acceptodds
Under review as a conference paper at ICLR 2027

Self-Consistent Stochastic Interpolants for Covariance Prior and Posterior Learning

Abstract

Given many noisy covariance estimates, each computed from a few Gaussian samples with its own underlying covariance, how do we sample the covariance that produced a given estimate when no clean covariance is ever seen? This is an inverse problem: we also know the sampling mechanism, so each estimate is a draw from an unknown prior, seen through a channel we can simulate. Expectation–maximization over distributions undoes that blur, but every step requires sampling the posterior of the current prior, which no formula provides. We show that a conditional stochastic interpolant trained on its own reconstructions supplies that sampler and turns the iteration into an algorithm. Two versions target the same posterior over the whole matrix: one transports the matrix itself; the other, for a rotationally invariant prior, learns only the eigenvalue distribution and draws the relative orientation from an explicit conditional law, a completion that adds no distributional error. We prove identifiability exactly when each estimate uses at least as many samples as dimensions, and show that progress in what is observable is weaker than recovery of the prior, which needs further conditions. On synthetic data we recover both invariant and structured priors from noisy estimates alone, and on EEG and market records the posterior improves calibration and decisions.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
covariance estimation, posterior sampling, empirical Bayes, stochastic interpolants, expectation--maximization, Wishart model, rotational invariance, Ledoit--Wolf shrinkage, Ledoit--Peche nonlinear shrinkage.

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_8JnvOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper learns a population covariance prior and task-specific posteriors from noisy covariance estimates using a known Wishart simulator. Conditional stochastic interpolants implement distribution-valued EM under exact population training and sampling. The paper establishes a sharp identifiability threshold, separates observable convergence from latent recovery, and derives an exact spectral reduction for rotationally invariant priors. Experiments cover synthetic distributions, EEG, and financial returns.

Critical strengths
  1. The theory identifies substantive limits of covariance-prior learning. Theorem 1 establishes identifiability exactly when N≥dN\ge d, with indistinguishable smooth, compactly supported, rotationally invariant priors below the threshold. Theorem 2 carefully separates observation-space convergence, weak latent recovery under compact spectral bounds, and stronger recovery requiring inverse stability. [pp. 3–4; Appendices B–D]

  2. The spectral reduction preserves full posterior uncertainty. The conditional orientation law is independent of the unknown spectral prior, so exact completion adds no KL error. The accompanying analysis distinguishes population eigenvalues, empirical-direction variances, and loss-specific Bayes estimates, avoiding the incorrect identification of aligned eigenvalue samples with matrix posterior samples. [Theorem 4, p. 6; Appendix G]

Critical weaknesses
  1. The likelihood and oracle specifications are inconsistent. Finance uses d=12d=12 and forecast degrees of freedom Nf=6N_f=6, but evaluates covariance targets formed from 21 trading days. Every simulated forecast has rank at most six, whereas those targets are generally full rank. Consequently, the specified singular predictive law cannot provide the claimed full-matrix predictive density. A covariance-dependent expression evaluated outside that support would be a surrogate score rather than the stated predictive density. Directional calibration and realized portfolio returns are not automatically invalidated. [Appendices J.1/J.4–J.5, pp. 42, 45–47; Figure 4c; Table 6] Separately, the main AR(1) model and quadrature oracle use lognormal decay γ\gamma, while Appendix J.2 specifies uniform correlation ρ\rho. These priors are not equivalent under ρ=exp⁡(−γ)\rho=\exp(-\gamma), leaving the largest reported oracle comparison unresolved. [pp. 7, 43]

  2. The main empirical claims exceed the stronger validation. Main-text EEG accuracies are 0.852 overall and 0.956 on the confident half; five-fold subject-disjoint evaluation gives 0.806 and 0.911, and the appendix explicitly states that the larger single-split gains do not carry over. [p. 8; Appendix J.3/Table 3, pp. 44–45] The principal empirical conclusions should reflect the subject-disjoint evaluation and distinguish it from the favorable single split.

Questions for authors
  1. What probability law produced the financial matrix log-scores at Nf=6N_f=6? Please provide a valid-support reevaluation, or identify and justify the surrogate score while separating its conclusions from portfolio returns.

  2. Which AR(1) prior generated Table 1 and its oracle? Please reconcile the generator and quadrature specification and establish whether corrected definitions leave the reported comparison unchanged.

  3. Can the main EEG comparison be stated using the five-fold subject-disjoint results, including baseline differences and variability? Resolving the score/oracle definitions and showing which gains survive these checks would support a rating of 3.

5 comments

We thank the reviewer for their careful reading. All three issues are correct, and we fixed them as follows.

Q1. Finance log-score with Nf=6<d=12N_f=6<d=12. The reviewer is right. Appendix A defines the channel density for N<dN<d relative to a reference measure, but that density lives on matrices of rank at most NfN_f. The realized 21-day matrices have full rank, so they fall outside its support. Since terms free of CC were dropped, what we computed for each posterior draw CjC_j is −Nf2[log⁡det⁡Cj+tr⁡(Cj−1Cf)]=Nf21∑t=121log⁡N(rt;0,Cj)+const,-\tfrac{N_f}{2}\big[\log\det C_j+\operatorname{tr}(C_j^{-1}C_f)\big]=\tfrac{N_f}{21}\sum_{t=1}^{21}\log\mathcal N(r_t;0,C_j)+\text{const}, a tempered Gaussian log-likelihood of the 21 forward daily returns rtr_t, with temperature s=Nf/21s=N_f/21. It is a surrogate score, not the stated predictive density.

In the revision we name it as such and add two proper scores of the 21 returns (results in a follow-up):

  • Gaussian, log⁡1J∑j∏tN(rt;0,Cj)\log\frac1J\sum_j\prod_t\mathcal N(r_t;0,C_j): up to terms free of CC, the Wishart density with 21 degrees of freedom, valid since 21≥d21\ge d;
  • Student-tt with covariance CjC_j and excess kurtosis κ=5\kappa=5. Since 2⋅21/(2+κ)=Nf2\cdot 21/(2+\kappa)=N_f, this law carries, to second order, the same extra dispersion as the forward channel used for coverage.

Neither score enters the portfolio. Coverage of forward directional variances is unaffected: given CC, v⊤Cfvv^\top C_f v is (v⊤Cv) χNf2/Nf(v^\top Cv)\,\chi^2_{N_f}/N_f.

We also clarify the role of NN and NfN_f. They are effective degrees of freedom, chosen together on a validation fold that precedes all test folds (Table 4). They absorb heavy tails and the drift between the estimation and forecast windows. A new result gives this a precise meaning: if non-Gaussian samples are treated as Gaussian, the iteration converges to the prior under which the Wishart channel best explains the observed covariances, and Theorem 2 holds relative to that prior. Choosing NN is then model selection between well-defined projections. Fig. 4c is labeled as a tempered likelihood and shown next to the proper scores (Fig. 4d).

Q2. AR(1) prior. The reviewer is right that the two descriptions are inconsistent. Table 1 and its quadrature oracle both use Cij=ρ∣i−j∣C_{ij}=\rho^{|i-j|} with ρ∼U(0.3,0.9)\rho\sim\mathcal U(0.3,0.9). The lognormal-γ\gamma description (§4.1, J.2) referred to a variant not used for Table 1 and is removed; the bound 1.0 in J.2 was a typo for 0.9. We also corrected the stated sample size: all synthetic runs use N=40N=40, not the N=28N=28 written in the submission.

In the rerun (longer training, fresh observations at every iteration), SC-SI is 0.03 nats from the oracle on AR(1), with every W1W_1 (median over tasks) at most 0.38 posterior s.d. IW-ML is 9.7 nats below the oracle (23.3 in the submission, whose log-scores used only J=256J=256 draws; we now use J=2⋅106J=2\cdot10^6). The conclusion of the submission is unchanged: an inverse-Wishart prior cannot follow this family, and SC-SI can.

Q3. EEG, five-fold evaluation. Agreed. The main text now reports five-fold subject-disjoint cross-validation over all 109 subjects, and the 29-subject split moves to the appendix. Its test subjects turned out to be easier for every method: the sample-covariance accuracy was 0.820 there, against 0.790 over all subjects.

Log-score and calibration, where the main evidence lies:

  • Log-score gain over the sample covariance: +5.2 [3.7, 6.4] nats per window in split-half and +17.3 [14.5, 20.7] in next-window. The paired margins over LW-NLS are +2.5 [1.2, 3.5] and +7.8 [6.1, 9.2].
  • Split-half 80% coverage: directional variances 0.76 against 0.64–0.67. With 95% subject-level intervals (new Table 2c), log-determinant 0.67 [0.64, 0.69] against at most 0.46 [0.44, 0.48], and log condition number 0.75 [0.71, 0.78] against at most 0.67 [0.65, 0.69].

Classification, pooled over the five folds (3052 windows):

  • Accuracy 0.804 against 0.790 / 0.776 / 0.787 for the sample covariance / OAS / LW-NLS. The paired gains are +1.4 [0.2, 2.7], +2.8 [0.6, 4.9] and +1.7 [0.4, 3.0] points.
  • NLL 0.442 against 0.469–0.497, with all paired differences significant; ECE 0.021 against 0.038–0.052.
  • On the 50% most confident windows, SC-SI is better than OAS (+3.9 [1.2, 6.4]). It is not separated from the sample covariance (−0.1 [−1.6, 1.3]) or LW-NLS (+1.0 [−1.0, 2.8]).

The main-text claim becomes: a clear gain in log-score and calibration, and a modest but significant gain in classification accuracy and NLL.

Summary. You asked us to resolve the score and oracle definitions and to show which gains survive. The AR(1) oracle is consistent and the comparison stands. The finance score is correctly named, with proper scores to follow, and the portfolio result does not depend on it. On all 109 EEG subjects the log-score and calibration gains survive, and the classification gain is smaller but significant. We hope this meets the conditions you set for a positive rating.

AuthorsOpenReview ↗

revised PDF

The revised PDF is uploaded, and the finance proper scores are in the general comment above. Scoring the 21 forward daily returns, SC-SI gains +1.21 [0.97, 1.52] nats per day over the sample covariance under a Gaussian predictive and +0.48 [0.45, 0.52] under a Student-tt predictive, and it is best under both. The paired gains over OAS (+0.27 [0.25, 0.29]) and IW-ML (+0.18 [0.16, 0.20]) are positive in every volatility tercile and every test decade. The AR(1) specification is corrected in §4.1 and Appendix J.2, and the five-fold EEG results are in §4.2 and Tables 2–3. We hope this settles the three points you raised.

Dear Reviewer, as the discussion period closes on Nov 18, we wanted to check whether our reply and the revised PDF address your concerns. In particular, the three points you listed (finance score, AR(1) oracle, EEG evaluation) are now resolved in the revision, with the proper finance scores in our general comment. We would be glad to clarify anything further before then. Thank you again for your review.

Reviewer_tCuxOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper tries to model a distribution with only noisy empirical covariances, i.e., without clean covariance observations. A self-consistent non-parametric EM iteration algorithm is implemented, using a conditional stochastic interpolant trained on reconstructed and then re-corrupted samples. A matrix-logarithm implementation is complemented by a spectral formulation for rotationally invariant priors, with an explicit conditional orientation law. The theoretical contributions include a sharp identifiability threshold, population EM analysis, and an exact KL identity for spectral completion. Experiments cover six synthetic priors, EEG, and financial returns.

Critical strengths

I find the problem setup well motivated and the designed experiments highly relevant. I also think it is a valuable insight to distinguish between a fit in the observation space and the recovery of the latent prior. It is also useful to see the sharp identifiability result from Theorem 1. The paper clearly acknowledges its relationship to existing distributional EM and self-consistent stochastic-interpolant methods. I also appreciate the discussion of the detailed experimental setups.

Critical weaknesses

The performance on prediction seems to be overstated and not backed up firmly by the experiments. For example. In Table 1, the noticeable gap in the Wasserstein distance between the reconstructed posterior and the oracle cannot support a method for posterior recovery. Another example along this line is about EEG. The main-text 29-subject split and Figure 3c support the reported 0.852 accuracy and 0.956 accuracy at 50% coverage. However, the five-fold full-cohort evaluation gives a smaller improvement: 0.806 versus 0.792 for sample covariance, with a paired gain of 1.4 percentage points and 95% CI [0.1, 2.8]. At 50% coverage, SC-SI has no clear advantage over sample covariance or LW-NLS.

Questions for authors
  1. Could the Wasserstein distance gap be linked to the acknowledged limitations of the unweighted empirical loss? Wonder if such cases can be scrutinized and, if possible, converted into diagnostics.
  2. Would appreciate more elaboration on the aforementioned EEG comparison.
  3. Does Figure 7 use Algorithm 2? Please specify its setup and provide one validation of conditional orientation and completed-matrix accuracy, rather than spectral marginals alone.
  4. Which AR(1) prior generated the data and oracle: uniform ρ on [0.3, 1), or Gaussian log γ? Please reconcile the descriptions in Section 4.1 and Appendix J.2.
  5. With d = 12 and N_f = 6, what density or surrogate produces the finite finance log-scores for generally full-rank 21-day realized matrices?
  6. Can one focused control, such as the small-ensemble exact EM reference proposed in H.4 or a sampler-resolution check, clarify the remaining posterior-recovery errors?
7 comments

We thank the reviewer for their thorough report. We agree that several empirical claims were stated more strongly than the strongest validation supports. We have corrected them, and the new results follow.

W1 gap to the oracle (weakness, Q1, Q6). We reran the synthetic benchmark at the same N=40N=40 with three changes, all stated in the revision: wider networks and longer training; log-scores estimated with J=2⋅106J=2\cdot10^6 posterior draws instead of 256 (the log-mean-exp estimator is biased at small JJ, so the submitted log-score gaps, IW-ML included, were not reliable); and a fresh batch of 5000 observations at every outer iteration (M=∞M=\infty, the setting of Theorem 2) instead of a fixed ensemble of M=5000M=5000. The last change separates the two sources of the gaps you point out.

  • Statistical. With a finite ensemble, exact EM does not target π⋆\pi_\star: a result we add shows that it descends toward the nonparametric MLE of the ensemble, which is atomic. This over-sharpening is why checkpoints must be selected, and it made the submitted comparison with the oracle mix statistical and algorithmic error.
  • Approximation, from the regression and the SDE sampler. With M=∞M=\infty this is what remains: every W1W_1, as a median over test tasks, is at most 0.40 oracle posterior s.d. (the submission reported means; medians are more stable, and the appendix gives both), and the log-score gap is at most 0.54 nats.

We revised the wording of the claims accordingly, and are running finite-ensemble and sampler-resolution controls for the final version.

Unweighted loss and diagnostics (Q1). Yes, the two are linked. The regression loss weights time uniformly, while the posterior error bound of Appendix F weights errors near t=1t=1 singularly, so a small loss does not control every functional. The diagnostic that needs no oracle is the per-functional PIT score we use for checkpoint selection, which takes the worst case over functionals so that one cannot be traded against another.

EEG (weakness, Q2). We agree. The main text now reports five-fold subject-disjoint cross-validation over all 109 subjects:

accacc@50%NLLECE
Sample cov.0.7900.9110.4690.052
OAS0.7760.8710.4970.038
LW-NLS0.7870.9000.4710.042
SC-SI0.8040.9100.4420.021
  • Accuracy. The paired gains are +1.4 [0.2, 2.7], +2.8 [0.6, 4.9] and +1.7 [0.4, 3.0] points over the sample covariance, OAS and LW-NLS. On the confident half, SC-SI is better than OAS only.
  • Log-score. The posterior advantage holds on all 109 subjects: +5.2 [3.7, 6.4] nats per window over the sample covariance in split-half, and +17.3 [14.5, 20.7] in next-window.
  • Calibration. In split-half, the 80% coverage of the log-determinant is 0.67 against 0.18–0.46, and that of the directional variances is 0.76 against 0.64–0.67.

The EEG section now leads with calibration and log-score, and presents classification as a modest gain.

Figure 7 (Q3). We agree that spectral marginals do not validate the orientation, and Figure 7 is removed. Algorithm 2 (learned spectral posterior, completed with the exact orientation law) now runs on the two rotationally invariant priors. Its learning curves appear in Figure 2 next to the matrix version. New rows of Table 1 evaluate the completed matrix against the oracle with the Table 1 metrics.

  • The mean W1W_1 is 0.066 on IW-RI and 0.19 on iso-diagonal, against 0.066 and 0.18 for the matrix version.
  • The correlation C01C_{01} and the variances along the top and bottom empirical eigenvectors (dvtopdv_{\rm top}, dvbotdv_{\rm bot}), which depend on the orientation, are within 0.05–0.18 posterior s.d.
  • On IW-RI its log-score gap is 0.70 nats, against 0.27 for the matrix version, so learning only the spectrum is cheaper but not free.

AR(1) prior (Q4). Table 1 uses Cij=ρ∣i−j∣C_{ij}=\rho^{|i-j|} with ρ∼U(0.3,0.9)\rho\sim\mathcal U(0.3,0.9) for both the data and the quadrature oracle. The lognormal-γ\gamma description referred to a variant that was not used for Table 1, and it has been removed. We also corrected the stated sample size: all synthetic runs use N=40N=40, not 28.

Finance log-score (Q5). With Nf=6<dN_f=6<d the Wishart forecast law is singular, so the reported quantity was not its density. Since terms free of CC were dropped, it equals Nf21∑t=121log⁡N(rt;0,C)\tfrac{N_f}{21}\sum_{t=1}^{21}\log\mathcal N(r_t;0,C), a tempered Gaussian log-likelihood of the forward returns. The revision names it as such and adds two proper scores of the 21 returns, under a Gaussian predictive and under a Student-tt predictive whose tails match the effective NfN_f; the results will follow in a comment. Coverage of the directional variances is unaffected, since v⊤Cfvv^\top C_f v given CC is a scaled χNf2\chi^2_{N_f}, and the portfolio results do not use the score.

We hope these results address your concerns, and would be grateful if you would reconsider your assessment.

AuthorsOpenReview ↗

revised PDF

The revised PDF is uploaded. The rerun synthetic benchmark is in §4.1, Fig. 2 and Table 1, including the rows for Algorithm 2; Table 1′ (Appendix J.2) gives the means over tasks. The five-fold EEG evaluation is in §4.2 and Tables 2–3. The finance proper scores are in the general comment above: SC-SI has the best score under both the Gaussian and the Student-tt predictive, with paired gains of +0.27 [0.25, 0.29] nats per day over OAS and +0.18 [0.16, 0.20] over IW-ML. We would be glad to answer any further question.

Dear Reviewer, as the discussion period closes on Nov 18, we wanted to check whether our reply and the revised PDF address your concerns. In particular, the rerun synthetic benchmark (§4.1, Fig. 2, Table 1) brings every W1 distance to the oracle below 0.40 posterior s.d., Algorithm 2 is now validated on the completed matrix, and the EEG section reports five-fold cross-validation over all 109 subjects. We would be glad to clarify anything further before then. Thank you again for your review.

Reviewer_RV5zOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This is the problem: Given a distribution π\pi over covariances, some samples C1,C2,⋯ ,CM∼πC_1, C_2, \cdots, C_M\sim\pi, and some finite sample covariances Ce,1,Ce,2,⋯ ,Ce,MC_{e,1}, C_{e, 2}, \cdots, C_{e,M} computed from samples drawn from C1,C2,⋯ ,CMC_1, C_2, \cdots, C_M, provide a sampler for π\pi (so a way to get a new CM+1C_{M+1}) and give a sampler from the population covariance CiC_i from any Ce,iC_{e, i}.

The authors prove that an EM algorithm (guess π\pi, sample CC, simulate sample covariances, use bayes rule to update π\pi) works well. The "ideal" algorithm updates π\pi density directly, but two practical versions are studied - one where that is based on a model that iterates a flow model that maps CeC_e to CC, and another that is based on a more explicit map that applies when π\pi is rotationally invariant (CC has the same density as QCQ⊤QCQ^\top for orthonormal QQ).

Critical strengths

I think the strongest result is the implementation via stochastic interpolates that uses a trained conditional flow. The idea is to take simulated pairs (C,Ce)(C, C_e) and train a model for those. These simulated pairs come from CC sampled from the current π\pi guess.

The ideal EM algorithm is shown to converge, and there is some effort to show this also for the practical algorithm using conditional flow with an error analysis.

The use of Expression (7) is nice. I dont think it is novel, but it is a clean, interpretable expression connecting some of the important quantities.

Critical weaknesses

The setting is very stylized - we want to learn a distribution over gaussians to the end also of understanding the posterior over the population covariance from a sample covariance.

Also I think some of this is similar to prior work (Aubin Frankowski et al., 2022).

Questions for authors

I discussed with AI the possibilities of extending this to general π\pi (that is, not just a distribution over gaussians). It seems like one immediate blocker is just a way to even represent distributions CC (now we can just use a covariance), and I imagine various analyses will be much harder.

But just from the perspective of an algorithm, it seems like the main difficulty is just representing the distribution for the stochastic interpolation? Would it make sense to test such an algorithm on a more general problem?

I also dont have a good understanding of how realistic the EEG experiments are.

I am also not sure important the N vs d result is for identifiability. It seems questionable to try estimate a covariance with fewer than dd observations anyway.

3 comments

We thank the reviewer for their careful summary. The question on general likelihoods led to a new result (point 2), which we add to the paper.

1. The problem is not academic. Estimating a covariance from a few samples, separately for each of many units, is routine in several fields:

  • finance: per-window covariances of hundreds of assets from a few months of returns;
  • EEG/BCI: per-trial covariances, the standard input of Riemannian classifiers;
  • fMRI: per-subject connectivity over 100–400 regions from a few hundred time points;
  • radar array processing: secondary-data covariances;
  • data assimilation: ensemble covariances from 50–100 members.

In each case the per-unit sample size is comparable to or smaller than dd. The downstream decision (a portfolio, a classifier, a filter gain) depends on the covariance, so its uncertainty matters. Shrinkage returns one matrix per unit and does not pool across units. Hierarchical inverse-Wishart priors pool, but they are rigid: in our synthetic benchmark they fall 1–10 nats below the oracle as soon as the prior leaves their family, where SC-SI stays within 0.6. Our method pools nonparametrically and returns a full posterior, from which a point estimate follows for any loss.

2. Beyond Gaussian samples. There are two distinct situations.

(i) A simulator of the true channel is available (e.g. heavy-tailed returns). The algorithm runs unchanged, because it only uses forward simulations of Ce∣CC_e\mid C. CeC_e is then no longer sufficient, which costs information but introduces no bias.

(ii) The data are non-Gaussian but treated as Gaussian. We now prove that the iteration minimizes F(π)=KL(μ⋆∥KNπ)F(\pi)=\mathrm{KL}(\mu_\star\|K_N\pi), where μ⋆\mu_\star is the actual law of CeC_e, and that Theorem 2 holds with π⋆\pi_\star replaced by a minimizer πˉ\bar\pi:

  • identity (7) becomes ∫KL(Pπˉ∥Pk) dμ⋆=KL(πˉ∥πk)−(Fk−F⋆)\int \mathrm{KL}(P_{\bar\pi}\|P_k)\,d\mu_\star=\mathrm{KL}(\bar\pi\|\pi_k)-(F_k-F^\star);
  • the rate becomes Fk−F⋆≤KL(πˉ∥π0)/(k+1)F_k-F^\star\le \mathrm{KL}(\bar\pi\|\pi_0)/(k+1);
  • for N≥dN\ge d and μ⋆\mu_\star with a density, πˉ\bar\pi is unique by Theorem 1, and πk⇒πˉ\pi_k\Rightarrow\bar\pi on the compact class.

The method is therefore consistent for the prior under which the Gaussian model best explains the observed CeC_e's, a nonparametric analogue of the pseudo-true parameter of a misspecified likelihood (White, 1982). Two cases show what πˉ\bar\pi is:

  • If the non-Gaussianity is a random scale shared within a unit, it is absorbed exactly: πˉ\bar\pi is the law of the rescaled covariance.
  • If the scale varies across samples with excess kurtosis κ\kappa, πˉ\bar\pi absorbs the extra dispersion. To second order this amounts to an effective sample size 2N/(2+κ)2N/(2+\kappa), the correction we use for financial returns.

For general likelihoods beyond covariances, we agree that the main difficulty is representation: one needs a summary statistic and an encoding of the latent parameter. Modi et al. (2026) treat the general black-box case. Our contribution is what the covariance structure makes possible: an exact identifiability threshold, a closed-form channel, and an exact spectral reduction.

3. Relation to Aubin-Frankowski et al. (2022). We build on their result that nonparametric EM is mirror descent with an O(1/k)O(1/k) rate in observation space, and we say so in the paper. We agree that (7) is in essence the classical EM decomposition. Its role here is to separate observation-space progress from prior recovery, now also under misspecification. What their framework leaves open, and what we add, is:

  • whether the prior is identifiable from the observations (Theorem 1);
  • when observation-space convergence implies recovery of the prior and posterior (Theorem 2(ii)–(iii));
  • how to perform the E-step, which has no closed form (Proposition 3);
  • the exact reduction to the spectrum under rotational invariance (Theorem 4).

4. Importance of N≥dN\ge d. The regime N<dN<d is the typical one in the applications above: 500 stocks with a year of daily data, or connectivity matrices from short scans. One might expect pooling many units (M→∞M\to\infty) to compensate for small NN. Theorem 1 shows it cannot: below dd, distinct smooth priors produce identical observation laws, whatever the number of units or the method. Proposition 5 shows that smaller classes remain recoverable there. In our finance experiment the forecast window already has fewer effective samples than assets.

5. EEG. The PhysioNet Motor Movement/Imagery dataset (109 subjects) is a standard BCI benchmark. Eyes-open versus eyes-closed is a classic alpha-band contrast, and window covariances are the standard features of Riemannian classifiers (Barachant et al., 2012). The setting is realistic but modest (d=8d=8). The revision evaluates it with five-fold subject-disjoint cross-validation over all 109 subjects.

We hope the new result and these clarifications strengthen the case for the paper.

AuthorsOpenReview ↗

revised PDF

The revised PDF is uploaded. Proposition 2′, on non-Gaussian samples treated as Gaussian, is stated at the end of §2, with the proof in Appendix C.4 and the effective-sample-size consequence used in §4.3. The contributions relative to Aubin-Frankowski et al. (2022) are stated in the introduction. We would be glad to answer any further question.

Dear Reviewer, as the discussion period closes on Nov 18, we wanted to check whether our reply and the revised PDF address your concerns. In particular, your question on likelihoods beyond the Gaussian case led to a new result (Proposition 2′, end of §2): the convergence guarantees hold for non-Gaussian samples treated as Gaussian, relative to the best prior under the Gaussian channel, which also justifies the effective sample sizes used on real data. We would be glad to clarify anything further before then. Thank you again for your review.

Comments on the paper

4 comments

We thank the reviewers for careful and constructive reports. Each reviewer receives a self-contained reply; this note summarizes the discussion.

What the reviewers valued.

  • The sharp identifiability threshold N≥dN\ge d (tCux, 8Jnv).
  • The separation between convergence in observation space and recovery of the latent prior (tCux, 8Jnv).
  • The exact spectral reduction, which keeps the full posterior uncertainty (8Jnv).
  • The self-consistent conditional interpolant as a practical E-step (RV5z).

Why the problem matters. Estimating a covariance from few samples, separately for each of many units, is routine: per-window covariances of asset returns, per-trial covariances in EEG/BCI, per-subject connectivity in fMRI, array processing, ensemble data assimilation. In all of these the per-unit sample size is comparable to the dimension, and the downstream decision (a portfolio, a classifier, a filter gain) depends on the covariance. Shrinkage returns one matrix per unit and does not pool across units; conjugate hierarchical priors pool but are rigid. Our method pools nonparametrically and returns a full posterior, and the identifiability threshold says exactly when this is possible. The Gaussian assumption is not restrictive either: the new result below shows that the method stays consistent for non-Gaussian samples.

What the reviewers questioned, and what we did to answer.

ConcernReviewersResolution
AR(1) prior described two ways8Jnv, tCuxA description error: data and oracle use one model, ρ∼U(0.3,0.9)\rho\sim\mathcal U(0.3,0.9). Text corrected, including the stated sample size (N=40N=40, not 28).
Finance score evaluated outside the support of its density (Nf<dN_f<d)8Jnv, tCuxRelabeled as a tempered likelihood. Two proper scores added (Gaussian and Student-tt), results in a follow-up. The portfolio and calibration results do not use the score.
EEG headline taken from a favourable 29-subject split8Jnv, tCuxThe main text now uses five-fold cross-validation over all 109 subjects. The log-score and calibration gains hold; the classification gain is smaller (+1.4 points) but significant.
Gap to the oracle on synthetic priorstCuxRetrained: every W1W_1, as a median over test tasks, is at most 0.40 posterior s.d. Outside the inverse-Wishart family, SC-SI is within 0.54 nats of the oracle, against 1.0–9.7 for IW-ML. Log-scores are now estimated with J=2⋅106J=2\cdot10^6 posterior draws instead of 256; the log-mean-exp estimator is biased at small JJ, which is why the IW-ML values also changed.
Spectral algorithm checked only on eigenvalue marginalstCuxNew rows of Table 1 validate Algorithm 2 on the completed matrix with orientation-dependent statistics. It is on par with the matrix version.
Setting seen as stylized; scope beyond Gaussian samples; relation to prior workRV5zRelevance argued above; new theoretical result (below). Contributions beyond Aubin-Frankowski et al. (2022) clarified.

New theoretical result. Suppose the samples are not Gaussian but are summarized by CeC_e and treated as Gaussian. The iteration then minimizes KL(μ⋆∥KNπ)\mathrm{KL}(\mu_\star\|K_N\pi) over priors, where μ⋆\mu_\star is the actual law of CeC_e. Theorem 2 holds verbatim with π⋆\pi_\star replaced by the minimizer πˉ\bar\pi, which is unique for N≥dN\ge d when μ⋆\mu_\star has a density. The method is therefore consistent for the best prior under the Gaussian channel, a nonparametric analogue of a pseudo-true parameter. The result also justifies the effective sample sizes used on real data, and it explains the over-sharpening with finite ensembles that motivates our checkpoint selection.

Revised paper. We will upload the revised PDF (10 pages) before the end of the discussion. It contains:

  • the corrected specifications;
  • the rerun synthetic benchmark (Fig. 2, Table 1) with Algorithm 2;
  • the five-fold EEG evaluation (Fig. 3, Table 4);
  • the relabeled finance score with the proper scores;
  • the new result, stated in §2 with the proof in the appendix.

Every main-text claim now matches the strongest validation we have.

Still in progress.

  • The finance proper scores, to be posted as a follow-up comment this week.
  • For the final version: ensemble-size and sampler-resolution controls that separate statistical from approximation error on the synthetic benchmark.
AuthorsOpenReview ↗

Revised paper and finance results

Revised paper and finance results

We have uploaded the revised PDF. It implements the changes described in our replies; the main text is 10 pages. Where to find them:

  • Proposition 2′ (non-Gaussian samples treated as Gaussian): end of §2, proof in Appendix C.4.
  • Synthetic benchmark: §4.1, Fig. 2, Table 1 (medians over tasks, with the rows for Algorithm 2), and Table 1′ in Appendix J.2 (means over tasks).
  • EEG, five-fold over all 109 subjects: §4.2, Fig. 3, Tables 2–3.
  • Finance: §4.3, Fig. 4, Tables 4–6.

Appendix lettering is unchanged. The former Appendix K (additional figures, including Fig. 7) is removed; its content is now in Fig. 2 and Table 1.

Finance: proper scores. As announced, we score the 21 forward daily returns under two predictives built from each posterior draw CjC_j (Fig. 4d, Table 6a). Gains are in nats per day over the sample covariance, with 95% month-block bootstrap intervals:

GaussianStudent-tt
OAS+0.51+0.21
LW-NLS+0.39+0.08
IW-ML+0.94+0.30
SC-SI+1.21 [0.97, 1.52]+0.48 [0.45, 0.52]

The Student-tt degrees of freedom are fitted on each fold's validation period and shared by all methods. In paired comparisons, SC-SI beats OAS by +0.27 [0.25, 0.29] and IW-ML by +0.18 [0.16, 0.20], and the advantage is positive in every volatility tercile and every test decade. The tempered score of the submission is kept in Fig. 4c, labeled as such.

For transparency: the finance network is now wider (1024) and its checkpoint is chosen by the Student-tt score on the validation period only. The earlier configuration (width 256, PIT selection) gives the same conclusions: a portfolio variance ratio to OAS of −0.030 [−0.039, −0.021] (now −0.032 [−0.041, −0.023]) and a Student-tt gain of +0.48 [0.44, 0.51].

Synthetic benchmark: what M=∞M=\infty means. Each outer iteration draws a fresh batch of 5000 observations, whereas the submission reused one ensemble of 5000. The gap to the oracle then measures the error of the algorithm rather than that of a finite ensemble, toward whose atomic maximum-likelihood prior exact EM over-sharpens (Proposition 2′). IW-ML is fitted on 5000 observations; refitting it on 40 times more data leaves its results unchanged. Finite-ensemble and sampler-resolution controls will be added to the final version.

Area_Chair_A29BOpenReview ↗

Author-Reviewer Discussion ends in 1 day, 12 hours

Dear authors and reviewers,

Thank you to the authors for their detailed responses, the revised manuscript (Nov 13) and the additional results posted since. The discussion period closes in about a day and a half, so this is the last chance for a direct exchange before the reviewer-AC discussion starts on Nov 19.

Reviewers RV5z and 8Jnv: I haven’t yet seen a reply from you. Please read the authors’ responses and the revision, which address several of your points directly. Please let us know whether your concerns have been addressed, and update your recommendation if your assessment has changed. Even a short acknowledgment helps.

Reviewer tCux: thank you for engaging. The authors answered your follow-up on Nov 17. Please let us know whether these settle your remaining points.

Authors: please be ready to answer any follow-up questions in the time left, focusing on the points most likely to affect the decision.

Thank you all for keeping the discussion constructive and timely.

Your AC

Thank you for coordinating the discussion. For convenience, here is what we posted after the revised PDF of Nov 13:

  • Finance proper scores (general comment): SC-SI has the best score under both a Gaussian and a Student-t predictive of the forward returns, with paired gains over OAS and IW-ML that are positive in every volatility regime and every test decade.
  • Sampler resolution (reply to tCux): halving or doubling the SDE step leaves the W1 distances to the oracle unchanged (at most 0.005) and the log-score within Monte Carlo error, so the sampler adds no measurable error.
  • Finite ensemble (reply to tCux): with a fixed ensemble of 5000 observations, checkpoint selection on held-out data, without the oracle, recovers the best posterior along the run; SC-SI then remains much closer to the oracle than IW-ML on the same data outside the conjugate family (median W1 0.36 vs 0.64 and 0.32 vs 0.79).

We are available to answer any further question before the discussion closes.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.