{"id":"4f01f7ae-c273-4548-b364-5b2dd842745c","arxiv_id":"2501.04791","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Normalising flows can evaluate Bayesian partition functions, entropies, and lower-order moments of non-Gaussian posteriors, but the proposed derivative-based flow expansion fails for skewness and kurtosis.","lead":"This paper uses normalising flows, a machine learning technique, to compute Bayesian evidence, entropies, and moments of non-Gaussian posterior distributions in cosmology. It applies the method to supernova data and recovers the known matter density and dark energy constraints, while reporting that higher-order moments are inaccurate.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The flow-expansion moment formula is formally correct, but the headline claim overreaches: the single-point derivative series is not controlled by the flow training objective and demonstrably fails at fourth order.","rationale":"The reader's weakest assumption identified the smoothness and fast convergence of the flow map as the vulnerable point of Equation 18, and the reader's verdict of CONDITIONAL is supported by the order-of-magnitude errors in kurtosis in Tables 4 and 5. My stress-test agrees with that reading but sharpens it: the load-bearing gap is not only whether the Taylor series converges quickly, but that the training objective in Equation 5 never targets the local derivatives at alpha = 0 that Equation 18 uses. Therefore even a smooth, well-trained flow could in principle have accurate samples and inaccurate high-order moments, as long as the derivatives at the origin are unconstrained. The paper's own appendix flags the divergence of higher derivatives, which is an explicit admission of the limitation. The evidence and entropy results are internally consistent and agree with independent estimates, so the core numerical machinery is not called into question; what fails is the stronger claim of analytical expressions beyond the Gaussian limit. For that reason the existing CONDITIONAL verdict remains appropriate: the paper should either restrict its claims to mean and variance, or demonstrate convergence of Equation 18 on a nontrivial distribution with known moments. I do not see a basis for outright rejection, because the formal identity in Equation 15 is standard and the entropy/evidence applications are sound.","tokens_in":13265,"tokens_out":5427,"duration_ms":55479,"concrete_test":"Train a FrEIA flow on a two-dimensional non-Gaussian distribution with known analytical moments (for example a skew-normal or a mixture with specified skewness and kurtosis), then evaluate Equation 18 at truncation orders k = 2, 4, 6, 8 and compare with the exact moments. If the kurtosis does not approach the exact value as the truncation order increases, or if successive terms do not decrease, the single-point expansion is not convergent and the 'beyond the Gaussian limit' claim fails even for smooth maps. As a complementary check, retrain a smooth normalising flow architecture (for example the smooth flows of Köhler et al. 2021) on the same supernova posterior and recompute Tables 4 and 5; if the kurtosis remains order-of-magnitude off, the failure is intrinsic to the expansion itself rather than to autodifferentiation instability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that normalising flows provide analytical expressions for posterior distributions beyond the Gaussian limit rests on Equation 18, which expresses posterior moments as an infinite derivative series of the learned map f(alpha) evaluated at alpha = 0. As a Gaussian integral identity, Equation 18 is valid if f is sufficiently smooth and the series converges. But the expansion is load-bearing in a way that the paper does not secure. First, the training loss in Equation 5 only fits the pushforward density to samples; it does not constrain the derivatives of f at alpha = 0, which are precisely the coefficients used in the expansion. A flow can have excellent sample fidelity while its local Taylor coefficients are inaccurate. Second, the paper's own Tables 4 and 5 demonstrate the failure: the expansion's kurtosis is 6.2 +/- 1.5 versus the emcee ground truth 0.55 +/- 0.05 for Omega_m, and 5.2 +/- 1.3 versus 0.21 +/- 0.04 for w0, roughly an order of magnitude off. The appendix explicitly attributes this to higher derivatives of the network diverging because of the FrEIA architecture, including its permutations. Thus the abstract's promise of posterior statistics 'beyond the Gaussian limit' is not demonstrated: only mean and variance are reliable. This is not merely a numerical implementation issue; the coefficients of the single-point Taylor expansion are not trained quantities, so even a perfectly smooth flow would not automatically make Equation 18 accurate. The entropy and evidence calculations are internally consistent and agree with benchmarks, but they do not support the stronger claim in the abstract.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes combining normalising flows with Bayesian partition functions to compute evidence, entropy, and posterior moments analytically. The central device is a Gaussian integral identity (Eq. 15) that expresses the partition function as a differential operator acting on the flow map evaluated at alpha=0, leading to a series expansion (Eq. 18) for posterior moments. The authors apply the method to a supernova cosmology example (Omega_m and w0) and compare flow-based entropy, mean, variance, skewness, and kurtosis against histogram, KDE, and emcee posterior samples. They also present a Gaussian toy model for verification.","tokens_in":13581,"tokens_out":3328,"duration_ms":35546,"significance":"If the central claim were fully established, the paper would offer a useful alternative route to Bayesian evidence, information entropy, and posterior moments without additional Monte Carlo sampling. The formal Gaussian-integral identity in Eq. (15) is correct, and the entropy computation via Eq. (24) is clean and agrees with independent estimates. The toy-model validation with an analytic Gaussian target is a strength, and the code is made available. However, the headline promise of 'analytical expressions for posterior distributions beyond the Gaussian limit' is only partially supported: the flow expansion is truncated at k=4 and produces kurtosis values that are wrong by an order of magnitude in the cosmological application. The paper is honest about this limitation in its summary, but the abstract overstates what is demonstrated.","major_comments":[{"comment":"The central claim of analytical posterior moments beyond the Gaussian limit is not supported by the numerical results. In Tables 4 and 5, the flow expansion truncated at k=4 gives kurtosis values 6.2±1.5 versus 0.55±0.05 for Omega_m, and 5.2±1.3 versus 0.21±0.04 for w0, i.e. errors of roughly an order of magnitude. Kurtosis is a fourth-order quantity and is exactly the kind of non-Gaussian statistic the abstract promises. The paper should either restrict its claims to mean, variance, and covariance, or demonstrate that the expansion can be made accurate at higher order.","section":"§2.4, Eq. (18) and Tables 4/5"},{"comment":"The expansion in Eq. (18) computes moments from derivatives of the learned flow map f at alpha=0, but the training loss in Eq. (5) only constrains the pushforward density on training samples; it does not constrain the Taylor coefficients at alpha=0. The appendix itself notes that higher derivatives diverge with the FrEIA architecture because of permutations and other non-smooth components. Consequently, the truncation at k=4 is not controlled by the training objective, and the formal identity in Eq. (18) does not guarantee accurate moments for a learned flow. The authors should provide convergence evidence (e.g. comparison of truncation orders k=2,3,4) or use a provably smooth flow and show that the expansion converges.","section":"§2.4, Eq. (18) and Appendix A"},{"comment":"The validation is partially circular. The normalising flow is trained on emcee posterior samples, and the same emcee posterior is later treated as ground truth in Tables 4 and 5. Since the expansion moments are deterministic functions of the trained flow, agreement with emcee values is partly built into the training procedure. The toy model in Table 7 provides an independent check only for a Gaussian target. The supernova validation should be repeated with an independent sampler, or the comparison should split the emcee samples into disjoint training and test sets so that the ground truth is not the same data used for training.","section":"§3, Tables 2, 4, 5"}],"minor_comments":[{"comment":"Several table captions are formatted inconsistently, e.g. 'T able 1', 'T able 4'; these should be corrected to 'Table 1', 'Table 4', etc.","section":"Tables 1-7"},{"comment":"The text says 'c.p. Foreman-Mackey et al. 2013'; this should read 'cf.' as a standard abbreviation for 'confer'.","section":"§3, near Eq. (22)"},{"comment":"In Eq. (23), the sample entropy estimate uses ln p(theta_i|y), but the paper does not specify how p(theta_i|y) is evaluated for the histogram and KDE estimates; a brief remark would improve reproducibility.","section":"§2.3, Eq. (23)"},{"comment":"The statement that the Fisher information matrix F is the identity is tied to the standard normal choice; if a non-unit covariance is used, the resulting linear transform should be spelled out for clarity.","section":"§2.2, near Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is transparent about its limitations, which I appreciate. The main issue is the mismatch between the abstract's promise of analytical non-Gaussian posterior statistics and the demonstrated performance, which is reliable only up to second-order moments. The formal derivation is correct, and the entropy and partition-function computations are solid. I believe the paper could be made publishable by reframing the claims as a proof-of-concept for low-order moments and evidence, and by adding convergence checks or using a smoother flow architecture to address the higher-order failures."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has one genuinely useful idea — using the differentiability of a trained normalising flow to evaluate posterior moments and entropy through a Gaussian-integral identity — and it validates that idea for entropy, mean, and variance. But the headline claim about analytical expressions 'beyond the Gaussian limit' is not supported: the expansion's own results for skewness and kurtosis are off by an order of magnitude, and the paper says so in the appendix. The reader's conditional verdict is about right.\n\nWhat is new: Equation 15 is a classical Gaussian integral identity; the application to flow-learned maps is the contribution. The entropy computation via change of variables (Eq. 24) is clean and agrees with histogram/KDE and the toy-model analytic value. Mean, variance, and covariance from the expansion match emcee ground truth. The toy model check is a good discipline. The partition-function plots vs KDE are convincing.\n\nWhere it goes soft: the abstract promises posterior distributions beyond Gaussian, but Tables 4 and 5 show kurtosis from the expansion is 6.2 vs 0.55 (Omega_m) and 5.2 vs 0.21 (w0). The stress-test note is correct: the training loss (Eq. 5) fits the pushforward density; it does not constrain the single-point Taylor coefficients that the expansion uses. So the failure is not just a numerical artifact of FrEIA; it is structural to the method as presented, though a smoother architecture could plausibly fix it. The appendix candidly blames higher-derivative divergence, so the authors are not hiding it. The paper would be stronger if the abstract and summary claimed reliable mean/variance plus a promising but unproven route to higher cumulants. Also, the 'available on GitHub' statement has no URL, and the uniform prior bounds are not specified in the text, which makes exact reproduction harder. These are minor but real.\n\nWho it is for: cosmologists using normalising flows for evidence and moments; readers interested in partition-function methods. It does not deserve a desk reject, but it needs a revision that aligns claims with demonstrated accuracy. I would send it to a referee, expecting moderate revision.","headline":"A clean entropy and mean/variance pipeline for normalising flows, but the advertised non-Gaussian moments do not survive contact with the paper's own tables.","tokens_in":14110,"tokens_out":1798,"would_cite":false,"duration_ms":18263,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["98.80.-k","02.50.-r"],"model":"deepseek-v4-flash","headline":"Normalising flows compute Bayesian partition functions by differentiating a learned map at one point.","keywords":["normalising flows","Bayesian partition function","Bayesian evidence","non-Gaussian posterior","information entropy","supernova cosmology","cumulants","MCMC"],"falsifier":"Train a normalising flow on a smooth, heavy-tailed one-dimensional distribution whose skewness and kurtosis are known analytically, then compute these moments from the flow expansion terminated at $k=4$; if the expansion values differ from the true values by more than the sampling noise, the claim that posterior moments follow from local derivatives of the flow is refuted for that class of distributions. The paper's Tables 4 and 5 already indicate such a failure for the supernova posterior kurtosis.","tokens_in":13070,"feed_emoji":"🌌","tokens_out":8230,"duration_ms":72345,"temperature":0.7,"pith_summary":"This paper argues that normalising flows—invertible, differentiable maps learned between a standard Gaussian and a non-Gaussian posterior—can serve as an analytical toolkit for Bayesian inference, not just a sampler. The key move is to rewrite the Bayesian partition function (the evidence and its temperature/source generalisations) as a Gaussian integral over the flow, then apply a derivative-operator identity that evaluates the integral by differentiating the flow map at a single point, $\\alpha=0$. This yields series expressions for posterior moments and a simple formula for the information entropy in terms of the Gaussian base density and the flow's Jacobian determinant. Applied to supernova constraints on the matter density $\\Omega_m$ and the dark-energy equation-of-state parameter $w_0$, the method reproduces the mean, variance, covariance and entropy obtained by MCMC; the same expansion currently fails for skewness and kurtosis because higher-order derivatives of the learned network become unreliable. If smoother flows can be trained, the approach would give parameter-free analytical access to non-Gaussian posterior statistics without additional sampling.","feed_headline":"Flow derivatives give posterior statistics without extra sampling","feed_subtitle":"For supernova dark-energy parameters, means and covariances match direct sampling; higher moments need smoother networks.","key_machinery":"The load-bearing object is the trained normalising flow $f(\\alpha)=\\theta(\\alpha)$, an invertible differentiable map with Jacobian $Df$ that Gaussianises the posterior. The mechanistic identity is the flow expansion of Eq. (15), which uses the operator identity $\\exp\\!\\left(\\frac{T}{2}\\delta^{\\rho\\sigma}\\partial_\\rho\\partial_\\sigma\\right)$ to convert the Gaussian integral over $\\alpha$ into derivatives of $g(\\alpha)=|\\det Df(\\alpha)|^{1-1/T}\\exp(J_\\gamma\\theta^\\gamma(\\alpha)/T)$ evaluated at $\\alpha=0$. Differentiating this expression with respect to the source $J$ generates the moment series, and the change-of-variables formula supplies the entropy formula; the same learned flow therefore carries the partition function, moments, and entropy from a single mapping.","core_discovery":"On its own terms, the paper's central discovery is that the normalising flow map $f(\\alpha)=\\theta(\\alpha)$, trained to send a standard normal $p(\\alpha)$ to the posterior $p(\\theta|y)$, can be differentiated through the partition function. After the change of variables, the Bayesian partition function becomes $$Z[T,J]=\\frac{(2\\pi T)^{n/2}}{N(T)}\\,\\exp\\!\\left(\\frac{T}{2}\\$delta^{{\\rho\\sigma}}$\\partial_\\rho\\partial_\\$\\sigma$\\right)g(\\$\\alpha$)\\big|_{\\$\\alpha$=0}$$ with $g(\\alpha)=|\\det Df(\\alpha)|^{1-1/T}\\exp(J_\\gamma\\theta^\\gamma(\\alpha)/T)$; the paper calls this the flow expansion. Posterior moments follow as $$\\langle \\$theta^{{\\gamma_1}}$\\cdots\\$theta^{{\\gamma_m}}$\\rangle=\\sum_{k=0}^\\infty \\frac{1}{2^k k!}\\bigl(\\$delta^{{\\rho\\sigma}}$\\partial_\\rho\\partial_\\$\\sigma$\\bigr)^k\\, \\$theta^{{\\gamma_1}}$(\\$\\alpha$)\\cdots\\$theta^{{\\gamma_m}}$(\\$\\alpha$)\\big|_{\\$\\alpha$=0},$$ while the entropy is computed as a sample average of $-\\ln p(\\alpha)-\\ln|\\det Df(\\alpha)|$ over the Gaussian base. The paper verifies the expansion on a Gaussian toy model and on the Union2.1 supernova posterior for $(\\Omega_m,w_0)$, where mean, variance, covariance and entropy match the emcee ground truth within errors. Its own tables show that skewness and kurtosis from the expansion are not reliable, diverging from the sampled values by up to an order of magnitude, which the paper attributes to unstable higher-order derivatives of the network.","pith_inferences":["A natural stress test is to apply the flow expansion to a one-dimensional mixture of Gaussians with known skewness and kurtosis: the single-point Taylor character means accuracy should degrade as the posterior becomes multimodal or heavy-tailed, and the order at which the series breaks down can be measured directly.","If smoother flow architectures deliver reliable higher derivatives, the same derivative machinery could compute the Fisher information and the surprise statistic (the Kullback-Leibler divergence between prior and posterior) without extra sampling, since both are expectation values over the posterior.","The free-energy compensation curves in the paper's Figs. 5 and 6 hint at a practical sampling strategy: choose $(T,J)$ along a constant-free-energy contour to anneal the posterior, which could be tested on a multimodal benchmark."],"forward_implications":["A single trained flow yields the Bayesian evidence, the temperature-extended partition function $Z[T,J]$, and posterior moments and entropy from one object, so evidence and parameter estimation no longer require separate sampling runs.","The dependence of the partition function on temperature $T$ and sources $J$ can be mapped out cheaply by sampling the Gaussian base and transforming through the flow, opening a thermodynamic view of posterior inference.","Because the free energy is available as a function of $(T,J_1,J_2)$, one can in principle compensate a change in temperature with a change in the sources while keeping the free energy fixed, which the paper suggests may improve sampling.","For smooth-enough flows, the expansion would produce cumulants of arbitrary order analytically, making skewness, kurtosis and related statistics direct outputs of the network rather than noisy sample estimates; the present bottleneck is network smoothness, not the formalism."],"supporting_citations":[{"why":"Introduced normalising flows as invertible differentiable maps, the foundation of the method used here.","marker":"Rezende and Mohamed (2015)"},{"why":"Supplies the review of normalising flow theory and training that motivates the density-transformation construction.","marker":"Papamakarios et al. (2019)"},{"why":"Provides the FrEIA software package used to train the flows and to supply Jacobian determinants and derivatives.","marker":"Ardizzone et al. (2018-2022)"},{"why":"Gives the operator-calculus identity used to turn the Gaussian integral into the single-point derivative expansion.","marker":"Rota and Doubilet (1975)"},{"why":"Demonstrates normalising-flow-based Bayesian evidence estimation, which this paper extends to temperature- and source-dependent partition functions.","marker":"Srinivasan et al. (2024)"},{"why":"Provides the partition-function and entropy framework used as the comparison baseline for the supernova application.","marker":"Kuntz et al. (2024)"},{"why":"Supplies the Union2.1 supernova dataset used for the cosmological posterior application.","marker":"Suzuki et al. (2012)"},{"why":"Provides the emcee sampler whose posterior samples serve as the ground truth for moment and entropy comparisons.","marker":"Foreman-Mackey et al. (2013)"}],"fun_headline_variants":["Normalising flows crack Bayesian partitions without extra MCMC","Flow derivatives yield posterior moments, but higher stats fail","Differentiable flows replace MCMC for non-Gaussian posteriors","Bayesian partition functions via normalising flows: fast stats, shaky tails","Flow expansion gives posterior moments; skewness and kurtosis unreliable"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The flow map is smooth enough that its Taylor expansion around $\\alpha=0$ converges quickly, so truncating the moment series at fourth order gives accurate statistics; the paper's own tables show this breaks for skewness and kurtosis.","fun_headline_variants_meta":{"raw":{"variants":["Normalising flows crack Bayesian partitions without extra MCMC","Flow derivatives yield posterior moments, but higher stats fail","Differentiable flows replace MCMC for non-Gaussian posteriors","Bayesian partition functions via normalising flows: fast stats, shaky tails","Flow expansion gives posterior moments; skewness and kurtosis unreliable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1605,"prompt_tokens":1048,"completion_tokens":557,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":664,"completion_tokens_details":{"reasoning_tokens":472}},"tokens_in":664,"tokens_out":557,"duration_ms":5372,"temperature":1.0,"reasoning_tokens":472,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:25:32.404950+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a normalising flow on a smooth, heavy-tailed one-dimensional distribution whose skewness and kurtosis are known analytically, then compute these moments from the flow expansion terminated at $k=4$; if the expansion values differ from the true values by more than the sampling noise, the claim that posterior moments follow from local derivatives of the flow is refuted for that class of distributions. The paper's Tables 4 and 5 already indicate such a failure for the supernova posterior kurtosis.","supporting_citations":[],"review_version":1}