{"id":"0094097e-96ea-40ea-ac97-8fb53b675548","arxiv_id":"2607.17232","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"This paper is a tutorial that assembles methods, mostly from the authors' prior work, for computing rate–distortion–perception trade-offs under f-divergence, α-divergence, Wasserstein, and perfect-realism constraints.","lead":"This is a tutorial on computing the rate–distortion–perception function (RDPF), which extends classic lossy-compression theory with a third constraint: reconstructions should look like real signals. It collects optimization and estimation methods for discrete, Gaussian, and general continuous sources, but most of the machinery comes from the authors' own prior papers.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RAM convergence (Thm 3) rests on an unconstructed auxiliary functional v; the hypothesis lim ||h^(n+1) − v^(n)|| = 0 at linear rate is neither instantiated nor verified, so the claimed general-purpose computability of the RDPF is not established.","rationale":"The reader's weakest-assumption analysis identifies the RAM convergence condition in Theorem 3, and my reading confirms this is the most load-bearing weakness in the computational claim. The central assertion of the tutorial is that the RDPF in Eq. (1) can be computed by general-purpose tools. NAM is restricted to twice-differentiable divergences and its inner Newton loop also lacks a proven global-convergence guarantee, but the paper is upfront about NAM's smoothness requirement. RAM is presented as the method that removes this restriction and handles nondifferentiable perception metrics such as TV, making its convergence theorem central to the tutorial's reach. However, Theorem 3's 'provided that' condition is not constructive: the auxiliary functional v is unrestricted in type, and no choice satisfying the linear-rate condition is exhibited. In fact, the condition is close to circular: it demands that v asymptotically match the implicit OAM update, which is precisely the hard part that RAM was introduced to circumvent. The numerical evidence in Figs. 2–3 is not accompanied by code, data, or a statement of the v used, so it cannot be checked. This does not warrant rejection of the tutorial: the paper explicitly labels itself a tutorial, attributes results to prior work [28,33,39], and candidly notes RAM's restricted multiplier range and SLB looseness. The reader's CONDITIONAL verdict is therefore appropriate. My concern does not move the verdict; it reinforces the condition that the computational claims, especially RAM, require either a constructive v with a proven linear-rate condition or explicit empirical verification of that condition.","tokens_in":25434,"tokens_out":8122,"duration_ms":78710,"concrete_test":"Extract from [28, Algorithm 2] (or the description for Fig. 2) the concrete auxiliary functional v used to generate the RAM curve in the Bernoulli(0.15)/Hamming/KL example. Then run that RAM iteration and compute δ_n = ||h^{(n+1)} − v^{(n)}|| and its successive ratios δ_{n+1}/δ_n. If no v is specified, or if for the specified v the ratios do not remain bounded below 1 in the limit, Theorem 3's linear-rate hypothesis is not satisfied and the claimed RAM convergence to R(D_s,P_s) is not established by the theorem.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central computational claim is that the RDPF can be computed by the NAM and RAM schemes for discrete sources with f-divergence constraints. Theorem 3 (RAM) is the main mechanism for nondifferentiable divergences such as TV, but its convergence guarantee is conditional on a hypothesis that is never operationalized. The theorem states that v[·] may be 'any functional that maps a probability distribution to a probability distribution,' and then requires that lim_{n→∞} ||h^{(n+1)} − v^{(n)}|| = 0 with at least a linear rate. No concrete v is constructed, and no general method for finding one is given. Moreover, the notation is ambiguous: h^{(n+1)} is not defined inside Theorem 3 (only the hatted \\h iterates are), so the reader cannot tell whether the condition involves the OAM update h^{(n+1)} from Theorem 1 or the RAM iterate \\hat{h}^{(n+1)}. Either way, the condition is essentially that v asymptotically predicts the exact OAM update, which is defined implicitly through the fixed-point equation (8). This is not a minor technical gap: without a constructive v that provably satisfies the linear-rate condition, Theorem 3 does not by itself establish that any implementable RAM iteration converges to R(D_s,P_s). The numerical figures (Figs. 2–3) are presented without code or a statement of the v used, so they do not fill the gap. The paper does honestly note in Remark 1 and the TV subsection that RAM restricts the feasible Lagrange-multiplier range and may not recover the full curve, which mitigates the overclaim, but the core 'guaranteed convergence' assertion in Theorem 3 remains unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This tutorial surveys rate-distortion-perception (RDP) theory, centered on the RDPF of Blau and Michaeli as defined in Eq. (1). It reviews coding theorems and then presents optimization methods for computing the RDPF: Newton-based alternating minimization (NAM) and relaxed alternating minimization (RAM) for finite alphabets with f-divergence perception constraints, Gauss-Seidel/water-filling methods for Gaussian sources under alpha-divergence and squared Wasserstein-2 constraints, and a copula-based method for continuous sources in the perfect-realism regime. The paper also discusses a Shannon lower bound for the perfect-realism RDPF and outlines future research directions. The central claim is that Eq. (1) is computable by these general-purpose tools for discrete and continuous sources.","tokens_in":25820,"tokens_out":12730,"duration_ms":109503,"significance":"If the presented algorithms are correct, the survey would be a useful reference, consolidating computational RDPF results that are currently scattered across the authors' prior papers and related literature. The paper is candid about several limitations, including Remark 1 on NAM-versus-RAM, the TV subsection's admission that RAM does not fully recover the RDP surface, and Remark 5 on the scope of the Shannon lower bound. It also properly attributes the Gaussian Wasserstein closed form to [34] and compares with recent Wasserstein-barycenter and primal-dual approaches. However, the paper does not provide proofs or code, and some of its central computational claims—especially those involving RAM—are stated with hypotheses that are not instantiated. The tutorial's value therefore depends on how carefully these conditional results are delimited.","major_comments":[{"comment":"The RAM result is the main mechanism offered for nondifferentiable perception metrics, but the convergence hypothesis is not operational. The theorem defines \\hat Q, \\hat h, and \\upsilon^{(n)}=\\upsilon[\\hat h^{(n)}], yet the condition reads lim ||h^{(n+1)}-\\upsilon^{(n)}||=0 with 'at least a linear rate'; h^{(n+1)} is undefined in the statement (only \\hat h^{(n+1)} is defined) and \\upsilon is allowed to be any distribution-to-distribution functional. Consequently the reader cannot check whether the hypothesis holds for the Bernoulli/TV examples in Figs. 2-3, and no constructive \\upsilon is given for any divergence. Because the abstract and Section III present RAM as part of general-purpose tools for computing (1), this is a load-bearing gap. Please either instantiate \\upsilon with a provable condition, or recast the RAM claims and Figs. 2-3 as conditional/heuristic.","section":"Theorem 3 (RAM), Sec. III-A"},{"comment":"The claimed closed-form linear realization \\hat X = aX + W is not explicit. In the first case a=\\max\\{1-D/\\sigma_W^2,0\\}, but \\sigma_W^2 is defined two lines below in terms of a. This is a circular definition, so the 'closed-form design parameters' are not actually closed-form. The classical RD limit suggests the intended expression is a=\\max\\{1-D/\\sigma_X^2,0\\}; if so, please correct. As printed, the complement to [34] (the realizing test channel) is not verifiable.","section":"Theorem 5, Eq. (19)"},{"comment":"The reduction of the Gaussian vector RDPF to the separable program (22) asserts that the lower bound (21) 'holds with equality without loss of generality' via simultaneous diagonalization of (A,\\Sigma_W,\\Sigma_X). Simultaneous diagonalizability of the optimal solution is not shown; the text itself notes the inequality is strict beyond i.i.d. vectors. Since (22) underlies Theorem 6 and the Gauss-Seidel algorithm, either provide a proof or state precisely the condition under which equality is guaranteed, with a specific citation.","section":"Sec. III-B, Eqs. (20)-(22)"}],"minor_comments":[{"comment":"Convention (i): 'f(0)=lim_{x->0+} f(0)' should read 'f(0)=lim_{x->0+} f(x)'.","section":"Definition 2"},{"comment":"The Lagrangian uses both h and q_{\\hat X}; define q_{\\hat X} before Eq. (3) and state that at the optimum h = q_{\\hat X}, otherwise the notation is confusing.","section":"Lemma 1, Eq. (3)"},{"comment":"If the unadorned h in the convergence hypothesis is the OAM update from Theorem 1, say so explicitly and specify the norm; if it is \\hat h, use consistent notation. Also define what 'at least a linear rate' means (linear in n, in which norm).","section":"Theorem 3"},{"comment":"The RAM row lists 'Convergence Rate: O(log(1/epsilon))' and 'Total Complexity: O(log(1/epsilon))'. A rate is not a number of iterations; as printed this is dimensionally confused and omits dependence on the alphabet size and the cost of evaluating A[upsilon].","section":"Table II"},{"comment":"The sentence claiming that RAM 'still guarantees convergence to a globally optimal solution' overstates the conditional Theorem 3. It should be qualified with 'provided the hypothesis of Theorem 3 holds.'","section":"Sec. III-A, before Theorem 3"},{"comment":"The captions do not state which RAM auxiliary functional v was used in the numerical studies. For a paper whose message is computability, this makes the figures hard to reproduce; please state the choice of v or describe the results as illustrative.","section":"Figs. 2-3"},{"comment":"The text refers to a 'bivariate exponential source with zero mean'. A genuine exponential marginal cannot have zero mean; specify the mean-shifted construction or change the wording.","section":"Sec. III-B, Fig. 8"}],"recommendation":"major_revision","confidential_remarks":"To the editor: this is a tutorial that largely restates the authors' own prior work, which is acceptable if the claims are properly delimited. The main risk is that the abstract and Section III overstate the RAM method: Theorem 3's convergence guarantee depends on an unconstructed auxiliary functional, and the notation in the hypothesis is ambiguous. I would support publication after the authors either supply an explicit RAM auxiliary functional or firmly demote the RAM part to a conditional/heuristic discussion, and after fixing the circularity in Eq. (19)."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: this is a tutorial, not a research paper. It restates the RDPF definition of Blau and Michaeli and then compiles the authors' prior work (plus a few other results) on computing it. The tutorial frame is the only real novelty, and the title's \"Redefining the Fundamental Limits\" oversells what is a synthesis. That said, for someone looking to get oriented in the RDP computation literature, this is a clean, honest map.\n\nWhat it does well: The organization is genuinely useful. The coding-theorem overview in Section II gives the right context, and the split between discrete f-divergence methods (NAM/RAM) and continuous Gaussian/copula methods is sensible. The paper is candid about RAM's restricted multiplier range and the Shannon lower bound's looseness in the perfect-realism regime. The small constructive addition in Theorem 5 — the linear realization parameters for the Wasserstein-2 closed form attributed to [34] — is real and useful.\n\nWhere the soft spots are: The main one is, as your stress-test note says, Theorem 3. The RAM convergence condition requires lim ||h^{(n+1)} - v^{(n)}|| = 0 with a linear rate, for an auxiliary functional v that is never instantiated. The notation is also ambiguous: h^{(n+1)} isn't defined inside the theorem, only \\hat h iterates are. Without a constructive v that provably satisfies this condition, Theorem 3 doesn't establish that any implementable RAM iteration converges to the RDPF. The figures don't fill that gap because no code or v selection is given. This is a real flaw, though it is partly mitigated by the paper's own Remark 1 and the TV discussion, which acknowledge that RAM may not cover the full tradeoff curve.\n\nA secondary issue: the computational claims are stated without proofs here, relying on self-cited theorems from [28,32,33,39]. That's fine for a tutorial, but it means the reader can't verify the claims without pulling the originals. Also, the 'Redefining the Fundamental Limits' phrasing in the title sets expectations the paper doesn't meet.\n\nVerdict: If this is submitted as a tutorial/survey to a journal, it deserves a serious referee. The synthesis is valuable, the attribution is mostly accurate, and the authors are honest about known gaps. The referee should push for a fix to the RAM gap (either instantiate v for the worked examples or state the theorem with the precondition more sharply) and for calmer title language. For a reading group, I'd bring it in; I'd cite it as a survey reference, though I'd still point to the originals for the proofs.","headline":"A competent tutorial that repackages the authors' own results; the title oversells it and the RAM convergence condition is not constructively instantiated, but as a survey it deserves referee time.","tokens_in":26321,"tokens_out":2124,"would_cite":true,"duration_ms":20377,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A34","94A17","62B10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The rate-distortion-perception function can be computed, not just defined from first principles.","keywords":["rate-distortion-perception function","lossy compression","perceptual fidelity","f-divergence","alternating minimization","Gaussian sources","perfect realism","copula methods"],"falsifier":"Run the relaxed alternating-minimization scheme on a Bernoulli source with Hamming distortion and total-variation perception, using the paper's own smooth approximation as a benchmark. If the relaxed scheme's limit misses the analytic RDPF in a regime where the linear-rate condition provably holds, the central convergence claim would be settled. A second observation: if a divergence satisfies the linear-rate condition but the scheme converges to different limits from different initial distributions, the claimed global optimality fails.","tokens_in":25291,"feed_emoji":"📉","tokens_out":6151,"duration_ms":55682,"temperature":0.7,"pith_summary":"This tutorial argues that the rate-distortion-perception function (RDPF), which gives the minimum bit rate for a lossy code whose reconstructions are both close to the source and statistically similar to it, is a computable object. For discrete sources with f-divergence perception constraints, it assembles alternating-minimization algorithms, including a Newton-based variant and a relaxed variant, and claims their iterates converge to the RDPF. For Gaussian sources, it gives closed-form expressions under alpha-divergence and squared optimal-transport perception, plus a water-filling interpretation. For continuous non-Gaussian sources under perfect realism, it develops a copula-based convex program that can be solved numerically. The reader should care because these algorithms turn a three-way trade-off central to modern generative compression into concrete numerical curves that can be measured, benchmarked, and designed against.","feed_headline":"Rate-distortion-perception limits are computable, not just theoretical","feed_subtitle":"For generative codecs, that means the ideal rate-perception-distortion trade-off can be computed, not guessed.","key_machinery":"The workhorse is the Lagrangian double-minimization reformulation of the RDPF: minimizing mutual information is replaced by minimizing a KL divergence plus distortion and perception penalty terms over a conditional distribution Q and an auxiliary reconstruction distribution h. The optimal h for fixed Q is the induced output distribution; the optimal Q for fixed h is a Boltzmann-like kernel weighted by an exponent involving the distortion and a function of the divergence. Because the h-update is implicit, the paper's two algorithms resolve it either by Jacobian-based root-finding (the NAM scheme) or by substituting an auxiliary functional (the RAM scheme). For Gaussian sources, the same machi","core_discovery":"The central claim is that the RDPF, as defined in (1), can be solved by alternating minimization. For finite-alphabet sources with f-divergence perception, the paper shows that an optimal alternating-minimization update converges to a point on the RDPF, and since the update is implicit, it develops two implementable variants: a root-finding iteration that uses the Jacobian of the update map and requires twice-differentiable divergences, and a relaxed iteration that introduces an auxiliary functional to make the update explicit while preserving global convergence under a linear-rate condition. For Gaussian sources, the paper derives parametric closed-form solutions for alpha-divergence and sq","pith_inferences":["If the RDPF becomes routinely computable, perception-aware codec design can shift from heuristics to targeting a computed rate-perception-distortion frontier; one could benchmark a generative codec by its distance to this surface.","The relaxed scheme's reliance on an auxiliary functional is a gap: a constructive rule for choosing that functional would make the algorithm fully general, and one could test whether running a few inner root-finding steps to define the auxiliary map yields the required linear-rate convergence.","The copula formulation suggests a natural estimator for the RDPF under non-tensorizable distortion measures, since the projection is expressed on the copula space rather than on the original signal space, potentially connecting to entropic optimal-transport algorithms in neural compression.","The Gaussian closed forms show that perfect-realism coding allocates rate to preserve marginal laws rather than only to reduce squared error, a design principle that could be tested in neural codecs by measuring per-component rate allocation against the water-filling prediction."],"forward_implications":["If the convergence theorems hold, the RDPF for discrete sources under any smooth f-divergence can be computed pointwise by the Newton-based scheme with exponential convergence, and by the relaxed scheme without smoothness requirements, for the multiplier range where the linear-rate condition holds.","For Gaussian sources, the closed-form RDPF expressions under alpha-divergence and squared optimal-transport perception imply that optimal encoders take linear realizations with closed-form parameters, connecting information-theoretic limits to lattice-based coding.","In the perfect-realism regime, the Gaussian water-filling allocation changes from a single global water level to component-dependent levels, so every source component must remain represented even at low rates.","The copula-based estimator provides a numerical route to the perfect-realism RDPF for non-Gaussian continuous sources, complementing the entropy-type lower bound that is tight only in the low-distortion regime.","The paper's comparison with barycenter and primal-dual methods suggests these approaches cover complementary regimes of the RDPF landscape rather than competing for the same problems."],"fun_headline_variants":["RDPF computed: alternating minimization unlocks perceptual compression limits","Alternating minimization makes RDPF computable in practice","RDPF solved: a recipe for computing rate-perception limits","Now you can compute the rate-distortion-perception trade-off"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The relaxed alternating-minimization convergence theorem requires an auxiliary functional that predicts the implicit next iterate closely enough to converge at a linear rate, and the paper gives no constructive recipe for such a functional; if no such choice exists for a given divergence, the scheme's global-convergence guarantee collapses.","fun_headline_variants_meta":{"raw":{"variants":["RDPF computed: alternating minimization unlocks perceptual compression limits","Alternating minimization makes RDPF computable in practice","RDPF solved: a recipe for computing rate-perception limits","Now you can compute the rate-distortion-perception trade-off"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001568,"raw_usage":{"total_tokens":6130,"prompt_tokens":806,"completion_tokens":5324,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":5266}},"tokens_in":550,"tokens_out":5324,"duration_ms":33559,"temperature":1.0,"reasoning_tokens":5266,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T18:37:42.211342+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the relaxed alternating-minimization scheme on a Bernoulli source with Hamming distortion and total-variation perception, using the paper's own smooth approximation as a benchmark. If the relaxed scheme's limit misses the analytic RDPF in a regime where the linear-rate condition provably holds, the central convergence claim would be settled. A second observation: if a divergence satisfies the linear-rate condition but the scheme converges to different limits from different initial distributions, the claimed global optimality fails.","supporting_citations":[],"review_version":1}