{"id":"896f1518-cd39-4fcb-8a73-308ee0d5f727","arxiv_id":"2509.05877","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Predictive uncertainty in Gaussian process latent variable models is decomposed into epistemic and aleatoric parts via the law of total variance and estimated with Monte Carlo samples and random Fourier features, but the epistemic estimator drops the squared mean term.","lead":"This paper splits a model's uncertainty into a part that comes from too little data and a part that comes from noise in the data, then tests that split on a fast approximation of a Gaussian process latent variable model. It matters because getting this split right helps high-stakes prediction systems know when to collect more data, but the proposed estimator leaves out a needed term.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (31) estimates the second moment, not the variance, so the reported epistemic uncertainties are biased; the paper's own linear-data experiment has nonzero predictive mean, making the omitted centering term non-negligible.","rationale":"The reader identified the same load-bearing concern: Eq. (31) estimates a second moment instead of a centered variance, so the epistemic uncertainty numbers are systematically overestimated when the predictive mean is nonzero. This is precisely the paper's linear-output case. The concern is not a stylistic preference; it invalidates the quantitative claims in Figures 1–2 and the stated method for computing epistemic uncertainty. A correct variance estimator requires subtracting |E[a]|^2, and the paper provides no justification for omitting it. The related issue in Eq. (23) reinforces that the derivation conflates variance over x* with joint variance over x* and θ, but the core defect remains the uncentered Monte Carlo estimator. No other concern needs to be raised: the synthetic setup and RFF approximations are not themselves the central failure. Since the reader's verdict of REJECT already follows from this unsupported central estimator, no change to the verdict is needed.","tokens_in":7053,"tokens_out":11650,"duration_ms":102580,"concrete_test":"Recompute the final summand of Eq. (31) on the paper's synthetic setup (Section 6, e.g., J = 100, M = 100, L = 100) using the centered estimator (1/(ML))Σ_{m,l} a_{m,l}^2 − ((1/(ML))Σ_{m,l} a_{m,l})^2 with a_{m,l} = φ(x*^{(m,l)})^T E[θ_d|X^{(m)},Y]. If the centered epistemic estimates are materially smaller than the published uncentered ones (expected because E[a] is nonzero for the linear output), the estimator in (31) is biased and the experimental insights are not supported. Independently, evaluating Eq. (23) for a scalar toy case, e.g., φ(x) = x, x ~ N(0,1), E[θ] = 1, Cov[θ] = 1, immediately shows the identity fails as written, confirming the derivation error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Monte Carlo estimator in Eq. (31) does not estimate the variance in Eq. (24). The final summand of (31) computes (up to averaging constants) (1/L)Σ(φ(x*)^T E[θ_d|X,Y])^2, which is the sample second moment of a = φ(x*)^T E[θ_d|X,Y]. The corresponding term in Eq. (24) is Var[a] = E[a^2] − (E[a])^2, and the squared-mean term is never subtracted. The zero-mean condition E[a] = 0 is neither stated nor implied by the RFF construction. In the paper's own synthetic example, y1 = w1^T x + ε, the posterior predictive mean is a nonzero linear function over the test-latent distribution, so E[a] ≠ 0. The paper also explicitly uses μ_φ = E[φ(x*)] in the first line of (31), so it cannot assume E[φ^T Eθ] = 0 without invalidating that term. The same conflation appears earlier: Eq. (23) writes Var_{x*}(φ^T E[θ|X,Y]) as E[φ^T Cov[θ|X,Y] φ] + Var_{x*}(φ^T E[θ|X,Y]), which is an identity only if the left side is the joint variance Var_{x*,θ}(φ^T θ), not the conditional variance over x* alone; the text does not supply that interpretation. Thus Eq. (24) is not established as written, and the reported epistemic uncertainties reflect E[a^2] rather than Var[a], so the central claim that Eqs. (31)–(32) provide correct epistemic and aleatoric uncertainty estimates is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for decomposing predictive uncertainty in Gaussian process latent variable models (GPLVMs) with Random Fourier Feature (RFF) approximations into aleatoric and epistemic components. It derives an expression for the epistemic variance using the law of total variance, introduces Monte Carlo estimators for both components, and reports synthetic experiments with four output types (linear, quadratic, periodic, step) to illustrate how the estimates behave as the number of RFF features varies. The central claim is that Eqs. (18), (24), (31), and (32) provide a correct decomposition of total predictive variance and that the proposed Monte Carlo estimates are valid.","tokens_in":7367,"tokens_out":4200,"duration_ms":39679,"significance":"If the derivation and estimator were correct, the paper would offer a useful practical recipe for uncertainty decomposition in a scalable GPLVM setting, and the synthetic study would provide some insight into how uncertainty estimates behave across output types. The authors also deserve credit for explicitly acknowledging that Monte Carlo variation is not accounted for, which is a common omission. However, the significance is undercut by a load-bearing error in the Monte Carlo estimator: as written, Eq. (31) estimates a second moment rather than a variance, so the reported epistemic uncertainties are biased. The associated theoretical derivation in Eqs. (23)--(24) also contains a notational/identity problem that must be repaired. Because these issues affect the paper's central claim, the current version cannot be accepted as a reliable contribution to UQ methodology.","major_comments":[{"comment":"Eq. (31) does not estimate the variance appearing in Eq. (24). The final term of (31) is, up to averaging constants, (1/L) Σ_ℓ (φ(x*^{(ℓ)})^T E[θ_d | X, Y])^2, which is the sample second moment of a = φ(x*)^T E[θ_d | X, Y]. The corresponding expression in Eq. (24) is Var[a] = E[a^2] − (E[a])^2, and the centering term is never subtracted. The paper neither states nor enforces the zero-mean condition E[a] = 0; indeed, in the synthetic linear output y1 = w1^T x + ε, the posterior predictive mean is generally nonzero over the test-latent distribution, so the omitted term is non-negligible. This is not a cosmetic issue: the reported epistemic uncertainties in Fig. 2 are biased upward by the squared mean of the predictive function, which invalidates the central claim that Eqs. (31)--(32) provide correct epistemic and aleatoric uncertainty estimates.","section":"§5, Eq. (31); §4, Eq. (24)"},{"comment":"Eq. (23) as printed is an identity only under an interpretation that conflicts with the notation used elsewhere. The left side, Var_{p(x*|...)}(φ(x*)^T E[θ_d|X,Y]), is identical to the second term on the right side, so the equation is circular as written. The intended identity presumably involves the joint variance Var_{x*,θ}(φ^T θ), in which case the correct decomposition is Var_{x*,θ}(φ^T θ) = E_{x*}[φ^T Cov[θ|X,Y] φ] + Var_{x*}(φ^T E[θ|X,Y]). The text does not state this interpretation, and Eq. (24) therefore does not follow from the equations as written. This derivation issue must be fixed for the theoretical formulation to be sound.","section":"§4, Eqs. (23)--(24)"},{"comment":"The experimental section does not validate the uncertainty estimates. The paper reports boxplots of estimated aleatoric and epistemic variances but provides no calibration metric (e.g., coverage of predictive intervals), no comparison against an exact GP or another baseline, and no numerical comparison with the known generative noise variance σ²_epsilon = 1, even though such a comparison is possible for the linear output. The statement that the aleatoric uncertainty for y1 was 'estimated accurately' is therefore not supported by the presented evidence. This matters because the estimator bias identified in Eq. (31) directly affects the plotted epistemic values.","section":"§6, Figs. 1--2"}],"minor_comments":[{"comment":"The prior notation p(Σ) should be p(Σ_y) for consistency with the earlier definition of Σ_y.","section":"§3, Eq. (10)"},{"comment":"The text contains an unresolved placeholder reference 'Fig. ??', which should be fixed before submission.","section":"§3, after Eq. (16)"},{"comment":"The quantities bθ_d and bσ²_o,*,d are described only as 'estimated values' without specifying whether they are MAP estimates, posterior means, or derived from a particular Monte Carlo sample; this should be clarified.","section":"§3, Eq. (17)"},{"comment":"The prose around Eq. (21) contains a typographical repetition ('variations in x* variations') and missing punctuation; these should be cleaned up.","section":"§4, Eqs. (20)--(21)"},{"comment":"The reference list is not consistently formatted (e.g., volume/page information is incomplete for several entries); this should be standardized.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The estimator error in Eq. (31) is serious, but it is fixable: the authors need to replace the second-moment term with a centered sample variance and rerun the experiments. Given that the central decomposition idea is salvageable, I am recommending major revision rather than rejection. If the authors cannot supply corrected experiments, the paper should be rejected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a straightforward application of the law of total variance to uncertainty quantification in GPLVMs with RFFs. The exposition is readable, and the authors correctly identify the three sources of epistemic uncertainty (training latents, test latent, GP weights) and propose a Monte Carlo scheme. The synthetic experiments are fine as sanity checks. That is the good part.\n\nThe problem is load-bearing: Eq. (31) does not estimate Eq. (24). The last summand of (31) is (roughly) the sample second moment of a = φ(x*)⊤E[θ|X,Y], averaged over X and x*. But Eq. (24) requires the variance of a, which is E[a²] − (E[a])². The centering term is never subtracted. In the paper's own linear example, the mean of a is nonzero, so the bias is real and not a technicality. Eq. (23) also mislabels the variance as being over x* alone when it only holds as a joint variance over θ and x*; the text doesn't clarify that. So the central claim that Eqs. (31)–(32) give correct epistemic and aleatoric estimates is unsupported.\n\nBeyond the math, the experimental section has no baselines, no calibration against actual prediction errors, no error bars on the uncertainty estimates, and no code. The authors themselves note that Monte Carlo induced uncertainty is unaccounted for, which is a significant omission for a UQ paper. The citation pattern also misses the obvious prior work on epistemic/aleatoric splits (Kendall & Gal, Depeweg et al.), so the novelty claim is overstated.\n\nWho is this for? A reader new to GPLVMs might learn something from the setup, but as a research contribution it is not there yet. The core estimator error is fixable, but as written the paper would mislead anyone who takes the epistemic numbers at face value. I would not send it to referees in its current state; a desk reject with an invitation to resubmit after fixing the estimator and adding baselines seems right. If revised, it might become a workshop-level paper.\n\nMy recommendation: don't engage further until the authors correct the centering and show calibrated results. I would not cite it in this form.","headline":"The epistemic uncertainty estimator in Eq. (31) computes a second moment, not a variance, so the paper's central quantity is biased; the rest is a clean but standard GPLVM UQ exposition.","tokens_in":7978,"tokens_out":3495,"would_cite":false,"duration_ms":32287,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Predictive uncertainty in GPLVMs decomposes exactly into aleatoric and epistemic parts.","keywords":["uncertainty quantification","epistemic uncertainty","aleatoric uncertainty","Gaussian process latent variable model","random Fourier features","Monte Carlo estimation","law of total variance","predictive variance decomposition"],"falsifier":"In the synthetic linear output $y_1 = w_1^\\top x + \\epsilon$, compute the mean of $\\phi(x_*)^\\top E[\\theta_1|X,Y]$ over the $L$ samples; if it is not close to zero, then Eq. (31)'s second term is a mean square rather than a variance, and the epistemic uncertainty reported for that output is overstated.","tokens_in":6745,"feed_emoji":"📊","tokens_out":7237,"duration_ms":61303,"temperature":0.7,"pith_summary":"This paper argues that in a Gaussian process latent variable model equipped with random Fourier features, the total predictive variance of a missing output can be decomposed exactly into an aleatoric part, the expected irreducible noise variance, and an epistemic part that collects uncertainty in the training latent positions, the test latent position, and the Gaussian process weights. The decomposition follows from the law of total variance, applied twice, so that epistemic uncertainty is further split into a contribution from parameter covariance and a contribution from test-input variability. To make the decomposition computable, the paper proposes a Monte Carlo estimator that draws samples of the training latents and of the test latent position and evaluates the random-feature predictive mean and noise variance. If correct, the framework gives practitioners two actionable numbers for every imputed value: how much uncertainty is intrinsic to the data and how much could shrink with better inference or more data.","feed_headline":"One formula splits prediction error into data noise and model doubt","feed_subtitle":"For Gaussian-process latent-variable models, a Monte Carlo recipe computes both parts and tells you which one more data would reduce.","key_machinery":"The load-bearing device is the random Fourier feature map $\\phi(x)$, which turns each Gaussian process output into a linear function $\\phi(x)^\\top \\theta_d$. Because the conditional predictive mean is linear in the weights, the law of total variance can be applied at the level of $\\phi(x_*)^\\top E[\\theta_d|X,Y]$, producing the expressions in Eqs. (20) through (24) that isolate contributions from posterior weight covariance, test-input variability, and training-latent variability. The same feature map makes the Monte Carlo estimator in Eq. (31) possible: samples $x_*^{(m,\\ell)}$ plug into $\\phi$, and the posterior weight moments $\\mathrm{Cov}[\\theta_d|X^{(m)},Y]$ come from the RFF-based Gaussian process fit.","core_discovery":"The paper's central claim is that Eqs. (18), (24), (31), and (32) give a correct and systematic accounting of where predictive uncertainty comes from in an RFF-based GPLVM. Starting from the conditional model $y_{d,*} \\sim N(\\phi(x_*)^\\top \\theta_d, \\sigma_d^2)$, the total variance $\\mathrm{Var}[y_{d,*}]$ equals the expectation of $\\sigma_d^2$ over the test-latent posterior, plus the variance of the conditional mean $\\phi(x_*)^\\top E[\\theta_d|X,Y]$. The epistemic part is then broken down by the law of total variance into an expected term involving $\\mathrm{Cov}[\\theta_d|X,Y]$, a trace term tracking fluctuations of $\\phi(x_*)$ around its mean, and a term tracking the variability of the conditional mean across training-latent posteriors. The proposed Monte Carlo scheme evaluates each of these terms using $M$ posterior draws of $X$ and $L$ draws of $x_*$ per $X$ draw. On the paper's own terms, this is a principled route to separate irreducible data noise from model uncertainty in a scalable Gaussian process approximation.","pith_inferences":["A diagnostic the paper does not report: comparing the sample second moment in Eq. (31) with a properly centered sample variance would quantify how much the zero-mean assumption on $\\phi(x_*)^\\top E[\\theta_d|X,Y]$ matters, and publishing that gap would let readers judge the estimator directly.","Because the decomposition only uses linearity of the conditional mean in $\\theta_d$, it transfers directly to Bayesian linear regression, Bayesian neural network last layers, or other feature-based linear models, where the same total-variance identities apply.","A follow-up experiment could split epistemic uncertainty by source for each of the four synthetic outputs and check whether the weight-covariance term correlates with $J$, the input-variability term with the smoothness of the function, and the training-latent term with $M$."],"forward_implications":["Each predicted missing output comes with a split: aleatoric uncertainty tells the floor set by data noise, and epistemic uncertainty tells how much remains available to shrink.","A practitioner can attribute epistemic uncertainty to training latents, test position, or GP weights and decide whether more data, better latent inference, or more random features is the right fix.","The RFF approximation makes the whole computation scalable, with cost controlled by the number of frequencies $J$ and the Monte Carlo sample sizes $M$ and $L$, so the method applies to larger datasets than exact GP GPLVMs.","Smooth outputs such as the periodic function keep epistemic uncertainty low, while a step-function output shows high epistemic uncertainty, reflecting the GP's known difficulty with discontinuities."],"supporting_citations":[{"why":"Defines the GPLVM and the posterior over latent variables that the uncertainty decomposition is built on.","marker":"Titsias and Lawrence (2010)"},{"why":"Introduces random Fourier features, which produce the finite feature map $\\phi(x)$ that makes the predictive mean linear in $\\theta_d$.","marker":"Rahimi and Recht (2007)"},{"why":"Supplies the Gaussian process predictive framework and kernel assumptions underlying the GPLVM likelihood.","marker":"Rasmussen and Williams (2006)"},{"why":"Provides Bochner's theorem, justifying sampling the RFF frequencies from the kernel's power spectral density.","marker":"Rudin (2017)"}],"fun_headline_variants":["Use total variance law to split data noise from model doubt","Monte Carlo estimates aleatoric and epistemic uncertainty","Decompose predictive variance into noise and model uncertainty","RFF-GPLVM: separate aleatoric from epistemic error","Equation unlocks breakdown of prediction uncertainty"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central estimator's second term is only a variance if the expected prediction, averaged over all plausible latent positions of the test point, is zero; the paper neither states nor enforces that condition.","fun_headline_variants_meta":{"raw":{"variants":["Use total variance law to split data noise from model doubt","Monte Carlo estimates aleatoric and epistemic uncertainty","Decompose predictive variance into noise and model uncertainty","RFF-GPLVM: separate aleatoric from epistemic error","Equation unlocks breakdown of prediction uncertainty"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1380,"prompt_tokens":905,"completion_tokens":475,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":399}},"tokens_in":521,"tokens_out":475,"duration_ms":4722,"temperature":1.0,"reasoning_tokens":399,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:20:28.541391+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the synthetic linear output $y_1 = w_1^\\top x + \\epsilon$, compute the mean of $\\phi(x_*)^\\top E[\\theta_1|X,Y]$ over the $L$ samples; if it is not close to zero, then Eq. (31)'s second term is a mean square rather than a variance, and the epistemic uncertainty reported for that output is overstated.","supporting_citations":[{"cited_title":"and Lawrence, N","cited_arxiv_id":null,"evidence_quote":"Defines the GPLVM and the posterior over latent variables that the uncertainty decomposition is built on."},{"cited_title":"and Recht, B","cited_arxiv_id":null,"evidence_quote":"Introduces random Fourier features, which produce the finite feature map $\\phi(x)$ that makes the predictive mean linear in $\\theta_d$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian process predictive framework and kernel assumptions underlying the GPLVM likelihood."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides Bochner's theorem, justifying sampling the RFF frequencies from the kernel's power spectral density."}],"review_version":2}