{"id":"c6b5132f-52dc-4ab7-89d8-7d23bef3688c","arxiv_id":"2411.16913","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Tsallis and Sharma-Mittal entropies of a Poisson distribution increase with λ, while two generalized Rényi entropies can be non-monotone in λ.","lead":"This paper studies six entropy measures for the Poisson distribution and shows that most of them increase with the intensity λ, while two generalized Rényi entropies can dip down for some parameter values. The result is a caution for anyone using these entropies on count data: not every entropy measure grows when the underlying distribution becomes more spread out.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central monotonicity/anomaly claims appear sound, but Theorem 4.1 items 4–5 assert false Tsallis/Sharma–Mittal lower bounds for α>1 and β>1; CONDITIONAL acceptance remains appropriate.","rationale":"The reader's conditional verdict is correct overall, but not for the primary reason stated. Proposition 5.4's inference can be made rigorous via the mean value theorem together with continuity of ρ′, so the small-α non-monotonicity claim does not rest on a logical gap; the large-α case is explicitly labeled numerical in Remark 5.7. However, Theorem 4.1 items 4 and 5 contain a concrete false inequality: for α>1 and β>1 the correct exponential factor is e^{(α−1)h(λ)} (respectively e^{(β−1)h(λ)}), not e^{−h(λ)}. This is verified numerically at α=2, λ=1, where HT(2,1)≈0.6915 is below the claimed lower bound ≈0.6954. Since the abstract advertises lower and upper bounds as a contribution, the manuscript needs correction before acceptance. The monotonicity results are unaffected, so a conditional verdict remains appropriate.","tokens_in":22714,"tokens_out":13495,"duration_ms":112598,"concrete_test":"Evaluate the claimed lower bound of Theorem 4.1 item 4 at α=2, λ=1: compute HT(2,1)=1−e^{−2}Σ_{i=0}^∞ 1/(i!)^2 and compare with 1−(2π)^{−1/2}e^{−h(1)}, where h(1)=0.5 log 2 − 1/13. If HT(2,1) < RHS, the inequality fails. Independently, re-derive the exponent by substituting (4.6) into (2.13) for α>1 and check whether the factor is e^{(α−1)h(λ)} rather than e^{−h(λ)}.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central anomaly claim survives scrutiny: in Proposition 5.4, after setting ρ(0)=0, strict positivity gives ρ(δ)>0; the mean value theorem yields c∈(0,δ) with ρ′(c)>0, and continuity of ρ′ gives an interval of positivity that can be placed arbitrarily close to 0 by choosing δ small. The paper's terse inference is therefore valid, and the large-α anomaly is explicitly labeled numerical in Remark 5.7. The genuine load-bearing error is in Theorem 4.1 items 4–5. For α>1, substituting the bound (4.6), μ(λ)<(2πλ)^{-1/2}e^{h(λ)}, into (2.13) yields HT(α,λ) ≥ 1/(α−1)[1−(2πλ)^{−(α−1)/2}e^{(α−1)h(λ)}], not the e^{−h(λ)} factor claimed. The claimed bound is stronger and is false: at α=2, λ=1, HT(2,1)=1−ψ(2,1)=1−e^{−2}(1+1+1/4+1/36+…)≈0.6915, while 1−(2π)^{−1/2}e^{−h(1)}≈0.6954, so HT lies below the stated lower bound. The same sign error propagates to item 5 for the Sharma–Mittal bound with β>1. This does not invalidate the monotonicity and anomaly theorems, but the advertised bounds are incorrect as stated and must be corrected.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies six entropy functionals—Shannon, Rényi, Tsallis, Sharma–Mittal, and two generalized Rényi entropies—for the Poisson distribution with intensity λ. It expresses all of them through the function ψ(α,λ)=Σ_i p_i(λ)^α, derives explicit formulas, asymptotic expansions as λ→∞, upper and lower bounds, and monotonicity results in λ. The main claims are that Shannon, Rényi, Tsallis, and Sharma–Mittal entropies are strictly increasing in λ for all admissible parameters, while the generalized Rényi entropies can be non-monotone for α near 0 (proved) and for large α (presented as numerical evidence), with the two-parameter version inheriting this non-monotonicity locally.","tokens_in":22984,"tokens_out":11750,"duration_ms":100068,"significance":"If correct, the monotonicity and anomaly results constitute a rigorous, parameter-free characterization of entropy behavior for Poisson count data, with practical cautionary value for researchers using generalized Rényi entropies as uncertainty measures. The ψ-based proof strategy is clean and yields explicit asymptotic constants; the main monotonicity theorems are correct, the α-near-0 anomaly proof is valid although terse, and the large-α anomaly is appropriately labeled numerical. The paper also provides useful explicit bounds, but the bound theorem contains a false statement for the Tsallis and Sharma–Mittal entropies in the α,β>1 regime; this does not undermine the monotonicity theorems but must be corrected before acceptance.","major_comments":[{"comment":"For α>1 and β>1, the stated replacements are false. Combining (4.6) with (2.13) gives H_T(α,λ) ≥ 1/(α−1)[1−(2πλ)^{−(α−1)/2} e^{(α−1)h(λ)}], not the factor e^{−h(λ)} claimed in item 4; the analogous factor for H_SM(α,β,λ) in item 5 is e^{(β−1)h(λ)}, not e^{−h(λ)}. The stated stronger bound is numerically violated: at α=2, λ=1, H_T(2,1)≈0.6915, while the claimed lower bound 1−(2π)^{−1/2}e^{−h(1)}≈0.6954. For 0<α<1 and 0<β<1 the claimed bounds are true but weaker than what follows from (4.6); the theorem should state the correct exponential factors for each parameter range.","section":"Theorem 4.1, items 4–5 (proof on p. 10; eqs. (2.13), (3.14)–(3.15), (4.6))"}],"minor_comments":[{"comment":"The inference that ρ must be increasing on an interval near α=0 is terse; please spell out the argument: for δ>0, the mean value theorem gives c∈(0,δ) with ρ′(c)=ρ(δ)/δ>0, and continuity of ρ′ then yields an interval around c, which can be chosen inside an arbitrarily small neighborhood of 0.","section":"Proposition 5.4, proof"},{"comment":"In the proof of the lower bound for H_GR(α,β,p), the displayed expression should be −log μ(p), not μ(p).","section":"Proposition 2.3, proof step 3"},{"comment":"For α>1 the text says the entropy is bounded by/increases to 1/(1−α); this should read 1/(α−1).","section":"Remark 5.2"},{"comment":"The anomalous behavior for large α is numerical evidence, not a theorem; please state explicitly in the abstract or introduction that the proved anomaly covers α near 0, while the large-α anomaly is observed numerically.","section":"Abstract and Remark 5.7"},{"comment":"The initial restriction γ∈(α,1) is incompatible with the later uniform choice γ≥γ* when α>γ*; rephrase the proof to allow γ∈(0,1) and rely on the uniform argument at the end of the lemma.","section":"Lemma A.6, part 1"},{"comment":"The proof covers β>1>α and β>1=α; please state explicitly that the α>1≥β case follows by the symmetry H_GR(α,β,λ)=H_GR(β,α,λ).","section":"Proposition 5.8, proof of item 1"}],"recommendation":"major_revision","confidential_remarks":"The main monotonicity and anomaly claims are sound; the paper should be publishable after Theorem 4.1 is corrected and the proof of Proposition 5.4 is expanded. I found no circularity: the use of [9] concerns an auxiliary published result, and the anomaly is derived independently from derivatives of log ψ."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline: this paper proves something new and worth knowing. For the Poisson distribution, Tsallis and Sharma-Mittal entropies are strictly increasing in the intensity lambda for all admissible parameters, and the proof is a clean reduction to the previously established monotonicity of psi(alpha, lambda). That part is correct and nicely complements the earlier Shannon/Renyi results. The genuinely interesting, and also new, finding is that the one- and two-parameter generalized Renyi entropies are not always monotone in lambda. The small-alpha anomaly is rigorously supported by the proof of Proposition 5.4, and the two-parameter extension in Proposition 5.8 follows naturally. The large-alpha anomaly appears only in numerical remarks; the authors say so explicitly in Remark 5.7, so that is a limitation, not a defect.\n\nThe reader's worry about Proposition 5.4 does not survive a close look. The argument is terse, but it is valid: rho is C^1, positive on (0,1), with rho(0)=rho(1)=0; strict positivity at a point plus the mean value theorem gives rho' > 0 somewhere arbitrarily close to 0, and continuity of rho' then gives an interval. It would be better to spell this out, but the logic holds.\n\nThe real problem is in Theorem 4.1, items 4 and 5. The substitution of the refined maximal-probability bound (4.6) into the general lower bounds (2.13) and (2.14) is done with the wrong exponent. For alpha > 1, the correct factor is e^{+(alpha-1)h(lambda)} in the subtracted term, not e^{-h(lambda)}; for alpha in (0,1) and for the Sharma-Mittal cases the same sign error appears. The claimed bound is stronger than what the derivation yields and is in fact false at, e.g., alpha=2, lambda=1 for Tsallis. This does not damage the monotonicity theorems or the anomaly claim, but the advertised estimates must be corrected.\n\nVerdict: this deserves a serious referee. The main results are clean, the anomaly is a useful caution for practitioners, and the errors are localized and fixable. The paper should go to review with a request to correct Theorem 4.1 and to make the numerical large-alpha anomaly explicitly conditional.","headline":"Solid, modest paper: the Tsallis/Sharma-Mittal monotonicity results are clean, the anomaly claim is plausible, but two advertised bounds in Theorem 4.1 have a wrong exponential factor and need correction.","tokens_in":23595,"tokens_out":2049,"would_cite":true,"duration_ms":18304,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A17","60E05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Tsallis and Sharma–Mittal entropies of the Poisson distribution always increase with intensity λ, but two generalized Rényi entropies can decrease over intervals of λ—an 'anomalous' non-monotonicity.","keywords":["Shannon entropy","Rényi entropy","Tsallis entropy","Sharma–Mittal entropy","generalized Rényi entropy","Poisson distribution","monotonicity","asymptotic bounds"],"falsifier":"Compute ρ(α)=α((Σ_{i≥1} i/(i!)^α)/(Σ_{i≥0}1/(i!)^α)−1) and ρ′(α) for α=$10^{{-k}}$, k=3,...,12, using the series or high-precision quadrature, and check whether ρ′(α)>0 on an entire interval near 0. Proposition 5.4 predicts such an interval; if ρ′≤0 for all sufficiently small α, the claimed decreasing behavior of H_GR(α,λ) near λ=1 would not follow from the given argument. Conversely, isolating an α with ∂_λ H_GR(α,λ)|_{λ=1}<0 would confirm the anomaly directly.","tokens_in":22458,"feed_emoji":"📉","tokens_out":9776,"duration_ms":77798,"temperature":0.7,"pith_summary":"The paper asks which entropy measures of a Poisson distribution with intensity λ keep their intuitive monotone behavior as λ grows. It proves that the Tsallis entropy H_T(α,λ) and the Sharma–Mittal entropy H_SM(α,β,λ) are strictly increasing in λ for every admissible parameter value, joining the already known monotonicity of the Shannon and Rényi entropies. In contrast, it proves that the one-parameter generalized Rényi entropy H_GR(α,λ) can decrease as λ increases: for each α in a small interval near 0, H_GR(α,λ) is decreasing in a neighborhood of λ=1, and numerical evidence indicates similar non-monotone behavior for large α. The two-parameter generalized Rényi entropy H_GR(α,β,λ) inherits this anomalous decrease when the parameters are close to a region where H_GR(α,·) decreases. The paper also gives leading-order asymptotics as λ→∞ and two-sided bounds for the entropies, the Shannon case matching the known (1/2)log(2πλ) growth.","feed_headline":"Tsallis and Sharma–Mittal climb, but two Rényi entropies dip","feed_subtitle":"For Poisson counts, these two Rényi variants can fall as the rate grows; Shannon, Tsallis, and Sharma–Mittal rise.","key_machinery":"The load-bearing object is ψ(α,λ)=Σ_{i≥0} p_i(λ)^α=$e^{{-αλ}}$Σ_{i≥0} $λ^{{iα}}$/(i!)^α. Every one of the six entropies is a functional of ψ and its logarithmic derivative: H_R=(1/(1−α))log ψ, H_T=(ψ−1)/(1−α), H_GR(α,β)=(log ψ(α,·)−log ψ(β,·))/(β−α), H_SM=($ψ^{{(1−β)/(1−α)}}$−1)/(1−β), and H_GR(α,·)=−∂_α log ψ. Since ψ is increasing in λ for 0<α<1 and decreasing for α>1, the signs in the definitions force Shannon, Rényi, Tsallis, and Sharma–Mittal entropies to increase in λ. The anomalous cases come from the mixed derivative: ∂_λ H_GR(α,λ)=−∂_α∂_λ log ψ(α,λ), evaluated at λ=1 through ρ(α)=α((Σ_{i≥1} i/(i!)^α)/(Σ_{i≥0}1/(i!)^α)−1); when ρ increases in α, the generalized Rényi entropy initially falls as λ rises.","core_discovery":"On the paper's own terms, the central result is a classification of 'normal' versus 'anomalous' behavior for six entropy families on Poisson probabilities p_i(λ)=$e^{{-λ}}$λ^i/i!. Proposition 5.1 (with earlier work) establishes that Shannon, Rényi, Tsallis, and Sharma–Mittal entropies are strictly increasing functions of λ for all admissible parameter values. Proposition 5.4 establishes the existence of an interval J⊂(0,1) near 0 such that for every α∈J, the generalized Rényi entropy H_GR(α,λ)=−∂_α log ψ(α,λ) is decreasing in λ in a neighborhood of λ=1, so monotonicity fails even though ρ(α)=∂_λ log ψ|_{λ=1} is positive on (0,1). Proposition 5.8 transfers this non-monotonicity to the two-parameter entropy H_GR(α,β,λ) for β near such an α. The same section reports numerical observations of non-monotonicity for large α. Theorems 3.4 and 4.1 give the λ→∞ asymptotics and bounds: Shannon, Rényi, and both generalized Rényi entropies diverge logarithmically; Tsallis entropy grows like a power for α∈(0,1) and converges to 1/(α−1) for α>1; Sharma–Mittal entropy does the same with β in place of α.","pith_inferences":["The derivative identities used here—expressing each entropy through ψ and log ψ—are general; for any discrete distribution where ψ(α,λ) can be controlled, the same argument would classify monotonicity, so the Poisson results are likely a template rather than an isolated example.","The large-α anomaly is presented numerically only; proving it analytically, for example by showing that the oscillating derivative ∂_λ H_GR(α,λ) crosses zero for every sufficiently large α, remains an open extension the paper does not attempt.","In Poisson applications, the decreasing interval near λ=1 for small α means that a generalized Rényi-based score can fall when more events are observed, so practitioners should check monotonicity for their parameter range before using H_GR as an uncertainty measure."],"forward_implications":["For a Poisson count model, the Tsallis and Sharma–Mittal entropies can be used as uncertainty measures that are guaranteed to grow with the rate λ, just as Shannon and Rényi entropies do.","The one-parameter generalized Rényi entropy is not a universally reliable uncertainty measure: for α in a small interval near 0, increasing λ can lower the entropy over an interval starting at λ=1.","For the two-parameter generalized Rényi entropy, any parameter pair close to an anomalous α inherits the decrease, so the phenomenon is not isolated to a single parameter choice.","The asymptotic formulas give practical approximations: for large λ, Shannon, Rényi, and generalized Rényi entropies grow like (1/2)log(2πλ) plus parameter-dependent constants, while Tsallis (0<α<1) and Sharma–Mittal (0<β<1) grow as power laws.","Two-sided estimates, e.g. for Shannon entropy L(λ) ≤ H_SH(λ) ≤ U_SH(λ), control the entropies uniformly for λ>1 and agree with the leading asymptotic term."],"supporting_citations":[{"why":"Establishes the monotonicity of ψ(α,λ) in λ and the already-known Shannon/Rényi monotonicity that this paper extends to Tsallis and Sharma–Mittal entropies.","marker":"[9]"},{"why":"Provides the definition of Rényi entropy on which the generalized forms are built.","marker":"[27]"},{"why":"Source of the Tsallis entropy definition whose Poisson behavior is proved to be monotone.","marker":"[36]"},{"why":"Source of the Sharma–Mittal entropy definition whose Poisson behavior is proved to be monotone.","marker":"[31]"},{"why":"Source of the one- and two-parameter generalized Rényi entropy definitions that exhibit anomalous non-monotonicity.","marker":"[1]"},{"why":"Supplies the known asymptotic (1/2)log(2πλ)+1/2 for the Shannon entropy used to anchor the asymptotic comparisons.","marker":"[8]"}],"fun_headline_variants":["Poisson entropy twist: two Rényi forms fall, others rise","Anomalous Rényi dips: Poisson entropy not always monotone","Rényi anomaly: entropy drops as rate rises for some variants","Non-monotone Rényi: Poisson entropy can decline with λ","Two Rényi entropies go anomalous in Poisson case"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Proposition 5.4 assumes that a positive continuously differentiable function on (0,1) which tends to 0 at 0 and equals 0 at 1 must be increasing on some interval near 0; continuity and endpoint limits alone do not force this, and the paper supplies no further analytic information about ρ(α).","fun_headline_variants_meta":{"raw":{"variants":["Poisson entropy twist: two Rényi forms fall, others rise","Anomalous Rényi dips: Poisson entropy not always monotone","Rényi anomaly: entropy drops as rate rises for some variants","Non-monotone Rényi: Poisson entropy can decline with λ","Two Rényi entropies go anomalous in Poisson case"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000251,"raw_usage":{"total_tokens":1559,"prompt_tokens":949,"completion_tokens":610,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":517}},"tokens_in":565,"tokens_out":610,"duration_ms":6219,"temperature":1.0,"reasoning_tokens":517,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:45:32.272784+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute ρ(α)=α((Σ_{i≥1} i/(i!)^α)/(Σ_{i≥0}1/(i!)^α)−1) and ρ′(α) for α=$10^{{-k}}$, k=3,...,12, using the series or high-precision quadrature, and check whether ρ′(α)>0 on an entire interval near 0. Proposition 5.4 predicts such an interval; if ρ′≤0 for all sufficiently small α, the claimed decreasing behavior of H_GR(α,λ) near λ=1 would not follow from the given argument. Conversely, isolating an α with ∂_λ H_GR(α,λ)|_{λ=1}<0 would confirm the anomaly directly.","supporting_citations":[{"cited_title":"Braiman, A","cited_arxiv_id":null,"evidence_quote":"Establishes the monotonicity of ψ(α,λ) in λ and the already-known Shannon/Rényi monotonicity that this paper extends to Tsallis and Sharma–Mittal entropies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the definition of Rényi entropy on which the generalized forms are built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the Tsallis entropy definition whose Poisson behavior is proved to be monotone."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the Sharma–Mittal entropy definition whose Poisson behavior is proved to be monotone."},{"cited_title":"Acz´ el and Z","cited_arxiv_id":null,"evidence_quote":"Source of the one- and two-parameter generalized Rényi entropy definitions that exhibit anomalous non-monotonicity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the known asymptotic (1/2)log(2πλ)+1/2 for the Shannon entropy used to anchor the asymptotic comparisons."}],"review_version":1}