Pith. sign in

REVIEW 2 major objections 3 minor 1 cited by

Three scalar bijections are shown to be simultaneously globally smooth, defined on all of R, analytically invertible in closed form, and expressive; radial flows built on them rival coupling flows with 1000x fewer parameters on radially str

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

2026-08-03 10:14 UTC pith:VS7MNKRA

load-bearing objection Three analytically invertible smooth scalar bijections, plus a radial-flow architecture that needs a small-but-real smoothness fix at its centers. the 2 major comments →

arxiv 2601.10774 v2 pith:VS7MNKRA submitted 2026-01-15 cs.LG hep-lat

Analytic Bijections for Smooth and Interpretable Normalizing Flows

Mathis Gerdes , Miranda C. N. Cheng This is my paper
classification cs.LG hep-lat
keywords normalizing flowsanalytic bijectionsclosed-form inversionradial flowscoupling flowsC-infinity smoothnessFourier parametrizationlattice field theory
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's central claim is that the usual trade-off between smoothness, invertibility, and expressivity in normalizing-flow scalar maps is avoidable: it introduces three families of scalar bijections—cubic rational, sinh, and cubic polynomial—that are globally C∞ on all of R, invertible in closed form via Cardano's formula or conjugation, and flexible enough to create both local deformations and global mass shifts. If the claim holds, any coupling flow can swap its affine or spline transformer for one of these maps without losing invertibility or needing bounded-domain tuning, and the resulting densities are smooth rather than piecewise smooth. The paper further claims that a radial-flow architecture built on these bijections—applying a scalar map only to the radius around learned centers—is unusually stable during training, geometrically interpretable, and can match coupling-flow quality on targets with radial structure using orders of magnitude fewer parameters, while its two-dimensional Fourier variant makes learned angular modes directly inspectable. These properties matter for density estimation where smoothness, stability, or interpretability are priorities, including a lattice-field-theory application where the bijections outperform affine baselines and a symmetric zero-mode bijection prevents mode collapse.

Core claim

In the paper's own terms, the discovery is that two construction principles—algebraic rational functions and conjugation with a monotonic map—yield scalar bijections satisfying all five desiderata at once: smoothness, global domain, analytic invertibility, tractable Jacobians, and expressive parametrization. The cubic rational map h(x)=x+λ(x−γ)/(1+(x−γ)^2/σ^2) has a derivative that stays positive for −1<λ<8 and an inverse given by the unique real root of a cubic; the sinh map h(x)=σ arcsinh(e^μ(e^ν sinh((x−γ)/σ)+δ))+γ has a closed-form inverse obtained by sign flips and supports a global shift; the cubic polynomial map conjugates ax+bx^3 with a translation and also inverts via Cardano. Appli

What carries the argument

The central mathematical objects are the three closed-form bijections and the radial-flow Jacobian identity. The machinery is: (i) the cubic rational bijection from the algebraic-ansatz construction, whose inverse is the real root of a cubic and whose monotonicity bound −1<λ<8 is derived from the derivative's minimum; (ii) the conjugation construction h=g^{−1}(g(x)+δ), whose inverse is obtained by flipping δ and whose log-Jacobian is log g'(x)−log g'(h(x)); (iii) the radial map g(x)=f(∥x∥)(x/∥x∥), whose log-determinant is log|f'(r)|+(n−1)log|f(r)/r|, making sphere-preserving volumetric change explicit; and (iv) in 2D, Fourier series in the angle parametrizing the radial bijection's parameter

Load-bearing premise

The radial-flow claims rest on the premise that enough centers and layers can redistribute angular mass; a single layer preserves every ray, and unless the target's angular mass distribution around the chosen centers can be matched by composition, radial flows cannot represent it (and for full smoothness at the origin the angular-dependent variant needs ∂f/∂r(0,x̂)=1, a condition the experiments relax).

What would settle it

Train a single-center radial flow on a 2D target shaped as one small off-center blob, so that nearly all target mass lies in one angular wedge from any plausible center while the Gaussian prior is isotropic; if the flow cannot raise the density of empty wedges without adding many centers while a coupling flow with hundreds of parameters succeeds, the ray-preservation constraint is decisive. A sharper check on the smoothness claim: evaluate the angular-dependent radial flow's density near the origin; if a cusp or directional discontinuity appears when ∂f/∂r(0,x̂)≠1, the learned map is not C∞ at

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

If this is right

  • Coupling flows can use these bijections as drop-in replacements for affine or spline maps, eliminating bounded-domain tuning while matching or exceeding spline forward-KL performance on the 2D benchmark.
  • Radial flows trained with direct parameters accept learning rates an order of magnitude higher than coupling flows and produce densities that avoid the 'folding' artifacts of alternating axis-aligned coupling layers.
  • On targets with radial structure, radial flows reach comparable density quality with roughly 1000x fewer parameters; even a 49-parameter Fourier radial flow captures significant structure on the spiral target.
  • In the φ4 lattice-field-theory setup, all three analytic bijections raise effective sample size by about 10 percentage points over the affine baseline, and a Z2-symmetric zero-mode bijection maintains balanced sampling across both magnetization modes when trained separately.
  • Because the zero-mode bijection preserves the sign of the spatially averaged field, it gives a concrete recipe for building symmetry-preserving flows that prevent mode collapse in bimodal physical systems.

Where Pith is reading between the lines

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

  • Editorial inference: if the conjugation construction is as general as it appears, the same trick should transfer to autoregressive flows, which consume conditioner outputs in the same scalar-bijection slot; the paper notes the connection but does not test it.
  • Editorial inference: the ray-preservation constraint means the 1000x parameter saving is a low-dimensional phenomenon; in higher dimensions, the number of centers or angular modes needed to redistribute mass will likely grow, so the practical niche is targets with genuine radial symmetry rather than a general replacement for coupling layers.
  • Editorial inference: the stability of radial flows may be understood as avoiding coordinate-axis folding; a hybrid architecture with radial layers first and coupling layers last could inherit both stability and high-dimensional flexibility, a concrete configuration worth testing.
  • Editorial inference: inspecting the learned Fourier coefficients offers a diagnostic for whether a target is actually radially decomposable—if high-order harmonics are needed, the radial inductive bias is a poor fit, and the coefficients themselves reveal that.

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

2 major / 3 minor

Summary. The paper introduces three families of scalar bijections — cubic rational, sinh, and cubic polynomial — and claims they simultaneously satisfy global C∞ smoothness, unbounded domain, closed-form analytic invertibility, and expressive parametrization. These bijections are used as drop-in replacements in coupling flows and as the basis of a new radial-flow architecture that transforms the radial coordinate while preserving angular direction. The authors report experiments in 1D and 2D density estimation, a 2D spiral and Gaussian-mixture target, and a φ4 lattice field theory application, arguing that radial flows yield smoother densities, better training stability, and interpretable parameters with orders-of-magnitude fewer parameters than coupling flows.

Significance. If the technical claims hold, the proposed scalar bijections are a useful contribution: they occupy a previously empty corner of the design space, and the closed-form inverses and proven monotonicity bounds are practical assets. The radial-flow architecture, with direct parametrization and Fourier angular dependence, is an original idea with appealing interpretability. The φ4 mode-collapse experiment is thoughtful and demonstrates a genuine application. However, the paper's most distinctive qualitative claim — that radial flows composed of these bijections are globally smooth — is not established by the presented construction, and the experimental evidence for it is partly visual. This weakens the advertised advantage over coupling flows and needs to be addressed before the paper can be accepted.

major comments (2)
  1. [§4.3.1, Eq. (19); §4.3.4] The claim that radial flows composed of smooth scalar bijections yield globally smooth densities is not supported. For g(x)=c+(f(r)/r)(x-c) with r=||s⊙(x-c)||, smoothness at c requires f(r)/r to extend to a smooth function of x. For a smooth f with f(0)=0, this requires all even Taylor coefficients of f to vanish at 0. The correction f~(r)=f(r)−f(0) enforces only f(0)=0; the learnable γ in Eq. (12) and δ in Eq. (14) generically create even terms. Example: f(r)=r+εr^2 gives g(x)=x+ε||x||x, which is C^1 but not C^2 at c. Thus the 'smoothness' advantage over coupling flows in §5.2 rests on an unverified assertion. Please impose parity constraints on the radial bijections or rephrase the claim as 'smooth away from centers'.
  2. [§4.4, Eq. (27)] The condition ∂f/∂r(0,x̂)=1 is presented as ensuring full smoothness at the origin. It is necessary but not sufficient: a C∞ radial map g(x)=(f(r,x̂)/r)x requires all mixed derivatives involving angular variations to vanish at r=0, and Eq. (27) only fixes the first radial term. The paper also explicitly relaxes even this condition in practice (§4.4). The visible non-smooth point at step 200 in Figure 6 is acknowledged, but the text continues to describe radial flows as producing globally smooth densities. This should be stated as an architectural limitation of the implemented angular-dependent variant rather than as a fully smooth construction.
minor comments (3)
  1. [§5 vs. §5.5.1] The opening of §5 states that results are averaged over 6 independent runs, but Figure 14 and the accompanying text report 3 runs. Please reconcile this inconsistency.
  2. [§5.3] The spline baseline uses a manually tuned domain. Please report the chosen domain and the sensitivity of the comparison to that choice, as it affects the strength of the 'match or exceed spline performance' claim.
  3. [§4.2, §4.4, §4.5] Minor typos: 'secton' in §4.2; 'Too see' in §4.4; 'F ourier' spacing in §4.5 heading. Please proofread.

Circularity Check

0 steps flagged

No significant circularity; the constructions are explicit, parameter bounds are derived in-text, and self-citations are not load-bearing.

full rationale

The paper's derivation chain is self-contained. The three scalar bijections are explicitly constructed (cubic rational, Eq. 12; sinh via conjugation, Eqs. 8–11 and 14; cubic polynomial, Eq. 17), and the claims of invertibility, smoothness, and Jacobians are supported by in-text algebra: Cardano-based inverses (Appendix A.1), derivative positivity deriving the parameter bounds −1<λ<8 (Eqs. 5–6 and Appendix A.1), and radial Jacobian formulas (Section 4.3.1 and Appendix A.2). No parameter is fitted to a subset of benchmark data and then renamed as a prediction; learned parameters are optimized on the actual target densities, and the '1000× fewer parameters' claim is an empirical comparison between architectures, not a construction from the result. Self-citations are limited to the authors' own software library [28] and lecture notes [30]; these are implementation/review pointers, not uniqueness theorems, ansatze, or load-bearing mathematical inputs. The paper itself discloses real limitations—angular-mass conservation for single-center radial flows (Section 4.3.4), the required condition ∂f/∂r(0,x̂)=1 for full smoothness of angular-dependent flows (Section 4.4), and low-dimensional tractability (Section 6). The skeptical concern that angle-independent radial flows are not C∞ at learned centers unless the radial bijection is odd is a correctness risk about a mathematical smoothness claim, not a circularity: it does not show that any result is equivalent to its inputs by construction. Under the hard rules, that concern belongs to correctness, not to circularity scoring. I therefore find no circular step.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central mathematical claims rest only on standard algebra and calculus. The empirical claims about radial flows depend on the target having radial structure, which is stated. No new physical entities are introduced.

free parameters (6)
  • Number of stacked bijections K = 3, 9, 27, 128, 256 (1D); 1, 3, 9, 27 (coupling)
    Hyperparameter chosen by hand; affects expressivity and is central to the claim that depth improves performance.
  • Number of radial centers L = 5, 10, 20, 40
    Hyperparameter controlling angular mixing; needed for multimodal targets.
  • Learning rate = 1e-3, 4e-4, 5e-3, 1e-2 depending on experiment
    Hand-chosen; radial flows tolerate larger rates, supporting the stability claim.
  • Spline baseline knots and domain = 8 knots, domain manually tuned
    Baseline configuration; manual tuning may bias comparison in favor of unbounded analytic bijections.
  • Fourier terms K = 1, 3, 5, 7
    Angular resolution in Fourier radial flows; chosen by hand.
  • Parametrization epsilon values = 1e-1, 1e-3
    Small constants to ensure strict positivity and numerical stability; chosen by hand.
axioms (5)
  • standard math Cardano's formula for solving cubic equations
    Used in Appendix A.1 for cubic rational inverse and in Section 3.2.3 for cubic polynomial inverse.
  • standard math Change-of-variables formula for normalizing flows
    Defines the density transformation in Eq. (1).
  • standard math Matrix determinant lemma
    Used in Appendix A.2.2 to derive the radial flow Jacobian.
  • domain assumption Single radial layer preserves angular marginal distribution
    Section 4.3.4 states a single radial layer leaves the angular marginal invariant; necessary for the interpretability and stability claims.
  • domain assumption Benchmark targets have radial structure
    Spiral and Gaussian-mixture targets are radially organized, which is why radial flows perform well; this is stated in Section 5.2.

pith-pipeline@v1.3.0-alltime-deepseek · 25782 in / 13605 out tokens · 131696 ms · 2026-08-03T10:14:54.405724+00:00 · methodology

0 comments
read the original abstract

A key challenge in normalizing flows is finding expressive invertible scalar bijections. Existing approaches face trade-offs: affine transformations are smooth and analytically invertible but lack expressivity; monotonic splines offer local control but are only piecewise smooth and act on bounded domains; residual flows achieve smoothness but need numerical inversion. We introduce three families of analytic bijections that are globally smooth ($C^\infty$), defined on all of $\mathbb{R}$, and analytically invertible in closed form, combining the favorable properties of prior approaches. Beyond serving as drop-in replacements in coupling flows, where they match or exceed spline performance, we develop radial flows: a novel architecture using direct parametrization that transforms the radial coordinate while preserving angular direction. Radial flows exhibit exceptional training stability, produce geometrically interpretable transformations, and on targets with radial structure can achieve comparable quality to coupling flows with $1000\times$ fewer parameters. We provide comprehensive evaluation on 1D and 2D benchmarks, and demonstrate applicability to higher-dimensional physics problems through experiments on $\phi^4$ lattice field theory, where our bijections outperform affine baselines and enable problem-specific designs that address mode collapse.

Figures

Figures reproduced from arXiv: 2601.10774 by Mathis Gerdes, Miranda C. N. Cheng.

Figure 1
Figure 1. Figure 1: Three types of bijection behavior. Top row: Bijection functions h(x) (solid) vs. identity (dashed). Bottom row: Resulting density (solid) when transforming a standard normal prior (dashed). Left (cubic rational): Local deformation creates bimodal structure while h(x) → x for large |x|. Thus, “stretching” in some parts is compensated by “compressing” in a nearby region. Middle (sinh): Local “stretching” dis… view at source ↗
Figure 2
Figure 2. Figure 2: 1D flow performance vs. number of stacked bijections. Both metrics tend to improve [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training dynamics for 1D flows. Each panel shows reverse KL divergence vs. training [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Transformation of points on a circle un [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Source distribution samples and where they are mapped to by the trained coupling flow [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Evolution of learned log-likelihood during training. Top: Radial Fourier flow develops [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Layer-by-layer progression through a 12-layer coupling flow. Colors show final location in [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Learned densities for a 2D Gaussian mixture (5 components). All flows use cubic bijec [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Forward KL divergence for 2D coupling flows. Our bijections (markers with error bars) improve with more copies up to N = 9, with cu￾bic outperforming both affine and spline baselines (horizontal lines). Error bars show one standard deviation over 6 random seeds. Quantitative results [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Training dynamics for 2D coupling flows. Each panel shows forward KL vs. training [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Forward KL divergence for radial flows as a function of number of centers and stacked [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Learned densities for Fourier radial flows with increasing angular resolution. Even with [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Radial transformation f(r) as a function of input radius r (y-axis) and angle ϕ (x-axis) for models with increasing Fourier terms. Color indicates output radius r ′ = f(r, ϕ). With one term (constant), the transformation is independent of ϕ, yielding vertical color bands. • Superior training stability (higher learning rates, no divergence). • Interpretable transformations (Section 5.2). • Smoother learned… view at source ↗
Figure 14
Figure 14. Figure 14: ESS during training for ϕ 4 theory on a 20 × 20 lattice. Analytic bijections (cubic, sinh, rational cubic) achieve around 10% higher ESS than the affine baseline. Median with interquar￾tile range (shaded region) over 3 runs, smoothed with 5-step moving average. Results [PITH_FULL_IMAGE:figures/full_fig_p026_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Training dynamics for bimodal ϕ 4 distribution. Top: ESS (log scale). Bottom: Fraction of samples with negative magnetiza￾tion. Naive training achieves high ESS but col￾lapses to one mode (100% negative). Pretrain￾ing the zero-mode bijection maintains balanced sampling (50%) while achieving the same final ESS. Vertical dashed line marks when coupling layers are activated for the pretrain approach. 1.0 0.9… view at source ↗
Figure 17
Figure 17. Figure 17: Lattice field configurations from trained models with 10 samples per row and same ran [PITH_FULL_IMAGE:figures/full_fig_p028_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Parameter sweeps for the three analytic bijection types. Each row varies one parameter [PITH_FULL_IMAGE:figures/full_fig_p038_18.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Weight-Space Physics: Interpretable Hypernetworks for Lattice Quantum Field Theories

    hep-lat 2026-07 conditional novelty 7.0

    A JEPA-based hypernetwork maps lattice field theory couplings to flow-model weights, and the geometry of those weights recovers the phase transition, intrinsic dimension, and Ising critical exponent of 2D scalar field...

Reference graph

Works this paper leans on

36 extracted references · 25 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Density Estimation Using Real NVP

    L. Dinh, J. Sohl-Dickstein, and S. Bengio. “Density Estimation Using Real NVP”. In:5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. 2017. arXiv:1605.08803 [cs.LG]. 30

    Pith/arXiv arXiv 2017

  2. [2]

    Neural Spline Flows

    C. Durkan, A. Bekasov, I. Murray, and G. Papamakarios. “Neural Spline Flows”. In:Advances in Neural Information Processing Systems. Vol. 32. Curran Associates, Inc., 2019. arXiv: 1906.04032 [stat.ML]

    Pith/arXiv arXiv 2019

  3. [3]

    Neural Autoregressive Flows

    C.-W. Huang, D. Krueger, A. Lacoste, and A. Courville. “Neural Autoregressive Flows”. In: Proceedings of the 35th International Conference on Machine Learning. International Confer- ence on Machine Learning. PMLR, July 3, 2018, pp. 2078–2087. arXiv:1804.00779 [cs.LG]

    Pith/arXiv arXiv 2018

  4. [4]

    Neural Networks with Cheap Differential Operators

    R. T. Q. Chen and D. K. Duvenaud. “Neural Networks with Cheap Differential Operators”. In:Advances in Neural Information Processing Systems. Vol. 32. Curran Associates, Inc.,

  5. [5]

    Invert- ible Residual Networks

    J. Behrmann, W. Grathwohl, R. T. Q. Chen, D. Duvenaud, and J.-H. Jacobsen. “Invert- ible Residual Networks”. In:Proceedings of the 36th International Conference on Machine Learning. International Conference on Machine Learning. PMLR, May 24, 2019, pp. 573–582. arXiv:1811.00995 [cs.LG]

    Pith/arXiv arXiv 2019

  6. [6]

    Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design

    J. Ho, X. Chen, A. Srinivas, Y. Duan, and P. Abbeel. “Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Design”. In:Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA. Ed. by K. Chaudhuri and R. Salakhutdinov. Vol. 97. Proceedings of Mac...

    Pith/arXiv arXiv 2019

  7. [7]

    Gaussianization Flows

    C. Meng, Y. Song, J. Song, and S. Ermon. “Gaussianization Flows”. In:Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics. International Conference on Artificial Intelligence and Statistics. PMLR, June 3, 2020, pp. 4336–4345. arXiv:2003.01941 [cs.LG]

    Pith/arXiv arXiv 2020

  8. [8]

    Smooth Normalizing Flows

    J. K¨ ohler, A. Kr¨ amer, and F. Noe. “Smooth Normalizing Flows”. In:Advances in Neural Information Processing Systems. Vol. 34. Curran Associates, Inc., 2021, pp. 2796–2809. arXiv: 2110.00351 [stat.ML]

    Pith/arXiv arXiv 2021

  9. [9]

    Density Estimation by Dual Ascent of the Log- Likelihood

    E. G. Tabak and E. Vanden-Eijnden. “Density Estimation by Dual Ascent of the Log- Likelihood”. In:Communications in Mathematical Sciences8.1 (Mar. 2010), pp. 217–233. issn: 1539-6746, 1945-0796

    2010

  10. [10]

    Normalizing Flows for Probabilistic Modeling and Inference

    G. Papamakarios, E. Nalisnick, D. J. Rezende, S. Mohamed, and B. Lakshminarayanan. “Normalizing Flows for Probabilistic Modeling and Inference”. In:J. Mach. Learn. Res.22.1 (Jan. 1, 2021), 57:2617–57:2680.issn: 1532-4435. arXiv:1912.02762 [stat]

    Pith/arXiv arXiv 2021

  11. [11]

    Normalizing Flows: An Introduction and Review of Current Methods

    I. Kobyzev, S. J. D. Prince, and M. A. Brubaker. “Normalizing Flows: An Introduction and Review of Current Methods”. In:IEEE Trans. Pattern Anal. Machine Intell.43.11 (Nov. 1, 2021), pp. 3964–3979.doi:10.1109/tpami.2020.2992934. arXiv:1908.09257 [stat.ML]

    arXiv 2021

  12. [12]

    Masked Autoregressive Flow for Density Estimation

    G. Papamakarios, I. Murray, and T. Pavlakou. “Masked Autoregressive Flow for Density Estimation”. In:Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA. Ed. by I. Guyon, U. von Luxburg, S. Bengio, H. M. Wallach, R. Fergus, S. V. N. Vishwanathan, and...

    Pith/arXiv arXiv 2017

  13. [13]

    Improved Variational Inference with Inverse Autoregressive Flow

    D. P. Kingma, T. Salimans, R. Jozefowicz, X. Chen, I. Sutskever, and M. Welling. “Improved Variational Inference with Inverse Autoregressive Flow”. In:Advances in Neural Information Processing Systems. Vol. 29. Curran Associates, Inc., 2016. arXiv:1606.04934.url:https: / / proceedings . neurips . cc / paper / 2016 / hash / ddeebdeefdb7e7e7a697e1c3e3d8ef54...

    Pith/arXiv arXiv 2016

  14. [14]

    NICE: Non-linear Independent Components Estima- tion

    L. Dinh, D. Krueger, and Y. Bengio. “NICE: Non-linear Independent Components Estima- tion”. In:3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Workshop Track Proceedings. Ed. by Y. Bengio and Y. LeCun

    2015

  15. [15]

    Glow: Generative Flow with Invertible 1x1 Convolutions

    D. P. Kingma and P. Dhariwal. “Glow: Generative Flow with Invertible 1x1 Convolutions”. In:Advances in Neural Information Processing Systems. Vol. 31. Curran Associates, Inc.,

  16. [16]

    Residual Flows for Invert- ible Generative Modeling

    T. Q. Chen, J. Behrmann, D. Duvenaud, and J.-H. Jacobsen. “Residual Flows for Invert- ible Generative Modeling”. In:Proceedings of the 33rd International Conference on Neural Information Processing Systems. Ed. by H. M. Wallach, H. Larochelle, A. Beygelzimer, F. d’Alch´ e-Buc, E. B. Fox, and R. Garnett. 2019, pp. 9913–9923. arXiv:1906.02735 [stat]

    Pith/arXiv arXiv 2019

  17. [17]

    Neural Ordinary Differ- ential Equations

    R. T. Q. Chen, Y. Rubanova, J. Bettencourt, and D. K. Duvenaud. “Neural Ordinary Differ- ential Equations”. In:Advances in Neural Information Processing Systems. Vol. 31. Curran Associates, Inc., 2018. arXiv:1806.07366 [cs]

    Pith/arXiv arXiv 2018

  18. [18]

    FFJORD: Free-Form Continuous Dynamics for Scalable Reversible Generative Models

    W. Grathwohl, R. T. Q. Chen, J. Bettencourt, I. Sutskever, and D. Duvenaud. “FFJORD: Free-Form Continuous Dynamics for Scalable Reversible Generative Models”. In:7th Inter- national Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. arXiv:1810.01367 [cs.LG]

    Pith/arXiv arXiv 2019

  19. [19]

    Invertible Generative Modeling Using Lin- ear Rational Splines

    H. M. Dolatabadi, S. M. Erfani, and C. Leckie. “Invertible Generative Modeling Using Lin- ear Rational Splines”. In:The 23rd International Conference on Artificial Intelligence and Statistics, AISTATS 2020, 26-28 August 2020, Online [Palermo, Sicily, Italy]. Ed. by S. Chi- appa and R. Calandra. Vol. 108. Proceedings of Machine Learning Research. PMLR, 202...

    Pith/arXiv arXiv 2020

  20. [20]

    Neural Diffeomorphic Non-uniform B-spline Flows

    S. Hong and S. Y. Chun. “Neural Diffeomorphic Non-uniform B-spline Flows”. In:Proceedings of the 37th AAAI Conference on Artificial Intelligence. Ed. by B. Williams, Y. Chen, and J. Neville. AAAI Press, 2023, pp. 12225–12233.doi:10.1609/AAAI.V37I10.26441. arXiv: 2304.04555 [cs]

    Pith/arXiv arXiv 2023

  21. [21]

    Variational Inference with Normalizing Flows

    D. J. Rezende and S. Mohamed. “Variational Inference with Normalizing Flows”. In:Proceed- ings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015. Ed. by F. R. Bach and D. M. Blei. Vol. 37. JMLR Workshop and Conference Proceedings. JMLR.org, 2015, pp. 1530–1538. arXiv:1505.05770 [stat.ML]

    Pith/arXiv arXiv 2015

  22. [22]

    Normalizing Flows on Tori and Spheres

    D. J. Rezende, G. Papamakarios, S. Racani` ere, M. S. Albergo, G. Kanwar, P. E. Shanahan, and K. Cranmer. “Normalizing Flows on Tori and Spheres”. In:Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event. Vol. 119. Proceedings of Machine Learning Research. PMLR, 2020, pp. 8083–8092. arXiv: 2002.024...

    Pith/arXiv arXiv 2020

  23. [23]

    SE(3) Equivariant Augmented Coupling Flows

    L. I. Midgley, V. Stimper, J. Antor´ an, E. Mathieu, B. Sch¨ olkopf, and J. M. Hern´ andez- Lobato. “SE(3) Equivariant Augmented Coupling Flows”. In:Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023. Ed. by A. Oh, T. Naumann, A. ...

    Pith/arXiv arXiv 2023

  24. [24]

    Representational Aspects of Depth and Conditioning in Normalizing Flows

    F. Koehler, V. Mehta, and A. Risteski. “Representational Aspects of Depth and Conditioning in Normalizing Flows”. In:Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event. Ed. by M. Meila and T. Zhang. Vol. 139. Proceedings of Machine Learning Research. PMLR, 2021, pp. 5628–5636. arXiv: 2010.01155 [cs]

    Pith/arXiv arXiv 2021

  25. [25]

    Andrade.Stable Training of Normalizing Flows for High-dimensional Variational Infer- ence

    D. Andrade.Stable Training of Normalizing Flows for High-dimensional Variational Infer- ence. Feb. 26, 2024. arXiv:2402.16408 [stat]. Pre-published

    Pith/arXiv arXiv 2024

  26. [26]

    On the Robustness of Normalizing Flows for Inverse Prob- lems in Imaging

    S. Hong, I. Park, and S. Y. Chun. “On the Robustness of Normalizing Flows for Inverse Prob- lems in Imaging”. In:IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023. IEEE, 2023, pp. 10711–10721.doi:10.1109/ICCV51070. 2023.00986. arXiv:2212.04319 [cs]

    arXiv 2023

  27. [27]

    Bradbury, R

    J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. Vander{P}las, S. Wanderman-{M}ilne, and Q. Zhang.JAX: Composable Transfor- mations of Python+NumPy Programs. Version 0.3.13. jax-ml, 2018.url:https://github. com/jax-ml/jax

    2018

  28. [28]

    Bijx: Bijections and Normalizing Flows with JAX/NNX

    M. Gerdes and M. C. N. Cheng. “Bijx: Bijections and Normalizing Flows with JAX/NNX”. In:J. Open Source Softw.10.116 (Dec. 6, 2025), p. 9521.issn: 2475-9066.doi:10.21105/ joss.09521

    2025

  29. [29]

    Flow-Based Sampling for Lattice Field Theories

    G. Kanwar. “Flow-Based Sampling for Lattice Field Theories”. In:Proceedings of The 40th International Symposium on Lattice Field Theory — PoS(LATTICE2023). The 40th Inter- national Symposium on Lattice Field Theory. Fermi National Accelerator Laboratory: Sissa Medialab, Oct. 11, 2024, p. 114.doi:10.22323/1.453.0114. arXiv:2401.01297 [hep-lat]

    Pith/arXiv arXiv 2024

  30. [30]

    Lecture Notes on Normalizing Flows for Lattice Quantum Field Theories

    M. C. N. Cheng and N. Stratikopoulou. “Lecture Notes on Normalizing Flows for Lattice Quantum Field Theories”. In:arXiv:2504.18126 [hep-lat](Apr. 25, 2025). arXiv:2504.18126 [hep-lat]

    arXiv 2025

  31. [31]

    Flow-Based Generative Models for Markov Chain Monte Carlo in Lattice Field Theory

    M. S. Albergo, G. Kanwar, and P. E. Shanahan. “Flow-Based Generative Models for Markov Chain Monte Carlo in Lattice Field Theory”. In:Phys. Rev. D100.3 (Aug. 23, 2019), p. 034515.issn: 2470-0010, 2470-0029.doi:10.1103/PhysRevD.100.034515. arXiv:1904. 12072 [hep-lat]

    doi:10.1103/physrevd.100.034515 2019

  32. [32]

    On the Universality of Volume-Preserving and Coupling-Based Normalizing Flows

    F. Draxler, S. Wahl, C. Schn¨ orr, and U. K¨ othe. “On the Universality of Volume-Preserving and Coupling-Based Normalizing Flows”. In:Proceedings of the 41st International Confer- ence on Machine Learning. Vol. 235. ICML’24. Vienna, Austria: JMLR.org, July 21, 2024, pp. 11613–11641

    2024

  33. [33]

    Free-Form Flows: Make Any Architecture a Normalizing Flow

    F. Draxler, P. Sorrenson, L. Zimmermann, A. Rousselot, and U. K¨ othe. “Free-Form Flows: Make Any Architecture a Normalizing Flow”. In:Proceedings of The 27th International Con- ference on Artificial Intelligence and Statistics. International Conference on Artificial Intelli- gence and Statistics. PMLR, Apr. 18, 2024, pp. 2197–2205. arXiv:2310.16624 [cs.L...

    Pith/arXiv arXiv 2024

  34. [2015]

    arXiv:1410.8516 [cs.LG]

    Pith/arXiv arXiv

  35. [2018]

    arXiv:1807.03039 [stat.ML]

    Pith/arXiv arXiv

  36. [2019]

    arXiv:1912.03579 [cs.LG]

    Pith/arXiv arXiv 1912