{"id":"42286ea7-1d27-4812-af59-f425ae8a2980","arxiv_id":"1908.02003","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Renyi entropy violates the Shore-Johnson subset-independence axiom, so using it for maximum-entropy inference with linear constraints introduces biases not present in the data.","lead":"This paper replies to a critique of the authors' earlier claim that Renyi entropy produces artificial biases in maximum-entropy inference. It argues that Renyi entropy still violates the subset-independence axiom and that the proposed fixes by Jizba and Korbel render entropy maximization redundant or inconsistent.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (5) uses the wrong subset posteriors: the q_{S_i} are not the Rényi MaxEnt maximizers for the stated subset means, so the claimed violation of Eq. (1) is an artifact.","rationale":"The reader identified the correctness of the q-exponential posteriors in Eq. (5) as the weakest assumption; this is exactly where the argument fails. A direct check shows that the subset posterior written with the global β and the subset mean as a shift does not have that subset mean. The true MaxEnt posterior on a subset is the restriction of the global posterior renormalized, because both are q-exponentials with the same shape parameter a and the restriction has the conditional mean. With these correct subset posteriors, Eq. (1) is satisfied, so the claimed subset-independence violation disappears. This is a concrete mathematical error in the central demonstration, not a disagreement about axioms or interpretation. The paper's additional claims about Schur-concavity and the JK composition rule may have merit, but the stated central result of the reply is unsupported as written. The unsupported Livesey-Skilling assertion is secondary but reinforces that the reply needs major revision before its main conclusion can be accepted.","tokens_in":5325,"tokens_out":47561,"duration_ms":425568,"concrete_test":"For q=0.55 and β=0.1, compute U_1 from Eq. (6) using the global q_D (with U=10 so that βU=1). Numerically integrate the q_{S1} given in Eq. (5) over [0,1] and compute its mean; if the mean differs from U_1 (predicted ≈0.48 vs ≈0.43), Eq. (5) is not the correct Rényi MaxEnt posterior for the subset. Then independently solve the subset MaxEnt problem variationally and verify that q_{S1}^{true}=q_D/m_1, which makes Eq. (1) hold identically.","verdict_should_be":"REJECT","load_bearing_attack":"The central demonstration that Rényi entropy violates subset independence rests on Eq. (5). The forms claimed for q_{S1}, q_{S2} are not the true MaxEnt posteriors when the same Lagrange multiplier β is used. For q_D on [0,∞) in Eq. (5), the mean is U only if βU=1; this follows by integrating x q_D. For S1=[0,1], the true maximizer of the Rényi entropy under ∫p=1 and ∫x p=U_1 is p(x)∝(a+x)^{-1/(1-q)} with a fixed by U_1, and its Lagrange multiplier is not the global β. The paper instead writes q_{S1} with the same β and a shift U_1, which is a different member of the family. Concretely, for q=0.55 and β=0.1, the global q_D has U=10 and Eq. (6) gives U_1≈0.43. Substituting q=0.55, β=0.1, U_1≈0.43 into the q_{S1} formula of Eq. (5) yields a distribution whose mean is ≈0.48, not 0.43. Hence the stated U_1 is not the mean of the q_{S1} in Eq. (5), so the comparison against Eq. (1) is not the subset-independence test. The correct subset posterior is the conditional q_D/m_i, which is itself a q-exponential with the same parameter a; with those conditionals, Eq. (1) becomes an identity. The paper also admits that the Livesey-Skilling claim is not proved here, so that supporting assertion is currently unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a reply to Jizba and Korbel's Comment on the authors' earlier claim that Rényi entropy yields artificial biases in maximum-entropy inference. The reply makes three main points: first, that Schur-concavity, as used by JK, does not correctly identify the normalizable range of the Rényi parameter q; second, that the Rényi entropy violates the Shore-Johnson subset-independence axiom, with a new continuous example comparing Rényi and Shannon posteriors; and third, that JK's modified composition rule for system independence makes the entropy-maximization procedure redundant and introduces biases. The paper concludes that only trace-form entropies satisfy subset independence per the Livesey-Skilling criterion, though it states that the proof of this criterion is deferred to a separate paper.","tokens_in":5739,"tokens_out":16995,"duration_ms":154535,"significance":"If the central claim is correct, the reply would substantially strengthen the case that Rényi entropy is unsuitable for maximum-entropy inference under linear constraints, directly rebutting JK's defense. The paper's explicit continuous example in Eqs. (5)-(6) and Fig. 1 is a concrete attempt to test the Shore-Johnson subset-independence axiom, and the identification of the dependence of the coefficients m(Si) on the whole-set mean U is a useful diagnostic. However, the central numerical demonstration is invalid as written because the subset posteriors in Eq. (5) are not the actual Rényi maximum-entropy distributions for the stated subset means, and the key Livesey-Skilling assertion is explicitly unproved in this reply. The significance is therefore conditional on correction of the demonstration and on resolution of the interpretive dispute over the SJ axioms.","major_comments":[{"comment":"The distributions qS1(x) and qS2(x) in Eq. (5) are not the Rényi maximum-entropy posteriors for the subsets S1 and S2 with means U1 and U2. For q=0.55 and β=0.1, using the value U1 ≈ 0.43 that follows from Eq. (6) and the global mean U=10, direct integration gives ∫0^1 x qS1(x) dx ≈ 0.49, not U1. The reason is that the parameter a=(1-q)β/q in Eq. (5) is inherited from the global problem; the true subset maximizer must have a subset-specific parameter determined by the constraint ∫x p = U1, which is generically different. Therefore the comparison of the l.h.s. and r.h.s. of Eq. (1) in Fig. 1b does not test the subset-independence axiom as the authors intend; the claimed violation is an artifact of using subset posteriors that do not satisfy the mean constraints stated in the paper.","section":"3. Subset independence axiom, Eqs. (5)-(6)"},{"comment":"The assertion that the Livesey-Skilling criterion shows that only trace-form entropies satisfy subset independence is explicitly stated to be outside the scope of this reply and deferred to a separate publication. This assertion is one of the main conclusions of the paper (item (ii) in the summary), and the reply does not provide the promised proof or a reference to an existing proof. Without this support, the paper's central claim is not established, and the conclusion should be weakened or the proof included.","section":"3. Subset independence axiom, final paragraph"},{"comment":"The criticism that JK 'are forced to change the Lagrange multipliers' is not a technical demonstration of an error. In the standard maximum-entropy formalism, Lagrange multipliers are determined by the constraints and are allowed to differ when the domain or the constraint set changes. Requiring the same β for the whole set and for subsets is a strong conditional-preservation property that is not a general requirement of the maximum-entropy method. The reply therefore does not substantively rebut JK's treatment of system independence; it merely restates a preference for a particular interpretation of how multipliers should behave.","section":"4. System independence axiom"}],"minor_comments":[{"comment":"The claim that q > 1 values are 'irrelevant' because 'the deformation parameter carries information about x' conflates the domain of the random variable with finite-size data; the normalization bound 0 < q < 1 + 1/xmax depends on the domain, not on the data set, so this argument needs clarification.","section":"1. Uniqueness axiom"},{"comment":"The notation for the q-exponential, ex_q := [1+(q-1)q^{-1}x]^{1/(q-1)}, uses x both as the function argument and as a subscript, which is confusing; consider writing exp_q(x) for clarity.","section":"3. Subset independence axiom, Eq. (5)"},{"comment":"The values of U, U1, and U2 used in the figure are not given explicitly; the text only states that β=0.1 and q=0.55 were chosen, so the reader cannot reproduce the plot without additional computation.","section":"3. Subset independence axiom, Fig. 1"}],"recommendation":"major_revision","confidential_remarks":"The main technical error in Eq. (5) is concrete and can in principle be fixed by deriving the correct subset maximizers, so I am not recommending rejection at this stage. If the corrected calculation still shows a violation of Eq. (1), the reply may be publishable after the Livesey-Skilling claim is either proved, properly referenced, or downgraded to a conjecture. If, however, the correct subset maximizers satisfy Eq. (1) under the authors' own criterion, then the central claim collapses and the paper should be rejected. The editor may also wish to consider whether a short reply is the right venue for introducing a new proof of the Livesey-Skilling criterion, which the authors themselves defer."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this reply is worth reading if you're following the Rényi-versus-Shannon MaxEnt debate, but its new centerpiece does not work. The continuous example meant to show that Rényi entropy violates Shore-Johnson subset independence is constructed with the wrong subset posteriors.\n\nWhat's actually new: the paper isolates JK's Lagrange-multiplier and composition-rule moves and objects to them. The observation that JK modify the multipliers to get their equivalence is fair, and the point that their composition rule makes the maximization redundant is also worth noting. These are legitimate criticisms, though they are extensions of the authors' earlier PRE work rather than fresh results.\n\nThe soft spot is load-bearing. In Eq. (5), the authors write q_S1 with the same beta as the global distribution and claim U_1 is its mean. It isn't. For their own numbers (q=0.55, beta=0.1, U=10), the global posterior gives m(S1) ≈ 0.365 and U_1 ≈ 0.43. Plugging U_1 into their q_S1 formula gives a distribution whose actual mean is about 0.48. So the q_S1 they test against Eq. (1) is not the Rényi MaxEnt posterior for the subset with mean U_1; it's just a member of the same family with the wrong parameter. The same issue afflicts q_S2. That means the violation they plot is an artifact of using distributions that don't satisfy the subset constraints. The conditional q_D/m(S1) would have the right mean and would be a legitimate candidate for the subset posterior, though whether it is the true MaxEnt posterior is itself the question. As it stands, the demonstration doesn't establish the violation.\n\nThe other announced support, the Livesey-Skilling criterion that only trace-form entropies satisfy subset independence, is asserted without proof here. That may be true, but the reader has to take it on faith.\n\nWho gets value: people tracking the SJ-axioms literature. The reply usefully catalogues JK's moves, but the numerical core is not sound. It should be reviewed because the debate matters and the flaws are correctable in principle, but a referee would need to insist on honest subset maximizers before publication.\n\nI would not cite this version.\n\nRegards.","headline":"The reply's new continuous example is built on incorrect subset posteriors—the q_S1 in Eq. (5) doesn't have the mean U_1 it claims—so the central demonstration of Rényi subset-independence violation does not hold up.","tokens_in":6212,"tokens_out":18019,"would_cite":false,"duration_ms":166360,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["05.20.-y","02.50.Tt","89.70.Cf"],"model":"deepseek-v4-flash","headline":"Rényi entropy violates the subset-independence axiom in maximum-entropy inference.","keywords":["Rényi entropy","maximum entropy inference","Shore-Johnson axioms","subset independence","artificial bias","Tsallis entropy","q-exponential distribution","Lagrange multipliers"],"falsifier":"Directly solve the constrained Rényi maximization for $D=[0,\\infty)$ with mean $U$ without assuming the paper's $q$-exponential form, for example by numerical optimization over a fine grid, and test whether the resulting $q_D$, $q_{S_1}$, $q_{S_2}$ satisfy $q_D=m(S_1)q_{S_1}+m(S_2)q_{S_2}$ with $m(S_i)=\\int_{S_i}q_D\\,dx$. If the true maximizers satisfy the identity, or if the coefficients turn out not to depend on $U$, the central claim is refuted. A simpler check: determine whether Eq. (5) satisfies the Euler–Lagrange stationarity condition for Rényi entropy with the linear mean constraint at the stated $q$ values.","tokens_in":5147,"feed_emoji":"📊","tokens_out":8674,"duration_ms":82821,"temperature":0.7,"pith_summary":"This reply defends the claim that using Rényi entropy as a maximum-entropy objective with linear constraints produces posterior distributions that violate the subset-independence axiom of consistent inference, meaning the inferred distribution for a subset inherits information from the whole data set that it should not. The authors exhibit a continuous two-subset example in which the Rényi maximum-entropy distributions do not satisfy the required identity, while the ordinary logarithmic entropy does. They also argue that the Comment's proposed corrections—a concavity-under-majorization criterion and a modified composition rule—are not fixes: the former misidentifies the admissible normalization range, and the latter fixes the functional form before any maximization, making entropy maximization redundant and introducing artificial bias. If the reply is right, any inference pipeline that maximizes Rényi entropy under ordinary linear constraints is unreliable for subset-consistent updating, and the standard logarithmic entropy is the unique consistent choice for such constraints.","feed_headline":"Rényi entropy violates the subset-independence axiom","feed_subtitle":"A two-subset example shows its inferred weights depend on the whole data mean, unlike the standard logarithmic entropy.","key_machinery":"The load-bearing object is the Shore–Johnson subset-independence identity $q_D(x)=m(S_1)q_{S_1}(x)+m(S_2)q_{S_2}(x)$ with $m(S_i)=\\int_{S_i}q_D(x)\\,dx$, together with the distinction between choosing an entropy before maximization (the axiom stage) and comparing entropies after maximization. The paper's explicit check uses the continuous MaxEnt posteriors: ordinary exponentials for the logarithmic entropy, and $q$-exponentials $e_x^q=[1+(q-1)q^{-1}x]^{1/(q-1)}$ for Rényi entropy. The identity fails because the Rényi mixing coefficients in Eq. (6) contain $U$, the global mean, and this $U$-dependence is the mechanism that creates artificial subset bias. The reply also uses the normalization conditions $q<1$ and $q>\\beta U/(1+\\beta U)$ (and analogously for subset means) to reject $q>1$ values admitted by the Comment's concavity-under-majorization criterion.","core_discovery":"The central claim is a violation, not a mathematical contradiction: when Rényi entropy is maximized under a linear mean constraint, the optimized distribution $q_D(x)$ on a set $D$ split into disjoint subsets $S_1$ and $S_2$ cannot be written as $q_D(x)=m(S_1)q_{S_1}(x)+m(S_2)q_{S_2}(x)$, where $m(S_i)$ is the probability mass the full optimum assigns to $S_i$. The paper computes the Rényi posteriors for $D=[0,\\infty)$, $S_1=[0,1)$, $S_2=[1,\\infty)$ in terms of $q$-exponentials, and finds that the coefficients $m(S_1)$ and $m(S_2)$ depend on the mean $U$ of the whole set $D$, not on the means of the subsets. Substituting the expressions into the axiom equation fails numerically at representative parameters ($\\beta=0.1$, $q=0.55$), whereas the analogous logarithmic-entropy posteriors satisfy the identity exactly. The reply reads the Shore–Johnson axioms as pre-maximization requirements: they constrain which functional may be maximized in the first place, so an equivalence that holds only after maximization, as in the Comment, is not sufficient.","pith_inferences":["The same two-subset test could be run as a general diagnostic on any candidate entropy: compute the true maximizers and compare $m(S_1)q_{S_1}+m(S_2)q_{S_2}$ with $q_D$. This would convert the paper's single example into a family of falsifiable checks for deformed entropies.","Because the paper finds that mixing coefficients carry the global mean $U$, a practical finite-sample signature is that Rényi-maximum-entropy fits will shift when the sampled range expands, even if the underlying distribution is unchanged; the paper argues this conceptually but does not present sample-size simulations.","The reply's pre-maximization reading of the axioms suggests a general principle: two entropies that are monotonically related can receive different consistency verdicts, so ordering entropies by information content alone does not settle which one may be maximized for inference."],"forward_implications":["Under linear mean constraints, Rényi-entropy maximum-entropy inference cannot be used for subset-consistent updating: the posterior for a part of the space depends on the global mean of the whole space.","The ordinary logarithmic (trace-form) entropy satisfies the subset-independence identity in the same example, so the paper supports the conclusion that for linear constraints it is the unique consistent choice.","The Comment's modified composition rule $g(p_{ij})=g(u_i)g(v_j)$ forces $g(x)\\sim x^q$, so the distribution is fixed before maximization; this makes the entropy maximization procedure redundant rather than a derivation.","The admissible $q>1$ range for Rényi maximizers should be discarded because $q$ then encodes the data interval $x_{\\max}$, making the deformation parameter data-dependent and not independent of the mathematical variable $x$.","Because Rényi entropy is monotonically related to Tsallis entropy yet only Rényi violates both subset and system independence under linear constraints, monotone equivalence does not preserve consistency properties."],"supporting_citations":[{"why":"The Comment under reply; supplies the concavity-under-majorization criterion, the subset-independence argument, and the modified composition rule that the reply challenges.","marker":"[1]"},{"why":"The authors' prior paper claiming Rényi entropy violates Shore–Johnson axioms; the reply restates and defends it against the Comment.","marker":"[2]"},{"why":"The original statement of the subset- and system-independence axioms and the criterion in Eq. (1) that the reply checks.","marker":"[3]"},{"why":"The Comment's modified axioms with the composition rule $g(p_{ij})=g(u_i)g(v_j)$; the redundancy argument targets this move.","marker":"[5]"},{"why":"Source for the pre-maximization reading of the axioms, used to reject post-maximization equivalence arguments.","marker":"[6]"},{"why":"Support for the claim that generalized $q$-product composition rules introduce biases not present in the data.","marker":"[7]"},{"why":"Result used to show the modified composition rule forces $g(x)\\sim x^q$, making maximization redundant.","marker":"[12]"}],"fun_headline_variants":["Rényi entropy breaks subset independence","Subset independence fails for Rényi entropy","Rényi inference biased by global mean","Rényi entropy violates Shore-Johnson axiom","Reply: Rényi inference is subset-dependent"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the $q$-exponential expressions in Eq. (5) are the true Rényi maximum-entropy distributions under the linear mean constraint and the stated normalization conditions; if another distribution maximizes Rényi entropy for those constraints, the $U$-dependence of the computed coefficients would not establish a violation of subset independence.","fun_headline_variants_meta":{"raw":{"variants":["Rényi entropy breaks subset independence","Subset independence fails for Rényi entropy","Rényi inference biased by global mean","Rényi entropy violates Shore-Johnson axiom","Reply: Rényi inference is subset-dependent"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1524,"prompt_tokens":967,"completion_tokens":557,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":488}},"tokens_in":583,"tokens_out":557,"duration_ms":5735,"temperature":1.0,"reasoning_tokens":488,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:56:31.987336+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Directly solve the constrained Rényi maximization for $D=[0,\\infty)$ with mean $U$ without assuming the paper's $q$-exponential form, for example by numerical optimization over a fine grid, and test whether the resulting $q_D$, $q_{S_1}$, $q_{S_2}$ satisfy $q_D=m(S_1)q_{S_1}+m(S_2)q_{S_2}$ with $m(S_i)=\\int_{S_i}q_D\\,dx$. If the true maximizers satisfy the identity, or if the coefficients turn out not to depend on $U$, the central claim is refuted. A simpler check: determine whether Eq. (5) satisfies the Euler–Lagrange stationarity condition for Rényi entropy with the linear mean constraint at the stated $q$ values.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Comment under reply; supplies the concavity-under-majorization criterion, the subset-independence argument, and the modified composition rule that the reply challenges."},{"cited_title":"Oikonomou and G.B","cited_arxiv_id":null,"evidence_quote":"The authors' prior paper claiming Rényi entropy violates Shore–Johnson axioms; the reply restates and defends it against the Comment."},{"cited_title":"The crux of their argument can be traced back to the idea that the maximization of f (∑ i g(pi)) should yield the same result as ∑ i g(pi)","cited_arxiv_id":null,"evidence_quote":"The original statement of the subset- and system-independence axioms and the criterion in Eq. (1) that the reply checks."},{"cited_title":"( 1) we verify that the latter is not satisfied","cited_arxiv_id":null,"evidence_quote":"The Comment's modified axioms with the composition rule $g(p_{ij})=g(u_i)g(v_j)$; the redundancy argument targets this move."},{"cited_title":"Comment on \"R\\'enyi entropy yields artifficial biases not in the data and incorrect updating due to the infinite-size data\"","cited_arxiv_id":"1905.00729","evidence_quote":"Source for the pre-maximization reading of the axioms, used to reject post-maximization equivalence arguments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Support for the claim that generalized $q$-product composition rules introduce biases not present in the data."},{"cited_title":"Karabulut, Eur","cited_arxiv_id":null,"evidence_quote":"Result used to show the modified composition rule forces $g(x)\\sim x^q$, making maximization redundant."}],"review_version":1}