{"id":"9347b8f7-36d3-48d8-b71f-79687dce9f7a","arxiv_id":"2501.13736","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Discrete layered entropy is a piecewise-linear lower bound on Shannon entropy with an exact conditioning property, yielding improved strong functional representation bounds.","lead":"This paper introduces a new entropy-like measure, discrete layered entropy, that stays within a logarithmic gap of Shannon entropy and has an exact conditioning rule. Using it, the author proves a tighter strong functional representation lemma, improving a known channel simulation bound from I+log(I+2)+2 to I+log(I+3.4)+1.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 14 rests on Appendix J's inequality (49), whose proof is an unshipped SymPy/Sturm computation plus a self-referential 'same arguments as Appendix J' step; until that certificate is independently reproducible, the I+1.29 bound is conditional.","rationale":"The reader's verdict is already CONDITIONAL, and the weakest assumption identified there is exactly the same load-bearing concern: inequality (49) and the verifiability of Appendix J. I agree with that assessment. The rest of the paper has real independent value: the definition and basic properties of Lambda are proved directly, the conditioning property is derived, and several applications follow from those facts. The improvement over the previous SFRL constant, however, depends on a single numerical inequality that is only certified by unshipped exact-arithmetic computations and by a self-referential sentence inside Appendix J. A failure of (49) would invalidate the claimed I+1.29 and I+log(I+3.4)+1 bounds, so the result should not be accepted as fully verified without a reproducible certificate. Since the gap is fixable and there is no evidence of falsehood, CONDITIONAL remains the appropriate disposition, and no change to the reader's verdict is needed.","tokens_in":11,"tokens_out":8371,"duration_ms":136960,"concrete_test":"Obtain or independently re-implement the exact-arithmetic certificate for (49). Build the rational upper bound (54) with m=18, k=5 and the lower bound for Lambda(Geom(1/2))+(t-1)g'(1) using (51)/(52) with m=20, k=8, then use Sturm sequences over [0.3,0.975] and [1.025,4] to certify strict positivity of the difference. Repeat for t<0.3 and for the concavity check on [0.975,1.025] with k=14, m=70. If the author's script is not released, evaluate (49) on a dense high-precision grid and at interval endpoints; a violation would be a counterexample, while agreement only makes the claim plausible, not certified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim Theorem 14 depends on inequality (49): g(t) <= Lambda(Geom(1/2)) + (t-1)g'(1) for all t>0. If this fails, the step at (47) collapses and with it the headline bounds (26) and (27). The Appendix J proof of (49) is not self-contained or auditable. Case 1 is delegated to a SymPy/Sturm exact-arithmetic check with parameters m=18, k=5 for the upper bound and m=20, k=8 for the derivative lower bound, but the verification script is not shipped. Case 3 says the same check is done 'in a similar manner', again without shipping the certificate. Case 4 asserts d2g/dt2 <= -0.013 on [0.975,1.025] using another unshipped Sturm computation with k=14, m=70. In addition, the proof chain contains the literal phrase 'By the same arguments as Appendix J' inside Appendix J itself, so the derivation of Lambda(Y|S) <= ... is circular as printed. The tail sum used in (52) is stated as -2^{-m-1}(m^2+2m+2) without derivation. None of this shows the inequality is false, but the main numerical improvement rests on unaudited computational certificates, and the paper gives no way to reproduce or check them from the text alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the discrete layered entropy Λ(p), a piecewise-linear approximation of Shannon entropy, and develops its properties: concavity, Schur concavity, approximate closeness to H(X) within a logarithmic gap, a conditioning property Λ(X|Y)=Λ(X\\Y), and operational connections to one-to-one non-prefix codes and conditional compression. The central application is a strengthened strong functional representation lemma (Theorem 14): for arbitrary X,Y there exists S independent of X with H(Y|X,S)=0 and Λ(Y|S) ≤ I(X;Y)+Λ(Geom(1/2)) < I(X;Y)+1.29, yielding H(Y|S) < I(X;Y)+log(I(X;Y)+3.4)+1. The proof in Appendix J reduces the new constant to a one-variable analytic inequality (49), verified by a case analysis using Sturm sequences and exact rational arithmetic.","tokens_in":35162,"tokens_out":6219,"duration_ms":52393,"significance":"If the main bound is valid, it is a genuine improvement over the previous best SFRL constants ([3], [14], [15]) and demonstrates that discrete layered entropy is a useful technical tool for one-shot channel simulation and lossy compression. The paper also contains several elegant structural results, such as the conditioning property, the axiomatic characterization in Theorems 12–13, and the operational interpretation of Λ as the length of optimal non-prefix codes. Many of the supporting propositions are proved carefully in appendices. However, the headline numerical improvement is conditional on a computer-assisted proof that is not reproducible from the manuscript: the Sturm/SymPy certificates are not shipped, and a key step in Appendix J contains a self-reference. These issues are fixable but prevent the main theorem from being audited as printed.","major_comments":[{"comment":"The chain leading to (47) begins with the sentence \"By the same arguments as Appendix J,\" which appears inside Appendix J itself. This is a self-reference that makes the derivation of Λ(Y|S) ≤ Λ(K|S) circular as printed. The authors should replace this with a direct argument or cite a specific prior result (e.g., Propositions 3 or 10, or the Poisson functional representation lemma from [14]). This step is load-bearing because it connects the construction of K to the bound on Λ(Y|S).","section":"Appendix J, proof of Theorem 14"},{"comment":"The verification of the key inequality g(t) ≤ Λ(Geom(1/2)) + (t−1)g′(1) is delegated to Sturm-sequence checks performed with SymPy, but no code, worksheet, certificate, or explicit polynomial is provided. Since (49) is the load-bearing inequality for the I+1.29 bound and for the derived bounds (26)–(27), the main theorem is not independently auditable from the manuscript. The authors should provide the verification script or a complete certificate (e.g., sign patterns of the Sturm sequence or certified rational bounds) in a supplement or appendix.","section":"Appendix J, cases 1, 3, 4 of the proof of (49)"},{"comment":"The assertion that d²g/dt² ≤ −0.013 on [0.975,1.025] via a Sturm computation with k=14, m=70 is also unshipped; without this concavity bound, the conclusion (49) on that interval does not follow. Additionally, the tail sum in (52), −2^{−m−1}(m²+2m+2), is stated without derivation. It is used to lower-bound g′(1) for t>1, so it should be derived (or a reference given) for the proof to be self-contained.","section":"Appendix J, case 4 and Eq. (52)"}],"minor_comments":[{"comment":"In the proof of Theorem 15, the displayed bound \"H(Y|X) ≤ I + log(I + 3.4) + 0.99\" should presumably read \"H(Y|S)\"; Theorem 14 gives a bound on H(Y|S), not on H(Y|X).","section":"Theorem 15 proof"},{"comment":"The text refers to \"Proposition 9\" when defining Λ(X) via Shannon entropy; the intended reference appears to be Theorem 9.","section":"Section V-C"},{"comment":"The rational upper bound in (54) is asserted to follow from the continued fraction bound (53), but the algebra is not shown; a brief derivation would help the reader verify that the case analysis indeed covers the stated interval.","section":"Appendix J, Eq. (54)"}],"recommendation":"major_revision","confidential_remarks":"The central framework and many of the secondary results are strong and would be a good fit for the journal, but the main SFRL improvement cannot be accepted until the computational proof of (49) is made auditable and the self-referential step in Appendix J is repaired. These are local but load-bearing issues, so major revision seems appropriate rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is real: discrete layered entropy Λ is a genuinely useful object, with a clean conditioning property, axiomatic characterization, and a linear programming formulation. The application to the strong functional representation lemma gives the best known constant, I + log(I+3.4)+1, improving on the I+log(I+2)+2 bound from 2024. That is worth a referee's time even before the proof is polished.\n\nThe paper does most of the work carefully. The alternative definitions in Proposition 2, the gap to Shannon entropy in Proposition 5, and the operational interpretation through non-prefix codes are all clear and extend prior work by the same author on continuous layered entropy. Theorems 12 and 13 are elegant: conditioning essentially pins down Λ.\n\nThe soft spot is exactly where the reader's report puts it. Theorem 14 rests on inequality (49), and the proof of (49) is not auditable as printed. Appendix J contains a step that literally says 'By the same arguments as Appendix J' — that is circular as written, and presumably should refer to an earlier appendix or to an omitted argument. The key bound is then verified by Sturm sequences and exact rational arithmetic in SymPy, but the code is not shipped, so an independent referee cannot check cases 1, 3, and 4 without reimplementing the computation. Case 4's assertion that the second derivative is ≤ −0.013 on [0.975, 1.025] is especially delicate because equality in (49) at t = 1 leaves no slack. Also, the tail sum in (52) is stated without derivation.\n\nNone of this suggests the result is false. The simple Rényi-layer bound already yields the log 3 constant, so the improvement is in the right direction, and the plots are believable. The issues are fixable: ship the SymPy certificates, replace the self-reference, and expand the tail-sum calculation. The central argument is not a fit; the parameters m, k, η are choices made to optimize a valid inequality.\n\nMy recommendation: send it to peer review. A competent referee should be able to verify (49) with the supplied certificates, and the rest of the paper is solid. I would cite the definition and the conditioning property regardless of what happens to the SFRL constant. For a reading group, it is a good paper to discuss, though the appendix proof is not a pleasant read.","headline":"A genuinely new information measure with a likely-correct but under-verified improvement to the strong functional representation lemma.","tokens_in":35622,"tokens_out":2577,"would_cite":true,"duration_ms":23419,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A17","94A29","60E15"],"pacs":[],"model":"deepseek-v4-flash","headline":"For any pair of random variables, a common randomness variable $S$ exists that recovers $Y$ from $X$ and $S$ while adding fewer than 1.29 bits to the mutual information.","keywords":["discrete layered entropy","strong functional representation lemma","one-shot channel simulation","conditional compression","non-prefix codes","Shannon entropy approximation","maximum entropy","information theory"],"falsifier":"Use exact rational or high-precision interval arithmetic to evaluate $g(t)=\\Lambda(\\mathrm{Geom}(t/(t+1)))+\\log t$ for all $t>0$ and test whether $g(t)\\le\\Lambda(\\mathrm{Geom}(1/2))+(t-1)g'(1)$ always holds; a counterexample would refute the main theorem. Alternatively, search over small finite alphabets for $X,Y$ and every candidate $S$ independent of $X$ with $H(Y|X,S)=0$, and check whether $\\Lambda(Y|S)-I(X;Y)$ is always at least $1.29$; exceeding that for every $S$ would falsify Theorem 14.","tokens_in":34610,"feed_emoji":"📡","tokens_out":11982,"duration_ms":90069,"temperature":0.7,"pith_summary":"This paper introduces the discrete layered entropy $\\Lambda$, defined by sorting a distribution's probabilities and weighting the $i$-th largest by the difference $i\\log i-(i-1)\\log(i-1)$. $\\Lambda$ is piecewise linear, always lies between min-entropy and Shannon entropy, and approximates Shannon entropy within a logarithmic gap, which makes it useful in linear programming and maximum-entropy problems. Its key property is conditional: $\\Lambda(X\\setminus Y)=\\Lambda(X|Y)$, so $\\Lambda$ treats the conditional random variable $X|Y$ as a genuine object. The paper uses this to prove that for every pair $X,Y$ there is a common randomness $S$ independent of $X$ with $Y$ recoverable from $(X,S)$ and $\\Lambda(Y|S)\\le I(X;Y)+\\Lambda(\\mathrm{Geom}(1/2))<I(X;Y)+1.29$, giving $H(Y|S)<I(X;Y)+\\log(I(X;Y)+3.4)+1$. This is a strict improvement over the best previously known strong functional representation lemma for all values of $I(X;Y)$.","feed_headline":"A layered entropy tightens channel simulation to I + log(I + 3.4) + 1","feed_subtitle":"For any correlated pair, the new bound beats prior channel-simulation results at every mutual information value.","key_machinery":"The load-bearing object is the discrete layered entropy $\\Lambda(p)=\\sum_{i=1}^{\\infty}p^{\\downarrow}(i)(i\\log i-(i-1)\\log(i-1))$, the upper concave envelope of the conditional min-entropy. It is the unique function that satisfies the conditioning property $\\Lambda(X\\setminus Y)=\\Lambda(X|Y)$ and equals $\\log k$ for uniform $X$ over $k$ outcomes, and it is also the largest such function that never exceeds Shannon entropy. The proof of the strong functional representation lemma reduces to a one-variable analytic inequality $g(t)\\le \\Lambda(\\mathrm{Geom}(1/2))+(t-1)g'(1)$ for $t>0$, where $g(t)=\\Lambda(\\mathrm{Geom}(t/(t+1)))+\\log t$; the paper verifies this inequality by a rigorous case analysis using exact rational arithmetic and Sturm sequences.","core_discovery":"The central discovery is a strengthened strong functional representation lemma. For any (not necessarily discrete) random variables $X,Y$, there exists a random variable $S$, independent of $X$, such that $H(Y|X,S)=0$ and $\\Lambda(Y|S)\\leq I(X;Y)+\\Lambda(\\mathrm{Geom}(1/2))$, where the constant $\\Lambda(\\mathrm{Geom}(1/2))<1.29$ is the discrete layered entropy of the geometric distribution with parameter $1/2$. Consequently $H(Y|S)<I(X;Y)+\\log(I(X;Y)+3.4)+1$, improving the previous best bound $I+\\log(I+2)+2$ for every mutual information $I$. The paper also shows the optimal constant for the non-prefix channel-simulation task lies between $0.086$ and $1.29$, so the extra cost over mutual information is a genuine non-zero constant of information theory rather than an artefact of prefix-free coding.","pith_inferences":["If the $I+\\log(I+3.4)+1$ bound is nearly tight, then one-shot channel simulation is essentially characterized by mutual information plus a universal additive term, and pinning down the exact constant would close the remaining gap.","The same pattern of proving a non-prefix bound first and converting it to a prefix bound through $\\Lambda$ could be carried over to other one-shot problems where prefix-freeness is an artificial constraint, such as privacy or randomness extraction.","The three-way closeness of $\\Lambda(X|Y)$, $H(X|Y)$, and $H(X\\setminus Y)$ suggests that in one-shot analyses conditional compression and conditional entropy can be interchanged up to logarithmic corrections, which may simplify future coding proofs.","Because $\\Lambda$ is the largest conditioning-compatible underestimate of $H$, an analogous construction might yield similar approximations for Renyi or other entropies, with comparable linear-programming and simulation payoffs."],"forward_implications":["For every pair $X,Y$, one-shot channel simulation with unlimited common randomness can be done with a non-prefix description length below $I(X;Y)+1.29$ bits, and with a prefix-free description length below $I(X;Y)+\\log(I(X;Y)+3.4)+1$.","These bounds beat the previous strongest functional representation lemma for all mutual information values; the prefix bound $I+\\log(I+3.4)+1$ is tighter than $I+\\log(I+2)+2$ for every $I$.","One-shot lossy source coding with prefix codes achieves expected length at most $R(D)+\\log(R(D)+3.4)+3$, improving the earlier $R(D)+\\log(R(D)+2)+4.01$.","The discrete layered entropy can replace Shannon entropy in maximum-entropy linear programs, and the resulting optimum is within a logarithmic gap of the true entropy optimum.","The exact optimal additive constant $c_n^*$ for non-prefix channel simulation is bracketed between $0.086$ and $1.29$, meaning the gap between mutual information and simulation cost is an intrinsic non-zero constant of entropy."],"supporting_citations":[{"why":"Introduces the strong functional representation lemma and the prior construction this paper improves; it is the baseline and the framework being modified.","marker":"[3]"},{"why":"Establishes the one-shot channel-simulation setting with unlimited common randomness and the generate-then-encode strategy used here.","marker":"[13]"},{"why":"Supplies the conditional geometric representation, a geometric random variable conditionally given $(X,Y)$, on which the proof of Theorem 14 relies.","marker":"[14]"},{"why":"Gives the previous best prefix-code bound $I+\\log(I+2)+2$ that Theorem 14 is compared against and improves.","marker":"[15]"},{"why":"Provides the continued-fraction bounds used in the rigorous exact-arithmetic verification of the key analytic inequality.","marker":"[55]"}],"fun_headline_variants":["Tighter channel-simulation bound: I + log(I+3.4)+1","Channel-simulation extra cost is constant, not prefix artefact","Strong functional representation lemma tightened to I+log(I+3.4)+1","Optimal channel-simulation constant is between 0.086 and 1.29 bits","New entropy measure yields tighter channel-simulation bound"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire improvement rests on the inequality that a particular function built from the geometric distribution never rises above its tangent line at $t=1$; if even one value of $t$ violates it, the $1.29$-bit and $\\log(I+3.4)+1$ bounds do not follow, and the appendix's verification of that inequality includes a step that refers back to the same appendix.","fun_headline_variants_meta":{"raw":{"variants":["Tighter channel-simulation bound: I + log(I+3.4)+1","Channel-simulation extra cost is constant, not prefix artefact","Strong functional representation lemma tightened to I+log(I+3.4)+1","Optimal channel-simulation constant is between 0.086 and 1.29 bits","New entropy measure yields tighter channel-simulation bound"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002383,"raw_usage":{"total_tokens":9138,"prompt_tokens":881,"completion_tokens":8257,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":8159}},"tokens_in":497,"tokens_out":8257,"duration_ms":52379,"temperature":1.0,"reasoning_tokens":8159,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:38:52.119395+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use exact rational or high-precision interval arithmetic to evaluate $g(t)=\\Lambda(\\mathrm{Geom}(t/(t+1)))+\\log t$ for all $t>0$ and test whether $g(t)\\le\\Lambda(\\mathrm{Geom}(1/2))+(t-1)g'(1)$ always holds; a counterexample would refute the main theorem. Alternatively, search over small finite alphabets for $X,Y$ and every candidate $S$ independent of $X$ with $H(Y|X,S)=0$, and check whether $\\Lambda(Y|S)-I(X;Y)$ is always at least $1.29$; exceeding that for every $S$ would falsify Theorem 14.","supporting_citations":[{"cited_title":"Pointwise redundancy in one-shot lossy compression via Poisson functional representation,","cited_arxiv_id":null,"evidence_quote":"Gives the previous best prefix-code bound $I+\\log(I+2)+2$ that Theorem 14 is compared against and improves."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the continued-fraction bounds used in the rigorous exact-arithmetic verification of the key analytic inequality."}],"review_version":1}