REVIEW 4 major objections 4 minor 39 references
Likelihood Training of Cascaded Diffusion Models via Hierarchical Volume-preserving Maps
T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read Cascaded diffusion models can be trained and evaluated by exact likelihood, without marginalizing their intermediate scales.
desk verdict The wavelet half of this paper is a real result; the Laplacian pyramid half rests on a change-of-variables misstep, and the reader's verdict is about right. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the hierarchical volume-preserving map, a smooth invertible multi-scale transformation $h$ whose Jacobian $A$ satisfies $\sqrt{\det(A^T A)} = 1$ at every point. The identity that carries the argument is the probabilistic invariance $\log p_\theta(x) = \log p_\theta(h(x))$, obtained from the change-of-variables formula for matrix volume. The paper shows that two standard decompositions are instances: the Laplacian pyramid, whose norm-preserving property makes it a tight frame, and the Haar wavelet transform, which is orthogonal. The role of this machinery is to convert the joint likelihood of a cascaded diffusion model into the data likelihood without any marginalization step.
What would settle it
On a small image size, compute the cascade's joint likelihood via the paper's formula and also by brute-force numerical marginalization over the intermediate Laplacian-pyramid scale; a systematic discrepancy in the LP case would show that the equality in Lemma 4.1 does not hold for the ambient likelihood.
Extended reading notes
Core claim
On the paper's own terms, the central discovery is Lemma 4.1: for any hierarchical volume-preserving map $h$, $\log p_\theta(x) = \log p_\theta(h(x))$. Because $h$ maps the image $x$ into the stacked scales $(z^{(1)}, \ldots, z^{(S)})$, the intractable marginalization over intermediate scales is unnecessary: the cascaded model's joint likelihood is exactly the data likelihood. The volume-preserving condition, $\sqrt{\det(A^T A)} = 1$ for the Jacobian $A$, is satisfied by at least two classical decompositions: the Laplacian pyramid, as a tight frame satisfying Parseval's equality, and the orthogonal Haar wavelet transform. Training such a cascaded diffusion model by maximizing the joint likelihood therefore amounts to maximum likelihood on the data. The paper further claims that this training loss controls a weighted sum of Earth Mover's Distances between the true and modeled score functions, giving a perceptual interpretation of the likelihood signal.
Load-bearing premise
The equality between the data likelihood and the joint scale likelihood assumes the model's probability density is supported on the image of the transform; for the Laplacian pyramid, which maps into a higher-dimensional space, the diffusion model can assign probability mass off that image, where the change-of-variables formula no longer applies.
Editorial extensions
If this is right
- Cascaded diffusion models can be trained by a variational bound whose terms sum to the log-likelihood of the data, making likelihood-based model selection, compression, and out-of-distribution scoring available in multi-scale settings.
- Laplacian pyramid and wavelet decompositions give exact likelihood preservation while remaining linear and cheap, so the method adds no expensive marginalization or auxiliary networks.
- The same training loss bounds the Earth Mover's Distance between true and estimated scores, suggesting that hierarchical likelihood training implicitly optimizes a perceptually relevant transport cost.
- On the benchmarks reported, the model attains lower bits-per-dimension than previous likelihood models on CIFAR10 and ImageNet 32, 64, and 128, and better AUROC for out-of-distribution detection.
- The lossless compression results follow from using the learned likelihood as an entropy model in bits-back coding.
Reading between the lines
- If Lemma 4.1 is extended to other tight-frame or orthogonal decompositions, the same exact-likelihood construction could apply beyond Laplacian pyramids and Haar wavelets, for example to steerable pyramids or higher-order wavelets; this is my extrapolation, not a claim in the paper.
- The Earth Mover's Distance bound being computable in linear time in the hierarchical basis suggests the framework could be tested as a fast perceptual training signal; one could compare sample quality under human preference or perceptual metrics with and without the hierarchical reparameterization, holding architecture constant.
- The overcompleteness of the Laplacian pyramid means the equality in Lemma 4.1 may hold as a density on the image submanifold rather than on the full ambient space; if the diffusion model assigns mass off this submanifold, a small-scale brute-force marginalization would reveal a discrepancy between the reported joint likelihood and the true data likelihood.
- A direct way to put the central claim under pressure is to measure whether the reported bits-per-dimension gap over single-scale variational diffusion persists when both models are restricted to the same parameter budget; the paper compares against published numbers rather than controlled same-budget baselines.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes hierarchical volume-preserving maps (HVPMs) as a way to make cascaded diffusion likelihoods tractable, claiming in Lemma 4.1 that for any HVPM h, log pθ(x) = log pθ(h(x)). It instantiates HVPMs with the Laplacian pyramid (overcomplete) and the Haar wavelet (orthogonal), builds PCDM variants (LP-PCDM and W-PCDM), reports state-of-the-art bits-per-dimension on CIFAR-10 and ImageNet 32/64/128, OOD detection AUROC, and lossless compression results, and derives an Earth Mover's Distance score-matching interpretation in Theorem 5.1. The orthogonal wavelet chain is a correct application of the change-of-variables formula; the Laplacian-pyramid branch is not, because the map is overcomplete and the model density is defined on an ambient space of higher dimension.
Significance. If the claims were restricted to the orthogonal wavelet case, the paper would make a clean and useful contribution: orthogonal multi-scale transforms let a cascaded diffusion model treat the joint likelihood over scales as the data likelihood, avoiding scale-level marginalization, with a coherent connection to wavelet-domain EMD score matching. The W-PCDM empirical results are strong, the code is released, and the theoretical derivation contains no fitted constants and no self-citation. However, the paper's headline contribution explicitly includes the Laplacian pyramid as a second valid HVPM, and LP-PCDM appears throughout Tables 1-3 and the EMD theorem. Because Lemma 4.1 is invalid for overcomplete maps, the LP likelihood claims, the LP rows of the empirical tables, and the LP version of Theorem 5.1 are unsupported. The paper would need substantial re-scoping to retain only the wavelet-based claims.
major comments (4)
- [Lemma 4.1 and Appendix A, Eq. (20)] The proof of Lemma 4.1 cites Ben-Israel's change-of-variables formula in Appendix A, Eq. (20). That formula is an area formula: it relates the density of X to a density of Y = h(X) with respect to Hausdorff measure on the image manifold, not to an ambient Lebesgue density. In the Laplacian-pyramid case h maps R^d into R^M with M > d (Section 4.2), while pθ in Section 5.1 is a diffusion density on the full product space Z^(1)×...×Z^(S), with Gaussian noise in every coordinate, including the M-d normal directions to h(R^d). Therefore pθ(h(x)) is not a density in x and Eq. (9) is not justified. The failure is concrete: for h(x) = (x,0), which also satisfies sqrt(det(A^T A)) = 1, and pθ = N(0,I_2), the asserted equality would require ∫ pθ(h(x)) dx = 1, but the integral is (2π)^(-1/2). Lemma 4.1 is valid in the equidimensional Haar-wavelet case, where Hausdorff measure and Lebesgue measure coincide, but not for the overcomplete Laplacian pyramid.
- [Definition 1 and Section 4.2] Definition 1 defines h as a homeomorphism but simultaneously permits dim(Z) > dim(X); a homeomorphism between open sets of Euclidean spaces preserves dimension. The Laplacian pyramid is invertible only as a reconstruction map on the d-dimensional image h(R^d), not as a bijection R^d → R^M. This is not merely a wording problem: it encodes the same conflation of a density supported on a submanifold with a density on the ambient space that breaks Lemma 4.1 for LP.
- [Theorem 5.1 and Appendix A, Eq. (31)] For the LP case, the proof of Theorem 5.1 asserts in Eq. (31) that h(∇x log q) = (∇_{z(1)} log q, ..., ∇_{z(S)} log q). For an overcomplete tight frame, applying h to a vector in R^d produces a vector in R^M, whereas the right-hand side is a vector of per-scale gradients whose total dimension is M; the two can be identified only in the orthogonal (wavelet) case, via ∇x = h^T ∇z. The later norm comparisons in the same proof compare ||z_lp^s||_2 with ||z_w^s||_2, i.e., norms of the coefficients themselves, not of the score residuals; hence the claim that the LP score-matching loss equals the wavelet loss does not follow. The LP version of Theorem 5.1 is therefore unsupported.
- [Appendix A, Eqs. (21)-(28)] The proof of Theorem 5.1 uses inconsistent signs for the ELBO components. In Eq. (21), L0, LT, and Lk are defined as log pθ, -KL, and -KL, respectively, so L0 and LT are not nonnegative as claimed: L0 is not guaranteed nonnegative and LT is nonpositive. Additionally, C(x) in Eq. (17) is defined as the sum of these ELBO terms, yet the displayed inequality αC(x) ≥ (a nonnegative expectation) only makes sense if C is a negative loss. As written, the proof does not establish the theorem; either C must be redefined or the signs of the loss components must be corrected.
minor comments (4)
- [Section 3, Eq. (5)] The expression contains 'log KL(...)' in the sum; this should be 'KL(...)' (or the corresponding loss with the opposite sign). This typo contributes to the sign ambiguity discussed above.
- [Throughout] There are several typos: 'must must', 'likelihoood', 'perpespective', and 'V AEs' should be corrected.
- [Sections 5.1 and 6.1] The paper should state clearly that the reported BPD values are variational upper bounds obtained from C(x) in Eq. (17), and that the OOD numbers use a 20-sample Monte Carlo estimate of the bound (Appendix B), rather than exact log-likelihoods; the abstract's 'exact likelihood' wording is stronger than what is evaluated.
- [Appendix A, Eqs. (35)-(39)] In the chain of inequalities for Lemma A.2, the summation index for the higher scales is written as ∇z(1) in several places; it should be ∇z(s).
Circularity Check
No significant circularity: the derivation reduces to external mathematical results, not to its own fitted inputs or self-citations.
full rationale
The paper's central claim, Lemma 4.1, is derived from the Ben-Israel (1999) change-of-variables formula in Appendix A, an external mathematical result. The equality log pθ(x) = log pθ[h(x)] follows from the volume-preserving condition sqrt(det(A^T A)) = 1; the condition is a premise, and the conclusion is a direct application of the cited change-of-variables formula. This is ordinary derivation, not a self-referential reduction. The EMD result, Theorem 5.1, relies on the Shirdhonkar and Jacobs (2008) linear-time approximation to the Earth Mover's Distance, also an external result, plus standard diffusion likelihood bounds from Ho et al. (2020) and Kingma et al. (2021), none of which are authored by the present paper's authors. No fitted constants are renamed as predictions: the models are trained with standard maximum-likelihood losses and evaluated against external benchmarks, and the reported BPD numbers are not recovered from any fitted parameter of the theory. There are effectively no self-citations in the reference list, and no uniqueness theorem or ansatz is imported from the authors' prior work. The skeptic's concern about the Laplacian pyramid being overcomplete, so that a Hausdorff density on the image of h is silently treated as an ambient Lebesgue density, is a mathematical correctness objection about the change-of-variables step, not a circularity objection: it does not identify a claim that is defined in terms of its own conclusion or a fitted input masquerading as a prediction. Because the paper is self-contained against external benchmarks and its theoretical claims reduce to prior, independently published mathematics, the circularity score is 0.
Assumptions & free parameters
assumptions (5)
- ad hoc to paper A density pθ defined on the ambient coordinate space R^M can be identified with a density on the image of an overcomplete map h via pθ(x)=pθ(h(x)).
- standard math The generalized change-of-variables formula with matrix volume (Ben-Israel 1999) applies to the model density as used.
- domain assumption The forward diffusion marginals are Gaussian, so score matching terms take the weighted L2 form.
- domain assumption The Shirdhonkar-Jacobs wavelet equivalence for EMD (Lemma A.1) applies to score functions.
- domain assumption The terms L0 and LT are nonnegative and can be dropped when bounding C(x).
Cite this review
Pith. "Pith review of Likelihood Training of Cascaded Diffusion Models via Hierarchical Volume-preserving Maps." pith.science (2026). https://pith.science/paper/PB62FUAC
@misc{pith2026250106999,
author = {Pith},
title = {Pith review of: Likelihood Training of Cascaded Diffusion Models via Hierarchical Volume-preserving Maps},
year = {2026},
howpublished = {\url{https://pith.science/paper/PB62FUAC}},
note = {Machine review of arXiv:2501.06999}
}
read the original abstract
Cascaded models are multi-scale generative models with a marked capacity for producing perceptually impressive samples at high resolutions. In this work, we show that they can also be excellent likelihood models, so long as we overcome a fundamental difficulty with probabilistic multi-scale models: the intractability of the likelihood function. Chiefly, in cascaded models each intermediary scale introduces extraneous variables that cannot be tractably marginalized out for likelihood evaluation. This issue vanishes by modeling the diffusion process on latent spaces induced by a class of transformations we call hierarchical volume-preserving maps, which decompose spatially structured data in a hierarchical fashion without introducing local distortions in the latent space. We demonstrate that two such maps are well-known in the literature for multiscale modeling: Laplacian pyramids and wavelet transforms. Not only do such reparameterizations allow the likelihood function to be directly expressed as a joint likelihood over the scales, we show that the Laplacian pyramid and wavelet transform also produces significant improvements to the state-of-the-art on a selection of benchmarks in likelihood modeling, including density estimation, lossless compression, and out-of-distribution detection. Investigating the theoretical basis of our empirical gains we uncover deep connections to score matching under the Earth Mover's Distance (EMD), which is a well-known surrogate for perceptual similarity. Code can be found at \href{https://github.com/lihenryhfl/pcdm}{this https url}.
Figures
Reference graph
Works this paper leans on
-
[1]
According to (Ho et al., 2020), the likelihood bound Eq
Proof. According to (Ho et al., 2020), the likelihood bound Eq. 5 L(z(1)) ≥ Ex1:T ∼q log pθ(z(1) 0 |z(1) 1 )| {z } L0(z(1)) − KL(q(z(1) T |z(1) 0 )||pθ(z(1) T ))| {z } LT (z(1)) − T −1X k=1 log KL(q(z(1) k |z(1) k+1, z(1) 0 )||pθ(z(1) k |z(1) k+1)) | {z } Lk(z(1)) , (21) can be decomposed into the three elements: a decoder reconstruction term L0(x), a pri...
work page 2020
-
[5]
Hierarchical autoregressive image models with auxiliary decoders
10 Published as a conference paper at ICLR 2024 Jeffrey De Fauw, Sander Dieleman, and Karen Simonyan. Hierarchical autoregressive image models with auxiliary decoders. arXiv preprint arXiv:1903.04933,
arXiv 2024
-
[6]
Nice: Non-linear independent components estimation
Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. arXiv preprint arXiv:1410.8516,
-
[8]
Learning energy-based models by diffusion recovery likelihood
Ruiqi Gao, Yang Song, Ben Poole, Ying Nian Wu, and Diederik P Kingma. Learning energy-based models by diffusion recovery likelihood. arXiv preprint arXiv:2012.08125,
arXiv 2012
-
[10]
Wavelet Score-Based Generative Modeling
Florentin Guth, Simon Coste, Valentin De Bortoli, and Stephane Mallat. Wavelet score-based generative modeling. arXiv preprint arXiv:2208.05003,
-
[12]
Deep anomaly detection with outlier exposure
Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. arXiv preprint arXiv:1812.04606,
-
[13]
Autoregressive diffusion models
Emiel Hoogeboom, Alexey A Gritsenko, Jasmijn Bastings, Ben Poole, Rianne van den Berg, and Tim Salimans. Autoregressive diffusion models. arXiv preprint arXiv:2110.02037,
-
[14]
Progressive growing of gans for improved quality, stability, and variation
11 Published as a conference paper at ICLR 2024 Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196,
arXiv 2024
Show all 39 references
-
[16]
Auto-encoding variational bayes
Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114,
-
[18]
Flow matching for generative modeling
Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747,
-
[19]
Generating high fidelity images with subscale pixel networks and multidimensional upscaling
Jacob Menick and Nal Kalchbrenner. Generating high fidelity images with subscale pixel networks and multidimensional upscaling. arXiv preprint arXiv:1812.01608,
-
[20]
Do deep generative models know what they don’t know? arXiv preprint arXiv:1810.09136,
Eric Nalisnick, Akihiro Matsukawa, Yee Whye Teh, Dilan Gorur, and Balaji Lakshminarayanan. Do deep generative models know what they don’t know? arXiv preprint arXiv:1810.09136,
-
[21]
Detecting out-of- distribution inputs to deep generative models using typicality
Eric Nalisnick, Akihiro Matsukawa, Yee Whye Teh, and Balaji Lakshminarayanan. Detecting out-of- distribution inputs to deep generative models using typicality. arXiv preprint arXiv:1906.02994,
1906 arXiv
-
[23]
Parallel multiscale autoregressive density estimation
12 Published as a conference paper at ICLR 2024 Scott Reed, Aäron Oord, Nal Kalchbrenner, Sergio Gómez Colmenarejo, Ziyu Wang, Yutian Chen, Dan Belov, and Nando Freitas. Parallel multiscale autoregressive density estimation. In Interna- tional Conference on Machine Learning, p...
2024
-
[25]
outlier
Alireza Shafaei, Mark Schmidt, and James J Little. Does your model know the digit 6 is not a cat? a less biased evaluation of" outlier" detectors. arXiv preprint arXiv:1809.04729,
-
[26]
Approximate earth mover’s distance in linear time
Sameer Shirdhonkar and David W Jacobs. Approximate earth mover’s distance in linear time. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pp. 1–8. IEEE,
2008
-
[27]
Flif: Free lossless image format based on maniac compression
Jon Sneyers and Pieter Wuille. Flif: Free lossless image format based on maniac compression. In 2016 IEEE international conference on image processing (ICIP), pp. 66–70. IEEE,
2016
-
[28]
Sliced score matching: A scalable approach to density and score estimation
Yang Song, Sahaj Garg, Jiaxin Shi, and Stefano Ermon. Sliced score matching: A scalable approach to density and score estimation. In Uncertainty in Artificial Intelligence, pp. 574–584. PMLR, 2020a. Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano E...
2011 arXiv
-
[30]
Pixel recurrent neural networks
13 Published as a conference paper at ICLR 2024 Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pp. 1747–1756. PMLR,
2024
-
[31]
Zero-shot image restoration using denoising diffusion null-space model
Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. arXiv preprint arXiv:2212.00490,
-
[32]
Moverscore: Text generation evaluating with contextualized embeddings and earth mover distance
Wei Zhao, Maxime Peyrard, Fei Liu, Yang Gao, Christian M Meyer, and Steffen Eger. Moverscore: Text generation evaluating with contextualized embeddings and earth mover distance. arXiv preprint arXiv:1909.02622,
1909 arXiv
-
[33]
Improved techniques for maximum likelihood estimation for diffusion odes
Kaiwen Zheng, Cheng Lu, Jianfei Chen, and Jun Zhu. Improved techniques for maximum likelihood estimation for diffusion odes. arXiv preprint arXiv:2305.03935,
-
[36]
at each scale s are also those of the wavelet hierarchy (Eq. 14). This is because the LLT convolution kernel with stride 2 is the bilinear downsampling operator. Thus d(·) =LLT ∗ (·). Now, since the wavelet operator W, d(·) and u(·) are norm preserving, ||z(s) lp ||2 = ||y(s) ...
2024
-
[38]
This is why we get differing behaviors between the two hierarchical maps
This allows us to upper bound the right side inequality in 2While the losses are the same for both models for each x, the gradients and therefore training trajectories are very different due to the different parameterizations of the representation space. This is why we get dif...
2024
-
[39]
All model architecture and hyper- parameters are retained from W-PCDM and LP-PCDM, except for the choice of z(1),
(i.e., Section 4.1). All model architecture and hyper- parameters are retained from W-PCDM and LP-PCDM, except for the choice of z(1), . . . ,z(S). In Tables 4, 5, and 6, we see that the removing volume-preservation property of h significantly degrades likelihood modeling perf...
2024
-
[1995]
Efficiently modeling long sequences with structured state spaces
Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396,
-
[2000]
Pixelcnn++: Improving the pixelcnn with discretized logistic mixture likelihood and other modifications
Tim Salimans, Andrej Karpathy, Xi Chen, and Diederik P Kingma. Pixelcnn++: Improving the pixelcnn with discretized logistic mixture likelihood and other modifications. arXiv preprint arXiv:1701.05517,
-
[2003]
Practical lossless compression with latent variables using bits back coding
James Townsend, Tom Bird, and David Barber. Practical lossless compression with latent variables using bits back coding. arXiv preprint arXiv:1901.04866,
1901 arXiv
-
[2008]
Lemma A.1 (From Theorem 2 in (Shirdhonkar & Jacobs, 2008))
using the notation in this text. Lemma A.1 (From Theorem 2 in (Shirdhonkar & Jacobs, 2008)). Consider the optimal transport cost Wp(x, y) (Eq
2008
-
[2009]
Score-based generative modeling secretly minimizes the wasserstein distance
Dohyun Kwon, Ying Fan, and Kangwook Lee. Score-based generative modeling secretly minimizes the wasserstein distance. arXiv preprint arXiv:2212.06359,
-
[2011]
Wavenet: A generative model for raw audio
Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499,
-
[2014]
Implicit generation and generalization in energy-based models
Yilun Du and Igor Mordatch. Implicit generation and generalization in energy-based models. arXiv preprint arXiv:1903.08689,
1903 arXiv
-
[2017]
Elucidating the design space of diffusion- based generative models
Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion- based generative models. arXiv preprint arXiv:2206.00364,
-
[2018]
Likelihood training of schr\" odinger bridge using forward-backward sdes theory
Tianrong Chen, Guan-Horng Liu, and Evangelos A Theodorou. Likelihood training of schr\" odinger bridge using forward-backward sdes theory. arXiv preprint arXiv:2110.11291,
-
[2019]
Waic, but why? generative ensembles for robust anomaly detection
Hyunsun Choi, Eric Jang, and Alexander A Alemi. Waic, but why? generative ensembles for robust anomaly detection. arXiv preprint arXiv:1810.01392,
-
[2020]
Generating long sequences with sparse transformers
Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509,
1904 arXiv
-
[2021]
Very deep vaes generalize autoregressive models and can outperform them on images
Rewon Child. Very deep vaes generalize autoregressive models and can outperform them on images. arXiv preprint arXiv:2011.10650,
2011 arXiv
-
[2022]
Laplacian pyramid-like autoencoder
Sangjun Han, Taeil Hur, and Youngmi Hur. Laplacian pyramid-like autoencoder. In Intelligent Computing: Proceedings of the 2022 Computing Conference, Volume 2, pp. 59–78. Springer,
2022
-
[2023]
Let h be a hierarchi- cal volume-preserving map such that h(x) = (z(1), z(2),
14 Published as a conference paper at ICLR 2024 A P ROOFS Lemma 4.1 (Probabilistic Invariance of Hierarchical V olume-preserving Maps). Let h be a hierarchi- cal volume-preserving map such that h(x) = (z(1), z(2), . . . ,z(S)), and pθ be a likelihood function on z(1), z(2), . ...
2024
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.