Pith. sign in

REVIEW 3 major objections 4 minor 70 references

The paper proves that factoring conditional generation through a sufficient latent representation moves the statistical bottleneck from the ambient dimension to the representation dimension.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 20:07 UTC pith:J4L3N7IM

load-bearing objection The KL/CMI decomposition and rate-separation idea are genuinely new and useful, but the paper's advertised guarantee assumes a fixed sufficient representation while the method learns one, and the missing proofs make the current version unverifiable. the 3 major comments →

arxiv 2607.16725 v1 pith:J4L3N7IM submitted 2026-07-18 stat.ML cs.LG

Semi-Supervised Conditional Generative Learning through Stochastic Interpolation and Sufficient Representations

Changyu Liu , Yuling Jiao , Jian Huang This is my paper
classification stat.ML cs.LG MSC 62G0562G0868T07
keywords conditional generative modelingsemi-supervised learningsufficient representationconditional mutual informationstochastic interpolationdeep neural networksminimax ratescurse of dimensionality
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes RepG, a two-stage conditional generative model built on stochastic interpolation. It wants to show that when data X and condition Y are linked through a low-dimensional sufficient representation R, so that Y→R→X, the high-dimensional conditional distribution can be learned from mostly unlabeled data: the label-dependent stage is confined to the low-dimensional pair (R,Y), and the label-free reconstruction stage X|R uses all samples. The main theorem decomposes the KL error into a structural bias equal to the conditional mutual information I(X;Y|R) plus two estimation errors, and the derived rates say that with enough unlabeled data the total error decays at a rate governed by k+q+1 rather than by the ambient d+q+1. A matching minimax lower bound for direct conditional SDE generation shows that this dimension reduction is what makes the improvement possible.

Core claim

On its own terms, the central discovery is an exact error decomposition (Theorem 2.1b): the expected KL divergence between the true joint conditional distribution and the RepG model equals I(X;Y|R), the structural bias from using the representation, plus the expected KL error in estimating P_{R|Y} plus the expected KL error in estimating P_{X|R}. Feeding non-asymptotic convergence rates for deep ReLU velocity estimators into this decomposition yields Corollary 4.2: if I(X;Y|R)=0 and the unlabeled sample size is at least roughly n^{(2α+d+k+1)/(2α+q+k+1)}−n, then the expected KL divergence is bounded by δ(ε)+n^{-2α/(2α+q+k+1)} poly(log n). The paper reads this as showing that the high-dimensio

What carries the argument

The load-bearing object is the Markov-chain factorization Y→R→X induced by a fixed deterministic sufficient representation R:R^d→R^k with X⊥⊥Y|R. The identity that carries the argument is the KL decomposition of Theorem 2.1(b), D_KL(P_{X,R|Y} ‖ P̂^{RepG}_{X,R|Y}) = I(X;Y|R) + E_Y D_KL(P_{R|Y} ‖ P̂_{R|Y}) + E_R D_KL(P_{X|R} ‖ P̂_{X|R}). The conditional mutual information term is the structural price of the representation, and the two stage errors are bounded by reducing conditional stochastic interpolation to least-squares velocity regression, then converting velocity estimates into SDE drift estimates through the Gaussian interpolation structure.

Load-bearing premise

The entire rate advantage rests on a fixed, low-dimensional representation R(X) that is truly sufficient, X⊥⊥Y|R; if no such representation exists, or if R is estimated from data rather than given, the conditional-mutual-information term I(X;Y|R) becomes a non-vanishing bias that can dominate the error no matter how many unlabeled samples are added.

What would settle it

Construct a distribution with a known sufficient scalar representation, e.g., X=(R, noise) with Y|R smooth, run RepG at n labeled and N unlabeled samples, and estimate the KL divergence to P_{X|Y}; if the error does not decay like n^{-2α/(2α+q+k+1)} once N exceeds the theorem's threshold, or if a distribution provably lacking any low-dimensional sufficient R still exhibits the advertised rate, the central claim is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If a sufficient representation exists, semi-supervised RepG achieves a KL rate n^{-2α/(2α+q+k+1)} poly(log n) once the unlabeled sample size is large enough, independent of the ambient dimension d.
  • Direct conditional SDE generation has a minimax KL lower bound of order n^{-2ω/(2ω+d+q+1)}, so the representation factorization is necessary for the improved dimension dependence.
  • Even in the fully supervised case N=0, the representation split yields a rate controlled by max{q,d}+k+1 instead of d+q+1, so the method is never worse in dimension than direct generation.
  • The conditional mutual information I(X;Y|R) is identified as the measurable cost of representation quality, making sufficiency a component of the statistical risk rather than an unverified assumption.
  • Empirically, MNIST digit-generation accuracy rises monotonically with the number of unlabeled samples, matching the theoretical prediction that unlabeled data pay for the reconstruction stage.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The theory treats R as fixed and pre-specified; in practice R is estimated from data, and the paper leaves open how estimation error in R inflates the CMI bias. A natural extension is to add a term measuring D_KL(P_{X|R} vs P_{X|R̂}) and to find regimes where plug-in representations still preserve the rate gain.
  • Corollary 4.2's threshold on unlabeled sample size can be read as an allocation rule: once N crosses that threshold, improving sufficiency (reducing I(X;Y|R)) is worth more than adding labels to the reconstruction stage.
  • The factorization argument is not tied to stochastic interpolation; any two-stage conditional generator with an exactly sufficient low-dimensional statistic should inherit the reduced dimension rate, so the analysis suggests a general design principle for semi-supervised conditional generation.
  • Because I(X;Y|R)=I(X;Y)−I(R(X);Y), maximizing the predictive information captured by R is equivalent to minimizing the structural bias; this points toward using estimated CMI as a practical representation-selection criterion, an option the paper mentions but does not develop.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes RepG, a two-stage semi-supervised conditional generative framework. The method assumes a fixed, pre-specified low-dimensional representation R(X) ∈ R^k such that X ⊥⊥ Y | R, factorizing the conditional density as p_{X|Y} = ∫ p_{X|R} p_{R|Y}. Stage 1 estimates the label-dependent latent distribution p_{R|Y} from n labeled pairs using conditional stochastic interpolation and a deep ReLU network; Stage 2 estimates the reconstruction distribution p_{X|R} from n labeled plus N unlabeled samples, thereby exploiting unlabeled data. The theoretical core is a KL error decomposition (Theorem 2.1) separating the error into a conditional-mutual-information structural bias and two estimation errors. The paper claims non-asymptotic rates for the velocity estimators (Theorem 4.2), an end-to-end KL bound (Theorem 4.4), a corollary showing that abundant unlabeled data reduces the effective dimension to k+q+1 (Corollary 4.2), and a minimax lower bound for direct conditional SDE generation in dimension d+q+1 (Theorem 4.5). Numerical experiments on synthetic data and MNIST illustrate the method, using a deep dimension reduction (DDR) representation learned from an auxiliary sample.

Significance. If the theoretical claims are correct, the paper makes a useful contribution to semi-supervised conditional generation: the KL decomposition cleanly identifies the role of representation sufficiency, and the rate comparison between upper bound (Corollary 4.2) and lower bound (Theorem 4.5) gives a formal sense in which a sufficient low-dimensional representation can circumvent the ambient curse of dimensionality. The idea of allocating the scarce labeled budget to the low-dimensional conditional latent sampler and using unlabeled data for the high-dimensional reconstruction is natural and practically relevant. The paper also includes explicit non-asymptotic rates and a minimax lower bound, which are welcome. However, as discussed below, the absence of proofs and the mismatch between the theoretical representation assumption and the algorithm actually implemented are serious issues that prevent the paper from being accepted in its current form.

major comments (3)
  1. [Entire Section 4] No proof is provided for any theorem, lemma, corollary, or proposition. This includes Theorem 2.1, Lemma 4.1, Corollary 4.1, Theorem 4.2, Lemma 4.3, Theorem 4.4, Corollary 4.2, Proposition 4.1, and Theorem 4.5. The statements are nontrivial: Lemma 4.1 relies on a Girsanov change of measure, Theorem 4.2 on FNN approximation under localized Hölder smoothness, and Theorem 4.5 on a minimax construction. Without proofs or even proof sketches, the central claims cannot be verified. A theoretical paper in a serious journal must include complete proofs in the main text or a supplement.
  2. [Section 2.1 vs. Sections 5.1–5.2] The theory is developed for a fixed, pre-specified deterministic representation R with X ⊥⊥ Y | R, and all rates are conditional on this exact sufficiency (I(X;Y|R)=0 in Corollary 4.2). The proposed algorithm, however, learns R from data using deep dimension reduction (DDR), trained on an auxiliary sample that is excluded from the labeled/unlabeled budgets n and N in the rates. Nothing in Theorem 2.1, Lemma 4.3, or Theorem 4.4 bounds I(X;Y|\hat R) or the estimation error of \hat R. For the realized learned representation, the CMI term in Theorem 2.1(b) is a structural bias that need not vanish as n,N → ∞, and if \hat R is only approximately sufficient this bias can dominate the advertised rate. Thus the central claim that RepG confines statistical complexity to k+q+1 is not established for the algorithm as implemented and tested. The paper should either extend the theory to estimated rep
  3. [Corollary 4.1 and Theorem 4.4] Corollary 4.1 shows that the KL divergence for the discretized SDE contains an additional time-discretization term, which the authors state becomes negligible for a sufficiently fine grid. However, Theorem 4.4 — the main end-to-end bound — contains no discretization error term at all, and no quantitative condition on H, ε, or the grid spacing is given. Algorithm 1 simulates both SDEs with H steps, but the theorem applies only to the idealized continuous-time processes in (3.7)–(3.8). To claim a convergence guarantee for the algorithm actually used, the discretization error must be carried into the final bound, at least as an additive term that is controlled by taking H large relative to n,N.
minor comments (4)
  1. [Section 4.2.1, Lemma 4.3] The quantity δ(ε) is introduced as I(R;Y) − I(R′;Y) and is stated to decrease to 0 as ε↓0, but no quantitative rate is given. Since the algorithm stops at a fixed ε>0, the bound in Theorem 4.4 contains an unquantified δ(ε). Please provide at least a bound in terms of ε, or state explicitly that δ(ε) is treated as an arbitrarily small early-stopping bias.
  2. [Section 1, Introduction] The abstract and introduction refer to 'sufficient representation' and 'the underlying assumption of perfect sufficiency holds only approximately', but the precise meaning of approximate sufficiency is not formalized until Theorem 2.1 via CMI. Consider defining the CMI criterion earlier to avoid ambiguity.
  3. [Section 5.2] Typo: 'the labeled date size' should be 'the labeled data size'. Also, the training details for the DDR representation (architecture, optimization, whether the auxiliary sample is labeled) are under-specified; please provide enough detail to reproduce the representation-learning step.
  4. [Theorem 4.5 and Remark 1] The +1 dimensionality penalty in the lower bound (4.3) versus the classical density-estimation rate is attributed to the time coordinate. This is an interesting claim, but it deserves a more careful discussion linking the path-space KL to the terminal marginal KL, since the lower bound is stated for path measures while the upper bounds in Proposition 4.1 concern early-stopped marginals.

Circularity Check

0 steps flagged

No circularity: the rate advantage is conditional on an explicitly assumed sufficient representation; estimation rates follow from standard nonparametric regression, not from fitted constants.

full rationale

The paper's derivation chain is self-contained and non-circular. Theorem 2.1 is an exact identity: the RepG KL error equals I(X;Y|R) plus the two velocity-estimation errors, obtained by expanding the conditional KL and using the Markov-chain factorization; it does not define the error to be what it later predicts. Theorem 4.2 gives FNN regression rates for the two velocity fields under stated Hölder/subexponential assumptions, using standard approximation arguments. Theorem 4.4 combines these with Girsanov path-space bounds; Corollary 4.2 explicitly assumes I(X;Y|R)=0 and N sufficiently large, so the advertised n^{-2α/(2α+q+k+1)} rate is a conditional upper bound rather than a fitted parameter renamed as a prediction. The lower bound Theorem 4.5 is a minimax argument over a direct SDE class and does not depend on the paper's own estimators. Citations to the authors' prior work (Huang et al. 2023, 2024; Gao et al. 2024) provide the stochastic-interpolation and DDR constructions, but the main theorems do not reduce to those citations: the interpolation identity is also attributed to the external stochastic-interpolants literature (Albergo et al. 2025), and DDR is used only for experiments/motivation. The gap identified by the skeptic—that the method learns R while the theory fixes R—is an explicit idealization (Section 2.1: 'a fixed, deterministic representation mapping that is pre-specified independently of the observed data') and is a scope limitation, not a circularity: the CMI bias I(X;Y|R) is honestly retained in Theorem 2.1(b) and Theorem 4.4. Therefore no circular step is exhibited.

Axiom & Free-Parameter Ledger

4 free parameters · 8 axioms · 0 invented entities

The central result rests on a strong structural assumption (sufficiency of a low-dimensional pre-specified R), standard interpolation/NN approximation machinery, and the unlabeled-data sampling assumption. No free parameters are fitted in the theory, but the empirical pipeline requires choosing k, ε, η, network sizes, and the representation estimator.

free parameters (4)
  • Latent representation dimension k = k=32 (MNIST); k=2/3 (simulations)
    User-chosen dimension of the representation; all rate advantages rely on k≪d, but no data-driven selection or sensitivity analysis for k is reported.
  • Early stopping parameter ε = not reported
    Controls truncation bias δ(ε) in Lemma 4.3 and Theorem 4.4; the theory requires ε↓0 but the exact value used in Algorithm 1 is not stated.
  • Diffusion level η = η=0.2 in MNIST experiments
    Pre-specified in the SDE; enters constants C_ε and the discretization error, but is not fitted in the theory.
  • Network width/depth (W1,L1,W2,L2) = not reported; Theorem 4.2 sets W1L1≍n^(...) log n and W2L2≍(n+N)^(...) log(n+N)
    The rates hold only for the specified width-depth scaling; practical architecture choices are not tied to these constants.
axioms (8)
  • domain assumption Assumption 1: interpolation schedules a_t,b_t are C^1 with boundary conditions a0=1,b0=0,a1=0,b1=1, monotone, bounded derivatives.
    Needed to define the Gaussian interpolation path and the plug-in drift (3.4); the linear rectified-flow schedule is the example used.
  • domain assumption Assumption 2: velocity fields are locally Hölder over compact hypercubes with Hölder norm growing at most polynomially in the radius.
    Required for the FNN approximation rates in Theorem 4.2; excludes velocity fields with stronger-than-polynomial tail growth.
  • domain assumption Assumption 3: subexponential moment bounds on X, R, Y.
    Controls tail contributions when truncating the target domain to a compact hypercube in the statistical analysis.
  • domain assumption Exponential integrability condition (4.1) for the true velocity field under the path measure.
    Needed to apply Girsanov's theorem in Lemma 4.1 and ensure absolute continuity of path measures.
  • domain assumption Existence of a fixed, deterministic, low-dimensional sufficient representation R with X⊥⊥Y|R (Section 2.1).
    This is the load-bearing structural assumption: if it fails, the CMI term I(X;Y|R) is a non-vanishing bias and the advertised rate advantage is lost.
  • domain assumption Unlabeled sample D_u is drawn from the same marginal P_X as the labeled sample.
    Justifies pooling labeled and unlabeled data in the second-stage augmented dataset D_all.
  • domain assumption The DDR estimator (Huang et al., 2024) recovers a sufficient representation in the experiments.
    Used in Section 5 to construct R; the paper does not prove finite-sample sufficiency of the learned DDR representation or account for its estimation error.
  • standard math Standard results: Girsanov theorem, FNN approximation capacity (Schmidt-Hieber, Stone), stochastic interpolant drift characterization.
    Explicitly or implicitly invoked in Lemmas 4.1, Theorem 4.2, and the plug-in drift construction; treated as known background.

pith-pipeline@v1.3.0-alltime-deepseek · 17349 in / 20623 out tokens · 191965 ms · 2026-08-01T20:07:44.090140+00:00 · methodology

0 comments
read the original abstract

Conditional generative modeling remains a challenging problem in semi-supervised settings where labeled data is scarce but unlabeled samples are abundant. To effectively leverage structural information embedded within the unlabeled dataset and compensate for sparse conditioning signals, we propose a semi-supervised framework combining conditional stochastic interpolation with low-dimensional latent representations. RepG decomposes generation into two stages: label-dependent latent sampling and high-dimensional reconstruction. This isolates the supervised learning of conditional dependencies to a low-dimensional space, requiring few labels while utilizing the abundant unlabeled data purely for reconstruction. Theoretically, we establish an error decomposition showing that the Kullback-Leibler divergence of RepG comprises stage-wise estimation errors and a structural bias quantified by conditional mutual information. For deep neural network estimators, we derive non-asymptotic convergence rates proving that RepG significantly improves sample complexity. By confining the supervised estimation burden to the low intrinsic dimension of the latent representation, RepG achieves a strictly faster convergence rate. Complemented by a minimax lower bound, our theoretical results demonstrate that this method effectively mitigates the curse of dimensionality inherent in direct ambient-space generative modeling.

Figures

Figures reproduced from arXiv: 2607.16725 by Changyu Liu, Jian Huang, Yuling Jiao.

Figure 1
Figure 1. Figure 1: Generative performance of RepG across four distinct scenarios. The top row [PITH_FULL_IMAGE:figures/full_fig_p027_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Real data (blue) and RepG generated samples (orange) across three distinct [PITH_FULL_IMAGE:figures/full_fig_p029_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of classifier accuracy between RepG (red) and baseline (green) against [PITH_FULL_IMAGE:figures/full_fig_p031_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of generation quality: (a) left panel [PITH_FULL_IMAGE:figures/full_fig_p032_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

70 extracted references · 4 linked inside Pith

  1. [1]

    Dennis , title =

    Cook, R. Dennis , title =. Communications in Statistics - Theory and Methods , year =

  2. [2]

    2006 , publisher =

    Gaussian Processes for Machine Learning , author =. 2006 , publisher =

    2006

  3. [3]

    , title =

    Bishop, Christopher M. , title =. 1995 , publisher =

    1995

  4. [4]

    Journal of Machine Learning Research , year =

    Meinshausen, Nicolai , title =. Journal of Machine Learning Research , year =

  5. [5]

    International Conference on Learning Representations , year=

    Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions , author=. International Conference on Learning Representations , year=

  6. [6]

    Nishant Jain and Tong Zhang , booktitle=. A Sharp

  7. [7]

    2009 , booktitle =

    Lyu, Siwei , title =. 2009 , booktitle =

    2009

  8. [8]

    International Conference on Learning Representations , year=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. International Conference on Learning Representations , year=

  9. [9]

    International Conference on Learning Representations , year=

    Flow Matching for Generative Modeling , author=. International Conference on Learning Representations , year=

  10. [10]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , year=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , year=

  11. [11]

    International Conference on Learning Representations , year=

    Solving Inverse Problems in Medical Imaging with Score-Based Generative Models , author=. International Conference on Learning Representations , year=

  12. [12]

    Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding , author=. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2019

  13. [13]

    NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications , year=

    Classifier-Free Diffusion Guidance , author=. NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications , year=

    2021

  14. [14]

    Proceedings of the 38th International Conference on Machine Learning , pages =

    Learning Transferable Visual Models From Natural Language Supervision , author =. Proceedings of the 38th International Conference on Machine Learning , pages =. 2021 , volume =

    2021

  15. [15]

    International Conference on Learning Representations , year =

    Auto-Encoding Variational Bayes , author =. International Conference on Learning Representations , year =

  16. [16]

    International Conference on Learning Representations , year=

    Score-Based Generative Modeling through Stochastic Differential Equations , author=. International Conference on Learning Representations , year=

  17. [17]

    Chen, Ricky T. Q. and Rubanova, Yulia and Bettencourt, Jesse and Duvenaud, David , title =. Advances in Neural Information Processing Systems , year =

  18. [18]

    Advances in Neural Information Processing Systems , year =

    Denoising Diffusion Probabilistic Models , author =. Advances in Neural Information Processing Systems , year =

  19. [19]

    Diffusion Models Beat

    Prafulla Dhariwal and Alexander Quinn Nichol , booktitle=. Diffusion Models Beat

  20. [20]

    Advances in Neural Information Processing Systems , year=

    Score-based Generative Modeling in Latent Space , author=. Advances in Neural Information Processing Systems , year=

  21. [21]

    Advances in Neural Information Processing Systems , year=

    On the Relation between Rectified Flows and Optimal Transport , author=. Advances in Neural Information Processing Systems , year=

  22. [22]

    Elucidating Flow Matching

    Zhengchao Wan and Qingsong Wang and Gal Mishne and Yusu Wang , booktitle=. Elucidating Flow Matching

  23. [23]

    International Conference on Learning Representations , year=

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author=. International Conference on Learning Representations , year=

  24. [24]

    International Conference on Learning Representations , year=

    Provable Memorization Capacity of Transformers , author=. International Conference on Learning Representations , year=

  25. [25]

    Bert: Pre-training of deep bidirectional transformers for language understanding , author=. Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , year=

    2019

  26. [26]

    2023 , booktitle =

    Chen, Minshuo and Huang, Kaixuan and Zhao, Tuo and Wang, Mengdi , title =. 2023 , booktitle =

    2023

  27. [27]

    Advances in Neural Information Processing Systems , year=

    Generative adversarial nets , author=. Advances in Neural Information Processing Systems , year=

  28. [28]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , year=

    Taming transformers for high-resolution image synthesis , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , year=

  29. [29]

    International Conference on Learning Representations , year=

    Building normalizing flows with stochastic interpolants , author=. International Conference on Learning Representations , year=

  30. [30]

    Proceedings of the 40th International Conference on Machine Learning , year =

    Improved Analysis of Score-based Generative Modeling: User-Friendly Bounds under Minimal Smoothness Assumptions , author =. Proceedings of the 40th International Conference on Machine Learning , year =

  31. [31]

    and Wasserman, Larry , title =

    Lei, Jing and G'Sell, Max and Rinaldo, Alessandro and Tibshirani, Ryan J. and Wasserman, Larry , title =. Journal of the American Statistical Association , volume =

  32. [32]

    , title =

    Gneiting, Tilmann and Raftery, Adrian E. , title =. Journal of the American Statistical Association , volume =

  33. [33]

    arXiv preprint arXiv:2512.18971 , year=

    On Conditional Stochastic Interpolation for Generative Nonlinear Sufficient Dimension Reduction , author=. arXiv preprint arXiv:2512.18971 , year=

    arXiv

  34. [34]

    arXiv preprint arXiv:2403.11968 , year=

    Unveil conditional diffusion models with classifier-free guidance: A sharp statistical theory , author=. arXiv preprint arXiv:2403.11968 , year=

    Pith/arXiv arXiv

  35. [35]

    Advances in Neural Information Processing Systems , volume=

    Understanding scaling laws with statistical and approximation theory for transformer neural networks on intrinsically low-dimensional data , author=. Advances in Neural Information Processing Systems , volume=

  36. [36]

    and Boffi, Nicholas M

    Albergo, Michael S. and Boffi, Nicholas M. and Vanden-Eijnden,Eric , TITLE =. Journal of Machine Learning Research , VOLUME =. 2025 , PAGES =

    2025

  37. [37]

    The Annals of Statistics , VOLUME =

    Lee, Kuang-Yao and Li, Bing and Chiaromonte, Francesca , TITLE =. The Annals of Statistics , VOLUME =. 2013 , PAGES =

    2013

  38. [38]

    Journal of Machine Learning Research , volume=

    Gaussian interpolation flows , author=. Journal of Machine Learning Research , volume=

  39. [39]

    Journal of the American Statistical Association , volume =

    Jianqing Fan and Yihong Gu , title =. Journal of the American Statistical Association , volume =

  40. [40]

    Measuring and testing dependence by correlation of distances , FJOURNAL =

    Sz\'. Measuring and testing dependence by correlation of distances , FJOURNAL =. The Annals of Statistics , VOLUME =. 2007 , PAGES =

    2007

  41. [41]

    van Erven, Tim and Harremos, Peter , journal=. R. 2014 , volume=

    2014

  42. [42]

    The Annals of Statistics , VOLUME =

    Kohler, Michael and Langer, Sophie , TITLE =. The Annals of Statistics , VOLUME =. 2021 , PAGES =

    2021

  43. [43]

    and Bottou, L

    LeCun, Y. and Bottou, L. and Bengio, Y. and Haffner, P. , journal=. Gradient-based learning applied to document recognition , year=

  44. [44]

    Hierarchical text-conditional image generation with

    Ramesh, Aditya and Dhariwal, Prafulla and Nichol, Alex and Chu, Casey and Chen, Mark , journal=. Hierarchical text-conditional image generation with

  45. [45]

    arXiv preprint arXiv:2511.03193 , year=

    Statistical Properties of Rectified Flow , author=. arXiv preprint arXiv:2511.03193 , year=

    arXiv

  46. [46]

    The Annals of Statistics , volume=

    Deep nonparametric regression on approximate manifolds: Nonasymptotic error bounds with polynomial prefactors , author=. The Annals of Statistics , volume=

  47. [47]

    and Gretton, Arthur and Fukumizu, Kenji and Sch\"

    Sriperumbudur, Bharath K. and Gretton, Arthur and Fukumizu, Kenji and Sch\". Hilbert space embeddings and metrics on probability measures , FJOURNAL =. Journal of Machine Learning Research , VOLUME =. 2010 , PAGES =

    2010

  48. [48]

    On the rate of convergence of a classifier based on a transformer encoder , FJOURNAL =

    Gurevych, Iryna and Kohler, Michael and. On the rate of convergence of a classifier based on a transformer encoder , FJOURNAL =. Institute of Electrical and Electronics Engineers. Transactions on Information Theory , VOLUME =. 2022 , PAGES =

    2022

  49. [49]

    The Annals of Statistics , VOLUME =

    Chen, YinFeng and Jiao, YuLing and Qiu, Rui and Yu, Zhou , TITLE =. The Annals of Statistics , VOLUME =. 2024 , PAGES =

    2024

  50. [50]

    Journal of the American Statistical Association , VOLUME =

    Li, Ker-Chau , TITLE =. Journal of the American Statistical Association , VOLUME =. 1991 , PAGES =

    1991

  51. [51]

    and Harvey, Nick and Liaw, Christopher and Mehrabian, Abbas , title =

    Bartlett, Peter L. and Harvey, Nick and Liaw, Christopher and Mehrabian, Abbas , title =. Journal of Machine Learning Research , year =

  52. [52]

    Deep Dimension Reduction for Supervised Representation Learning , year=

    Huang, Jian and Jiao, Yuling and Liao, Xu and Liu, Jin and Yu, Zhou , journal=. Deep Dimension Reduction for Supervised Representation Learning , year=

  53. [53]

    , TITLE =

    Stone, Charles J. , TITLE =. The Annals of Statistics , VOLUME =. 1982 , PAGES =

    1982

  54. [54]

    Lu, Jianfeng and Shen, Zuowei and Yang, Haizhao and Zhang, Shijun , TITLE =. SIAM J. Math. Anal. , FJOURNAL =. 2021 , NUMBER =

    2021

  55. [55]

    The Annals of Statistics , VOLUME =

    Schmidt-Hieber, Johannes , TITLE =. The Annals of Statistics , VOLUME =. 2020 , PAGES =

    2020

  56. [56]

    arXiv preprint arXiv:2603.04223 , year=

    Semi-Supervised Generative Learning via Latent Space Distribution Matching , author=. arXiv preprint arXiv:2603.04223 , year=

    arXiv

  57. [57]

    Neural Networks , volume =

    Dmitry Yarotsky , title =. Neural Networks , volume =

  58. [58]

    arXiv preprint arXiv:2404.02538 , year=

    Convergence analysis of flow matching in latent space with transformers , author=. arXiv preprint arXiv:2404.02538 , year=

    Pith/arXiv arXiv

  59. [59]

    Latent Schr

    Jiao, Yuling and Kang, Lican and Lin, Huazhen and Liu, Jin and Zuo, Heng , journal=. Latent Schr

  60. [60]

    arXiv preprint arXiv:2402.01460 , year=

    Deep conditional distribution learning via conditional F " ollmer flow , author=. arXiv preprint arXiv:2402.01460 , year=

    arXiv

  61. [61]

    arXiv preprint arXiv:2312.05579 , year=

    Conditional stochastic interpolation for generative learning , author=. arXiv preprint arXiv:2312.05579 , year=

    Pith/arXiv arXiv

  62. [62]

    arXiv preprint arXiv:2404.00551 , year=

    Convergence of continuous normalizing flows for learning probability distributions , author=. arXiv preprint arXiv:2404.00551 , year=

    Pith/arXiv arXiv

  63. [63]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Representation learning: A review and new perspectives , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2013 , doi=

    2013

  64. [64]

    and Baraniuk, Richard G

    Ongie, Gregory and Jalal, Ajil and Metzler, Christopher A. and Baraniuk, Richard G. and Dimakis, Alexandros G. and Willett, Rebecca , journal=. Deep Learning Techniques for Inverse Problems in Imaging , year=

  65. [65]

    ExDiffusion: Classifier-Guidance Diffusion Model for Extreme Load Scenario Generation With Extreme Value Theory , year=

    Yang, Ruizhang and Li, Yujia and Hu, Chenxi and Hou, Francis Yunhe , journal=. ExDiffusion: Classifier-Guidance Diffusion Model for Extreme Load Scenario Generation With Extreme Value Theory , year=

  66. [66]

    2003 , edition =

    Stochastic Differential Equations: An Introduction with Applications , publisher =. 2003 , edition =

    2003

  67. [67]

    2002 , publisher=

    A Distribution-Free Theory of Nonparametric Regression , author=. 2002 , publisher=

    2002

  68. [68]

    Brownian Motion, Martingales, and Stochastic Calculus , SERIES =

    Le Gall, Jean-Fran. Brownian Motion, Martingales, and Stochastic Calculus , SERIES =. 2016 , PAGES =

    2016

  69. [69]

    , TITLE =

    Anthony, Martin and Bartlett, Peter L. , TITLE =. 1999 , PAGES =

    1999

  70. [70]

    , TITLE =

    Tsybakov, Alexandre B. , TITLE =