's worry about Lemma 3.2 does not land. Worst case theta = 1/(2n), N = n: the cited Levin–Peres bound gives 36n^3(4 log n + 1/4) ≈ 144 n^3 log n, which is below n^4 for n ≥ 1024 (144·6.93 ≈ 998 < 1024), and larger theta only shrinks the product. So the dependence on the external bound is a reproducibility concern, not an internal inconsistency; a referee should ask the author to spell out that verification.\n\nMinor: the Markov-property extension from t = T to all t ≥ T is unnecessary, since (4.4) applies directly at every t ≥ T. Harmless. The AI-usage disclosure is honest and I take it without prejudice, but it does mean independent human checking of constants matters more than usual — and it paid off here, since the Lemma 4.1 slip is exactly what a human eye was needed for.\n\nClear thinking, honest engagement, no self-citation games. The paper deserves a serious referee: conditional acceptance, requiring the Lemma 4.1 fix and requesting the mixing verification. I'd bring it to reading group and cite it once the correction is in.","headline":"Real result — resolves Coester's conjecture with genuinely new machinery — carrying one genuine constant slip in Lemma 4.1 that the slack absorbs; the flagged mixing concern does not survive the arithmetic.","tokens_in":12456,"tokens_out":12312,"duration_ms":61195},"pith_extraction":{"msc":["60J10","68P10","68W40"],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"After polynomially many requests, the transposition rule guarantees expected query cost at most OPT + 1250, uniformly over all request distributions and all starting list orders.","keywords":["list update problem","transposition rule","self-organizing list","IID queries","excess cost","adaptive buffer","biased exclusion process","entropy decay"],"falsifier":"Simulate the biased exclusion process with $N = n$ items and bias $\\tanh(\\theta/2)$ at $\\theta = 1/(2n)$ from a worst-case starting word and measure total-variation distance to stationarity at time $n^4$: if it exceeds $n^{-2}$ for some $n \\geq 1024$, the quoted estimate that fixes the exponent fails. A complementary check runs the transposition rule itself on the difficult distribution $p_i \\propto e^{-i/\\sqrt{n}}$ from the worst order and measures expected excess cost $F(\\sigma_t)$ as a function of $t$; the theorem predicts a polynomial collapse below 1250 well before $n^{29}$, so observing a residual excess above 1250 (or a collapse requiring an exponent much larger than 29) would force the quantitative claim to be rescaled.","tokens_in":13026,"feed_emoji":"🔁","tokens_out":20688,"duration_ms":162085,"temperature":0.7,"pith_summary":"In the list update problem, the cost of a query is the position of the accessed item, and under independent identically distributed requests the best static order is by decreasing request probability, with optimal expected cost OPT. The transposition rule, which moves each queried item one position forward, was known to be within OPT + 1 of optimal once the process reaches its stationary distribution, but that distribution can take arbitrarily long to appear. This paper proves that the waiting is unnecessary: after $t \\geq n^{29}$ queries, from any starting order and under any probability distribution, the expected query cost is at most OPT + 1250. The guarantee is uniform — the constant 1250 does not depend on the distribution, the initial list, or the time (beyond the stated threshold). The result confirms the conjecture attached to the earlier stationarity bound and removes the need for the underlying list chain to have mixed.","feed_headline":"Transposition rule: OPT + O(1) after polynomially many queries","feed_subtitle":"Expected query cost stays within OPT plus 1250 for any request odds or starting order, even if the chain barely mixes.","key_machinery":"The motor of the proof is the threshold decomposition $F(\\sigma) = \\int_{-\\infty}^{0} e^{x} I_x(\\sigma)\\,dx$, where $y_i = \\log p_i$ and $I_x(\\sigma)$ counts inverted pairs $(i,j)$ with $\\sigma(i) > \\sigma(j)$ and $y_j \\leq x < y_i$. Around each threshold the paper places an adaptive buffer $\\theta(x)$, defined by $\\sum_i (2\\theta(x) - |x - y_i|)_+ = 1$ and capped at $1/2$; this ensures $\\theta(x) \\geq 1/(2n)$, makes $\\theta$ Lipschitz, and forces the integral $\\int_{-\\infty}^{0} e^x / \\theta(x)^2\\,dx$ to be an absolute constant. Inversions split into unsurprising pairs inside the buffer window, whose number is deterministically at most $\\theta^{-2}$, and surprising pairs separated by a probability ratio at least $e^{\\theta}$. Surprising inversions are controlled by an auxiliary chain that freezes the intermediate-probability items and evolves the remaining low/high word as a biased exclusion process with bias $\\tanh(\\theta/2)$; a canonical-paths comparison bounds the auxiliary chain's Dirichlet form by $2n^2 e^{-x}$ times the original chain's, and a discrete-time entropy-decay lemma bounds that Dirichlet form at time $t$ by roughly $\\sqrt{H/t}$ with $H \\leq n^3/2$, where $H$ is the log of the reciprocal of the minimum stationary weight. A final perturbation of $p$ by $e^{-n}$ handles zero probabilities, and the quoted polynomial mixing bound for the exclusion process is what fixes the $n^{29}$ exponent.","core_discovery":"The central claim is Theorem 1.2: for every $n \\geq 1$, every probability vector $p$ on the $n$ items, every initial permutation $\\sigma_0$, and every integer $t \\geq n^{29}$, the expected cost $\\mathbb{E}[\\mathrm{Cost}(\\sigma_t)]$ is at most $\\mathsf{OPT} + 1250$, where $\\sigma_t$ is the list after $t$ steps of the transposition rule. The proof writes the excess cost $F(\\sigma) = \\mathrm{Cost}(\\sigma) - \\mathsf{OPT}$ as a sum of probability-weighted inversions, $\\sum_{i<j}(p_i - p_j)\\mathbf{1}\\{\\sigma(i) > \\sigma(j)\\}$, and then as an integral over logarithmic thresholds $x \\in (-\\infty, 0]$ of the count $I_x(\\sigma)$ of inverted pairs whose log-probabilities straddle $x$. Inversions whose probabilities differ by less than a factor $e^{\\theta(x)}$ are few and cheap; inversions whose probabilities differ more are handled by comparison with a biased exclusion process that provably sorts in polynomial time. The remaining error from the chain not being at stationarity is bounded by a discrete-time entropy-decay inequality that shrinks like a power of $1/t$, so the cost guarantee does not wait for mixing. Read in the most direct way, the result says the slow-mixing regimes of the transposition chain are exactly the regimes where the unmixed tail carries negligible expected cost.","pith_inferences":["The paper declares the exponent 29 and the constant 1250 deliberately unoptimized; counting the polynomial factors in the proof — the $n^4$ exclusion mixing time, an $n^2$ comparison factor, and an $n^5$ Dirichlet-form factor against a $t^{-1/4}$ decay — suggests the real cost-stabilization threshold may be orders of magnitude smaller, and that OPT + O(1) may hold at $t = n^{O(1)}$ with exponent f","Because the adaptive buffer simply discards the slow-mixing tail, the argument suggests a transferable principle: for any data structure whose cost objective is concentrated on the high-probability items, polynomial mixing of the heavy subset might suffice for near-optimal cost, even if the full state space mixes arbitrarily slowly.","The same threshold-integral machinery could plausibly be aimed at other update rules, such as move-to-front, or at non-IID request streams, since nothing in the decomposition uses the one-position-forward update except the stationary measure and the comparison paths; this is a natural test of the method's scope, not something the paper claims.","A concrete experiment worth running is on the distribution $p_i \\propto e^{-i/\\sqrt{n}}$ that the development singled out as difficult: measuring the time at which expected excess cost first drops below a fixed constant would calibrate how much slack the $n^{29}$ bound carries, and the proof's structure predicts the observed threshold would be far smaller."],"forward_implications":["For any request distribution and any starting order, the transposition rule is within an additive constant 1250 of optimal expected query cost after at most $n^{29}$ requests, with no a priori estimate of the distribution needed.","Slow mixing of the full chain is no longer an obstacle to the cost guarantee: the distributions on which the chain mixes arbitrarily slowly are precisely handled because their unmixed tails are cheap.","The guarantee is time-uniform: by the Markov property the same bound holds at every $t \\geq n^{29}$, not just at one checkpoint.","The result upgrades the earlier 'eventually at most OPT + 1' stationarity theorem to 'at most OPT + 1250 within polynomial time', confirming the conjecture attached to that theorem.","All constants in the proof are explicit, so the bound OPT + 1250 at $t \\geq n^{29}$ is a fully quantitative claim about the transposition rule, not an asymptotic existence statement."],"supporting_citations":[{"why":"Provides the stationarity result (expected cost at most OPT + 1) that this paper extends, and the conjecture the paper confirms.","marker":"[2]"},{"why":"Provides the canonical-paths comparison theorems used to transfer the auxiliary chain's Dirichlet form bound to the original transposition chain.","marker":"[3]"},{"why":"Supplies the quoted polynomial mixing-time bound for the biased exclusion process that Lemma 3.2 depends on; this quantitative estimate fixes the exponent in the main theorem.","marker":"[6]"},{"why":"Gives the local-stationarity and entropy framework whose discrete-time analogue (Lemma 4.2) controls the cost defect before the chain mixes.","marker":"[7]"},{"why":"Establishes the stationary distribution of the transposition rule, which the paper uses as its reference measure throughout.","marker":"[10]"}],"fun_headline_variants":["Transposition hits OPT+O(1) in polynomial time","Polynomial-time near-optimal list update via transposition","Transposition rule: bounded cost without fast mixing","Slow mixing no barrier: transposition achieves OPT+1250 fast","List update: transposition converges in cost, not distribution"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The timing of the whole argument rests on a bound, quoted from a cited theorem and not re-derived in this paper, that a biased swapping process on $N \\leq n$ items with bias $\\tanh(\\theta/2)$ reaches within $n^{-2}$ of stationarity in at most $n^4$ steps from any starting arrangement; if the true worst-case settling time were a higher power of $n$, the $n^{29}$ threshold would have to be rescaled.","fun_headline_variants_meta":{"raw":{"variants":["Transposition hits OPT+O(1) in polynomial time","Polynomial-time near-optimal list update via transposition","Transposition rule: bounded cost without fast mixing","Slow mixing no barrier: transposition achieves OPT+1250 fast","List update: transposition converges in cost, not distribution"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000166,"raw_usage":{"cost":0.000166068,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000166068,"upstream_inference_prompt_cost":0.000124348,"upstream_inference_completions_cost":4.172e-05},"total_tokens":1242,"prompt_tokens":1093,"completion_tokens":149,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":63}},"tokens_in":1093,"tokens_out":149,"cost_source":"openrouter_usage","duration_ms":2462,"image_pages":[],"temperature":1.0,"reasoning_tokens":63,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-09-24T02:35:08.337347+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Simulate the biased exclusion process with $N = n$ items and bias $\\tanh(\\theta/2)$ at $\\theta = 1/(2n)$ from a worst-case starting word and measure total-variation distance to stationarity at time $n^4$: if it exceeds $n^{-2}$ for some $n \\geq 1024$, the quoted estimate that fixes the exponent fails. A complementary check runs the transposition rule itself on the difficult distribution $p_i \\propto e^{-i/\\sqrt{n}}$ from the worst order and measures expected excess cost $F(\\sigma_t)$ as a function of $t$; the theorem predicts a polynomial collapse below 1250 well before $n^{29}$, so observing a residual excess above 1250 (or a collapse requiring an exponent much larger than 29) would force the quantitative claim to be rescaled.","supporting_citations":[{"cited_title":"Operations Research , volume=","cited_arxiv_id":null,"evidence_quote":"Provides the stationarity result (expected cost at most OPT + 1) that this paper extends, and the conjecture the paper confirms."},{"cited_title":", title=","cited_arxiv_id":null,"evidence_quote":"Provides the canonical-paths comparison theorems used to transfer the auxiliary chain's Dirichlet form bound to the original transposition chain."},{"cited_title":"and Tarjan, Robert E","cited_arxiv_id":null,"evidence_quote":"Supplies the quoted polynomial mixing-time bound for the biased exclusion process that Lemma 3.2 depends on; this quantitative estimate fixes the exponent in the main theorem."},{"cited_title":"Proceedings of the 36th International Conference on Algorithmic Learning Theory , series=","cited_arxiv_id":null,"evidence_quote":"Gives the local-stationarity and entropy framework whose discrete-time analogue (Lemma 4.2) controls the cost defect before the chain mixes."},{"cited_title":"The Annals of Applied Probability , volume=","cited_arxiv_id":null,"evidence_quote":"Establishes the stationary distribution of the transposition rule, which the paper uses as its reference measure throughout."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}