{"id":"94798e48-50c4-4d28-a5a8-52637114fe2a","arxiv_id":"2411.17241","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For a wide class of f-divergences, asymptotic Markov chain contraction rates are characterized by the input-dependent chi-squared contraction coefficient, and the result extends to Petz f-divergences in quantum settings.","lead":"This paper proves that for many divergence measures, the speed at which a Markov chain forgets its initial state is governed by a single computable quantity: the chi-squared contraction coefficient. It also extends the same conclusion to quantum systems, where the quantity is identified but not yet efficiently computable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 50 proves only an upper bound; the advertised claim that η_{χ²}(W,π) characterizes the asymptotic contraction rate is refuted by non-reversible chains, where the true χ² rate can be strictly smaller.","rationale":"The reader's CONDITIONAL verdict is appropriate: the paper contains a useful and likely correct upper-bound theorem, but the advertised central claim needs revision. My concern differs from the reader's weakest_assumption. The reader focused on uniform convergence of κ↑_f(W^n p, W^n π) and aperiodicity; that gap is fixable because irreducibility plus aperiodicity gives uniform TV convergence, and indecomposability with a unique full-support stationary distribution effectively forces irreducibility and aperiodicity. The more load-bearing problem is that even after repairing that step, Theorem 50 only bounds the asymptotic rate from above by η_{χ²}(W,π). The one-step lower bound η_f(W,π) ≥ η_{χ²}(W,π) from [26] does not supply a matching asymptotic lower bound: contraction coefficients are submultiplicative, not multiplicative, and for non-normal transition operators the n-th root limit of η_{χ²}(W^n,π) is the squared spectral radius, which can be strictly less than the squared norm η_{χ²}(W,π). The explicit 3-state example above demonstrates this concretely: W is scrambling, so Theorem 50 applies, yet the χ² rate is 0.44 < 0.654 ≈ η_{χ²}(W,π). Therefore the abstract's 'characterized by' and the reader's combined implication are not supported. The correct claim is an efficiently computable upper bound on the asymptotic contraction rate, with tightness only for reversible chains as the paper notes. Since this is a substantive overclaim but the underlying theorem and technique remain valuable, the reader's CONDITIONAL verdict should stand, with the revision requirement now clearly including a weakening of the central claim.","tokens_in":47065,"tokens_out":27730,"duration_ms":252866,"concrete_test":"For W = [[0.9,0.1,0],[0,0.8,0.2],[0.4,0,0.6]] and π=(4,2,1)/7, compute η_{χ²}(W^n,π)^{1/n} numerically for n up to, say, 100 and compare with η_{χ²}(W,π). The sequence should converge to 0.44 while η_{χ²}(W,π) ≈ 0.654. If it does, the 'characterization' claim fails even for χ²-divergence itself, and only the upper-bound reading of Theorem 50 survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim, in the abstract and in the reader's strongest_claim, is that η_{χ²}(W,π) determines the asymptotic contraction rate for smooth f-divergences. But Theorem 50, Eq. (44), establishes only limsup_n η_f(W^n,π)^{1/n} ≤ η_{χ²}(W,π). The cited lower bound [26] is one-step: η_f(W,π) ≥ η_{χ²}(W,π). Applying it to W^n and taking n-th roots gives only liminf_n η_f(W^n,π)^{1/n} ≥ lim_n η_{χ²}(W^n,π)^{1/n}. For a non-reversible (non-normal) chain, the right-hand side is the square of the spectral radius of the normalized transition operator restricted to the mean-zero subspace, which can be strictly smaller than η_{χ²}(W,π), the square of its norm. A concrete example is W = [[0.9,0.1,0],[0,0.8,0.2],[0.4,0,0.6]] on three states, with stationary distribution π=(4,2,1)/7. W is scrambling, so Theorem 50 applies. For the normalized operator A = D^{1/2} W D^{-1/2}, the non-unit eigenvalues have modulus² = 0.44, so lim_n η_{χ²}(W^n,π)^{1/n} = 0.44, while η_{χ²}(W,π) = ||A||² ≈ 0.654. Thus even for f corresponding to χ²-divergence itself, the actual asymptotic rate is strictly below η_{χ²}(W,π). The theorem can support an upper-bound statement, not a characterization. The uniformity gap in the proof flagged by the reader is real but secondary: under the stated hypotheses, uniform convergence of W^n p to π is obtainable, whereas the missing lower bound is not a proof gap but a false conclusion for non-reversible chains.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops Taylor-expansion-based inequalities for twice-differentiable f-divergences: Pinsker-type lower bounds (Theorems 7 and 20) and two-sided bounds in terms of the chi-squared divergence (Theorem 31). These tools are then applied to input-dependent contraction coefficients of time-homogeneous Markov chains. The central advertised result, Theorem 50, asserts that for many f-divergences the asymptotic contraction rate is characterized by the input-dependent chi-squared contraction coefficient, which is efficiently computable. The paper also extends the divergence inequalities and ergodic applications to Petz f-divergences in quantum information theory, though without an efficient computation guarantee.","tokens_in":77,"tokens_out":12267,"duration_ms":350094,"significance":"If the main application were correct as stated, the paper would be a substantial unification: most smooth f-divergence contraction rates would be governed by the efficiently computable chi-squared coefficient, with clean consequences for mixing times. The underlying Pinsker and chi-squared comparison inequalities in Sections III and IV are clearly useful, appear to be proved correctly, and generalize several known results. The quantum extension in Section VI is also of interest, even though computability is explicitly not guaranteed. However, the central characterization claim in the abstract and Section V is not supported by Theorem 50, which proves only an upper bound; for non-reversible chains the actual asymptotic rate can be strictly smaller than the chi-squared contraction coefficient. This overclaim is load-bearing and must be corrected or qualified before the paper can be accepted.","major_comments":[{"comment":"The central claim that eta_{chi^2}(W,pi) characterizes the asymptotic contraction rate is not established and is false as stated. Theorem 50 establishes at most limsup_n eta_f(W^n,pi)^{1/n} <= eta_{chi^2}(W,pi); for non-reversible chains the true rate can be strictly smaller. For the scrambling chain W = [[0.9,0.1,0],[0,0.8,0.2],[0.4,0,0.6]] with stationary distribution pi=(4,2,1)/7, the normalized operator A = D^{1/2} W D^{-1/2} has ||A||^2 = eta_{chi^2}(W,pi) ≈ 0.654, while the squared modulus of the non-unit eigenvalues is 0.44, so lim_n eta_{chi^2}(W^n,pi)^{1/n} = 0.44 for the chi^2-divergence itself. Thus the rate is not characterized by eta_{chi^2}(W,pi) except in the reversible case. The abstract's 'characterized by' and Section I's 'scale at a rate given by' must be replaced by an upper-bound statement, with tightness explicitly restricted to reversible chains.","section":"Abstract; Section V, Theorem 50, Eq. (44)"},{"comment":"The proof that condition 3 (W indecomposable, pi full support) implies convergence of W^n p to pi is invalid. The two-state swap chain W = [[0,1],[1,0]] with pi=(1/2,1/2) is indecomposable with full-support stationary distribution, but eta_{chi^2}(W,pi)=1 and W^n p does not converge for p ≠ pi. Inequalities (47)-(49) only yield convergence when eta_{chi^2}(W,pi)<1, which is not implied by the stated hypotheses. Consequently, the assertion 'Thus, in all cases, lim_n W^n p = pi' is false. Moreover, the passage to lim_n sup_{p} kappa^up_f(W^n p, W^n pi) = f''(1) in Eqs. (50)-(52) requires uniform convergence over p in P(X); for conditions 1 and 2 this can be recovered from uniform TV bounds, but for condition 3 it is not established. The theorem should either add eta_{chi^2}(W,pi)<1 (or an equivalent aperiodicity-type assumption) and prove the uniformity, or remove condition 3.","section":"Section V, proof of Theorem 50, condition 3"},{"comment":"The theorem states a limit, lim_{n to infinity} eta_f(W^n,pi)^{1/n} <= eta_{chi^2}(W,pi), but the proof establishes at most a limsup inequality: the sequence eta_f(W^n,pi)^{1/n} need not converge for every f and W. The statement and the limit laws used in Eqs. (50)-(52) should be rewritten with limsup; otherwise the theorem asserts existence of a limit that is not proved.","section":"Section V, Theorem 50 statement"}],"minor_comments":[{"comment":"In the displayed chain of inequalities, the symbol 'L_f ,1' appears in the denominator; it should be 'L_f'.","section":"Section V, proof of Lemma 49"},{"comment":"The text says the chain 'increases the value by one modulo 3' for a four-state alphabet; this should be 'modulo 4'.","section":"Section V, Example 48"},{"comment":"The sentence 'Moreover, we know the above bound can be tight given Theorem 50 and Proposition 65' is not justified as stated: Proposition 65 only reduces the quantum contraction coefficient to the classical one for classical-to-classical channels, and tightness in Theorem 50 is proved only for reversible Markov chains. The claim should be qualified or removed.","section":"Section VI, Theorem 72"},{"comment":"The constant L_f appears in the assumptions before it is defined; the statement should explicitly say that L_f is a positive real number satisfying condition (8) or (9).","section":"Section III, Theorem 7"},{"comment":"In Eq. (50) the exponent on eta_f(W^n,pi) is displayed as n rather than 1/n; the displayed limit should read lim_n eta_f(W^n,pi)^{1/n}.","section":"Section V, proof of Theorem 50, Eq. (50)"}],"recommendation":"major_revision","confidential_remarks":"The divergence inequalities in Sections III and IV are strong and likely publishable, and the paper clearly aims at an important unification. However, the advertised characterization of ergodic contraction rates by the chi-squared coefficient is not what Theorem 50 proves, and the non-reversible counterexample is decisive. The authors should be asked to correct the abstract and introductory claims, repair the hypotheses and proof of Theorem 50, and resubmit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful core of this paper is the Taylor-based machinery in Sections III and IV. The Pinsker-style bounds in Tables I and II are clean, explicit, and genuinely convenient; the constants check out, and the method is more general than Gilardoni's in a real way. The derivation of divergence relations from an integral form of Taylor's theorem is straightforward, reproducible, and clearly explained. I'd happily point a student at Theorem 31 and Corollary 33 as a tidy way to bound f-divergences by χ². The quantum lift to Petz f-divergences is mostly a corollary ride on the Nussbaum-Skoła representation, but it is honest about what it does and does not give: they never claim efficiency there, and they flag it.\n\nThe soft spot is the central ergodic claim. Theorem 50 proves an upper bound: limsup of the nth root of η_f(W^n,π) is at most η_{χ²}(W,π). The lower bound imported from [26] is one-step, and applying it to W^n and taking nth roots only gives a liminf bound in terms of the asymptotic spectral radius of the normalized operator. For non-reversible chains, that spectral radius can be strictly smaller than η_{χ²}(W,π), as the three-state scrambling example in the stress-test shows. So the abstract's phrase \"rate of contraction ... is characterized by the input-dependent contraction coefficient of the χ²-divergence\" is too strong. The theorem supports a one-sided rate statement, not a characterization. That is a real gap between the advertised result and the proved result, and it is not just a missing uniformity argument. The uniformity issue the reader flagged is secondary; under the stated hypotheses that part can be patched, whereas the missing lower bound is a false conclusion for non-reversible chains.\n\nThe paper also leans on the authors' prior ITW work without, as far as I can tell, clearly flagging the degree of overlap in the divergence-inequality core. That matters for a journal submission, though it is fixable by disclosure rather than by changing the math.\n\nWho is this for? People working on contraction coefficients, strong data processing inequalities, and computable mixing-time bounds. The divergence inequalities are a solid reference contribution. The ergodic characterization needs to be restated as an upper bound or restricted to the reversible case where the tightness actually holds. With that revision, this deserves a serious referee.","headline":"Useful Taylor-based divergence inequalities and a clean upper bound on f-divergence contraction rates, but the advertised characterization of the asymptotic rate by the χ² coefficient overstates what the proof delivers for non-reversible chains.","tokens_in":48007,"tokens_out":906,"would_cite":true,"duration_ms":11052,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A17","60J10","62B10"],"pacs":[],"model":"deepseek-v4-flash","headline":"For ergodic Markov chains, most smooth f-divergences contract asymptotically at the efficiently computable χ² rate.","keywords":["f-divergences","χ² divergence","contraction coefficients","data processing inequality","Pinsker inequality","Markov chain mixing times","Petz f-divergences","ergodic theory"],"falsifier":"Numerically evaluate, for the 4-state noisy-typewriter chain of Example 48 or any small irreducible aperiodic chain, the quantities $\\eta_f(W^n,\\pi)^{1/n}$ for $f(t)=t\\log t$ and $\\eta_{\\chi^2}(W,\\pi)$ to high precision for large n, while tracking $\\sup_p\\kappa_f^{\\uparrow}(W^n p,\\pi)$ over starting distributions concentrated near the boundary of the simplex. If the supremum curvature does not converge to $f''(1)$, or if the n-th root stabilizes strictly above $\\eta_{\\chi^2}(W,\\pi)$, the uniform-convergence premise is violated and Eq. (52) would need a correction term.","tokens_in":46818,"feed_emoji":"📉","tokens_out":9511,"duration_ms":82942,"temperature":0.7,"pith_summary":"Most smooth f-divergences can be sandwiched between constant multiples of the χ²-divergence once their second-derivative curvature is bounded. Iterating a contractive channel makes those curvature factors converge to f''(1), leaving the χ² contraction coefficient as the n-th-root rate for most f-divergences. Because the χ² coefficient is the squared maximal correlation of the channel, the rate is computable and is the fastest a divergence in this class can contract. This gives explicit mixing-time bounds and a quantum version for Petz f-divergences.","feed_headline":"One χ² rate governs contraction for most smooth f-divergences","feed_subtitle":"Ergodic Markov chains: many smooth f-divergences share the efficiently computable χ² contraction rate.","key_machinery":"The central object is the second-order Taylor integral representation of an f-divergence, $$D_f(p\\|q)=\\$int_0^{1}$(1-t)\\sum_{i\\in\\operatorname{supp}(q)} $q_i^{{-1}}$ f''\\left(1+t\\left(\\frac{p_i}{q_i}-1\\right)\\right)(p_i-q_i)^2\\,dt,$$ which immediately gives $$\\frac{\\$kappa_f^{{\\downarrow}}$(p,q)}{2}\\$chi^{2}$(p\\|q)\\leq D_f(p\\|q)\\leq \\frac{\\$kappa_f^{{\\uparrow}}$(p,q)}{2}\\$chi^{2}$(p\\|q),$$ where $\\kappa_f^{\\uparrow}$ and $\\kappa_f^{\\downarrow}$ are the maximum and minimum of f'' along the likelihood-ratio segments between p and q. All later results flow from this sandwich: contraction-coefficient ratios are bounded by χ² ratios times curvature ratios, and after n channel iterations the curvature ratios move to f''(1). The χ² side is then evaluated through the maximal-correlation identity $\\eta_{\\chi^2}(W,\\pi)=\\rho_m(X,Y)^2$, which is spectral and efficiently computable.","core_discovery":"The load-bearing result is that for a finite Markov chain W with unique stationary distribution π, and any twice continuously differentiable convex f with f(1)=0, f''(1)>0 and finite positive L_f, the asymptotic contraction exponent obeys $\\lim_{n\\to\\infty}\\eta_f(W^n,\\pi)^{1/n}\\leq \\eta_{\\chi^2}(W,\\pi)$ whenever W is irreducible and aperiodic, or scrambling, or indecomposable with π of full support. Together with the known lower bound $\\eta_f(W,\\pi)\\geq \\eta_{\\chi^2}(W,\\pi)$, this identifies the input-dependent χ² contraction coefficient as the canonical asymptotic exponent for smooth f-divergences, and in reversible chains the bound is attained. The proof passes through new sandwiched inequalities $D_f(p\\|q)\\in[\\kappa_f^{\\downarrow}(p,q)/2,\\kappa_f^{\\uparrow}(p,q)/2]\\chi^2(p\\|q)$, then shows the curvature factors converge to f''(1) under channel iteration. The same mechanism yields computable mixing times in total variation and in f-divergence, and an analogue for Petz f-divergences on quantum mixing channels.","pith_inferences":["A testable extension is the time-inhomogeneous case: iterating different channels should make the n-th-root rate the geometric mean of the per-step χ² contraction coefficients rather than a single spectral number.","A natural next step is to supply an explicit uniform-convergence argument for $\\kappa_f^{\\uparrow}(W^n p,W^n\\pi)$ over the simplex; a Doeblin-coefficient bound would likely provide it.","On the quantum side, any future efficient estimate of the Petz χ² contraction coefficient would immediately turn Theorem 72 into an operational mixing-time tool, a computational question the paper leaves open.","The sandwich inequalities are not tied to Markov chains and could be used to estimate f-divergences under any iterative data-processing map, including approximate Bayesian updates and privacy mechanisms."],"forward_implications":["For irreducible aperiodic, scrambling, or indecomposable full-support stationary Markov chains, the asymptotic contraction rate of any smooth f-divergence is bounded by the χ² rate; for reversible chains the two rates coincide.","Because $\\eta_{\\chi^2}(W,\\pi)$ is the squared maximal correlation, the universal rate is computable from the spectrum of the channel's centered joint matrix.","The refined convergence corollary gives explicit finite-n bounds on total-variation distance to stationarity in terms of $\\eta_{\\chi^2}(W,\\pi)$, $\\pi_{\\min}$, and the target accuracy, without requiring irreducibility.","For f-divergences with concave $g(t)=(f(t)-f(0))/t$, mixing times obey the same $\\log(1/\\eta_{\\chi^2})$ scaling up to constants.","The analogous theorem holds for Petz f-divergences on quantum mixing channels, with the Petz χ² contraction coefficient replacing the classical one, though no efficient computation is claimed."],"supporting_citations":[{"why":"It supplies the input-dependent χ² contraction coefficient as the squared maximal correlation, along with its efficient computation and its status as the fastest contraction rate among f-divergences.","marker":"[26]"},{"why":"It states the lower bound $\\eta_f(W,\\pi)\\geq\\eta_{\\chi^2}(W,\\pi)$, proves the KL analogue that Theorem 50 generalizes, and gives the multiplicativity used for reversible chains.","marker":"[27]"},{"why":"It provides the linear f-divergence to χ² bound used in the mixing-time corollaries and in the linear contraction-coefficient proposition.","marker":"[25]"},{"why":"It gives the previous tight Pinsker-inequality framework that Theorems 7 and 20 extend to arbitrary twice-differentiable f with computable constants.","marker":"[21]"},{"why":"It documents the local behavior of f-divergences approaching the χ² divergence, which motivates the sandwich and the $\\kappa\\to f''(1)$ limit.","marker":"[16]"},{"why":"It introduces the Petz quasi-entropies, the quantum family to which Section VI extends the classical results.","marker":"[5]"},{"why":"It supplies the Nussbaum-Szkoła distribution representation of Petz f-divergences and the data-processing condition used in the quantum theorem.","marker":"[82]"},{"why":"It introduces the Petz χ²-divergence and the previous quantum mixing-time bounds that the quantum rate theorem builds on.","marker":"[51]"}],"fun_headline_variants":["Many f-divergences share χ²'s contraction rate","χ² sets the pace for f-divergence contraction","Smooth f-divergences: χ² contraction is the rule","One χ² rate decides contraction for smooth f-divergences","For Markov chains, χ² contraction rate bounds all smooth f-divergences"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof needs $\\sup_{p\\in\\mathcal{P}(X)}\\kappa_f^{\\uparrow}(W^n p,W^n\\pi)\\to f''(1)$: the curvature factor must converge uniformly over all starting distributions so that its n-th root tends to one. The manuscript invokes the classical convergence theorem, which gives pointwise total-variation convergence, but does not supply a uniformity argument, and it does not show that indecomposability with full-support π implies aperiodicity. If uniformity fails, Eq. (52) acquires an extra subexponential or exponential factor and the stated rate is not the whole story.","fun_headline_variants_meta":{"raw":{"variants":["Many f-divergences share χ²'s contraction rate","χ² sets the pace for f-divergence contraction","Smooth f-divergences: χ² contraction is the rule","One χ² rate decides contraction for smooth f-divergences","For Markov chains, χ² contraction rate bounds all smooth f-divergences"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001686,"raw_usage":{"total_tokens":6690,"prompt_tokens":960,"completion_tokens":5730,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":5642}},"tokens_in":576,"tokens_out":5730,"duration_ms":38287,"temperature":1.0,"reasoning_tokens":5642,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:23:02.863292+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Numerically evaluate, for the 4-state noisy-typewriter chain of Example 48 or any small irreducible aperiodic chain, the quantities $\\eta_f(W^n,\\pi)^{1/n}$ for $f(t)=t\\log t$ and $\\eta_{\\chi^2}(W,\\pi)$ to high precision for large n, while tracking $\\sup_p\\kappa_f^{\\uparrow}(W^n p,\\pi)$ over starting distributions concentrated near the boundary of the simplex. If the supremum curvature does not converge to $f''(1)$, or if the n-th root stabilizes strictly above $\\eta_{\\chi^2}(W,\\pi)$, the uniform-convergence premise is violated and Eq. (52) would need a correction term.","supporting_citations":[{"cited_title":"Strong data-processing inequalities for channels and bayesian networks","cited_arxiv_id":null,"evidence_quote":"It supplies the input-dependent χ² contraction coefficient as the squared maximal correlation, along with its efficient computation and its status as the fastest contraction rate among f-divergences."},{"cited_title":"Comparison of contract ion coefficients for f-divergences","cited_arxiv_id":null,"evidence_quote":"It states the lower bound $\\eta_f(W,\\pi)\\geq\\eta_{\\chi^2}(W,\\pi)$, proves the KL analogue that Theorem 50 generalizes, and gives the multiplicativity used for reversible chains."},{"cited_title":"Strong data processing inequalities a nd phi-sobolev inequalities for discrete channels","cited_arxiv_id":null,"evidence_quote":"It provides the linear f-divergence to χ² bound used in the mixing-time corollaries and in the linear contraction-coefficient proposition."},{"cited_title":"On Pinsker’s and Vajda’s type ineq ualities for csisz´ ar’s f -divergences","cited_arxiv_id":null,"evidence_quote":"It gives the previous tight Pinsker-inequality framework that Theorems 7 and 20 extend to arbitrary twice-differentiable f with computable constants."},{"cited_title":"On f-divergences: Integral representatio ns, local behavior, and inequalities","cited_arxiv_id":null,"evidence_quote":"It documents the local behavior of f-divergences approaching the χ² divergence, which motivates the sandwich and the $\\kappa\\to f''(1)$ limit."},{"cited_title":"Different quantum f-div ergences and the reversibility of quantum operations","cited_arxiv_id":null,"evidence_quote":"It supplies the Nussbaum-Szkoła distribution representation of Petz f-divergences and the data-processing condition used in the quantum theorem."},{"cited_title":"The χ 2-divergence and mixing times of quantum Markov processes","cited_arxiv_id":null,"evidence_quote":"It introduces the Petz χ²-divergence and the previous quantum mixing-time bounds that the quantum rate theorem builds on."}],"review_version":1}