Dynamic Frechet Regression with Feature Selection for Distributional Data

Amrutha Dinesh; Kiran Adhikari; Mathew Kuttolamadom; Ying Lin

REVIEW 3 major objections 4 minor 4 references

Dynamic Frechet Regression with Feature Selection for Distributional Data

T0 review · 3 major / 4 minor · reviewed 2026-07-13 · grok-4.5

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

Pith's one-line read Dynamic Fréchet Regression predicts sequences of distributions from scalar predictors by weighting Fréchet means on both covariate similarity and index proximity, with sparse metric learning for feature selection.

desk verdict Solid, usable extension of Fréchet regression to index-dependent distributional trajectories; the FoSR-weight transfer is heuristic but the empirical case is clean enough to take seriously. read the letter →

arxiv 2607.09613 v1 pith:I4ZUT6UY submitted 2026-07-10 stat.ME stat.ML

Kiran Adhikari , Amrutha Dinesh , Mathew Kuttolamadom , Ying Lin This is my paper

classification stat.MEstat.ML MSC 62R2062G0862H12

keywords DynamicFréchetRegressiondistributionaldataWassersteinspaceindex-awareweightingsparsemetriclearningfunction-on-scalaradditivemanufacturing

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Many processes produce not a single number but a probability distribution that itself changes over layers, time, or depth. The paper introduces Dynamic Fréchet Regression (DFR) to regress those evolving distributional trajectories on ordinary scalar predictors. At each index the prediction is a weighted Fréchet mean in the native metric of the distributions (commonly 2-Wasserstein), so geometry is preserved. The weights combine how similar the predictors are and how close the indices are, letting the model borrow strength from neighboring layers while remaining index-specific. Because classical regression coefficients do not exist for distribution-valued responses, predictor importance is learned by sparsifying a metric that governs the similarity weights. Simulations and a directed-energy-deposition melt-pool study show lower Wasserstein error and cleaner recovery of active predictors than global Fréchet regression or ordinary function-on-scalar regression.

What carries the argument

The index-aware weight w_il(x0,j) (Eq. 4 / Property 1), obtained from the smoother matrix of roughness-penalized function-on-scalar regression and then transferred to a general metric; together with the sparse positive-semidefinite metric M that multiplies the predictor deviations inside those weights (Eq. 6–8).

What would settle it

On data whose true index dependence cannot be captured by the spline-derived weight matrix (e.g., abrupt jumps or non-smooth periodic regimes), DFR’s Wasserstein prediction error should cease to improve on Global Fréchet Regression and the sparse-metric stage should recover the wrong active set.

Watch

Extended reading notes

Core claim

Replacing the static weights of Global Fréchet Regression with index-aware weights derived from a roughness-penalized function-on-scalar smoother, and embedding sparse metric learning inside those weights, produces more accurate index-specific distributional predictions and recovers the predictors that drive the dynamics, without ever leaving the metric space of the responses.

Load-bearing premise

The Euclidean smoother weights remain a valid and geometry-preserving scheme once squared Euclidean distance is replaced by a general metric on distributions; the paper motivates the transfer by analogy rather than proving it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The real contribution here is a practical way to do Fréchet regression when the responses are sequences of distributions that evolve smoothly over an ordered index. They take the Global Fréchet idea, replace the static weights with an index-aware smoother borrowed from roughness-penalized function-on-scalar regression, and then bolt on sparse metric learning so you can do feature selection without Euclidean coefficients. That combination is new and fills a genuine gap between static object regression and Euclidean FoSR.

What they do well is keep the geometry honest (Wasserstein / quantile-space isotonic regression) and show clear gains. Simulations (N=50, p=20, L=10, three evolution patterns) give perfect recall of the six active predictors and lower Wasserstein MSE than both GFR and FoSR. The DED melt-pool case study (25 builds, LOOCV) recovers the expected laser-power quadratic and interactions and produces more stable layer-wise distributional predictions. The two-stage workflow (select with λ=0, then refit and tune roughness) is transparent and the math is easy to follow.

The soft spot the stress-test flags is real but not fatal: the transfer of the FoSR weight matrix (Property 1 / Eq. 4) into a general metric space is pure analogy; there is no consistency argument for the resulting weighted Fréchet mean under a dynamic distributional model, and feature selection freezes λ=0 so it inherits that form. That is a modeling assumption, not a proof. Still, the empirical design is clean enough that the gains are not obviously circular, and the free parameters (λ, μ, τ) are ordinary tuning knobs chosen by LOOCV. No public code or data is a practical annoyance, not a conceptual flaw.

This is for people who already work with Fréchet or Wasserstein regression and need something that respects both geometry and index dependence—especially process-monitoring or manufacturing audiences. It is not a foundational rewrite of the field, but it is a competent, usable method paper. I would send it to referees; the heuristic weight transfer and missing asymptotics are exactly the kind of things a good referee can force them to clarify or bound. Worth engaging if you care about distributional trajectories.

Referee Report

3 major / 4 minor

Summary. The paper proposes Dynamic Fréchet Regression (DFR) for regressing index-dependent sequences of distribution-valued responses on scalar predictors. DFR extends Global Fréchet Regression by defining index-specific predictions as weighted Fréchet means (Eq. 2) whose weights jointly encode predictor similarity and index proximity; the weights are obtained by transferring the smoother matrix of roughness-penalized function-on-scalar regression (Property 1, Eq. 4) to a general metric space. A geometry-aware feature-selection stage learns a sparse positive-semidefinite metric M via group-lasso penalization of the reconstruction loss (Eq. 8), with selection performed under λ = 0 and subsequent refitting with tuned roughness. For univariate distributions the procedure reduces to weighted isotonic regression in quantile space. Simulations (N = 50, p = 20, L = 10, three evolution scenarios, 25 Monte Carlo replications) and a leave-one-out study on Directed Energy Deposition melt-pool data report lower Wasserstein MSE than GFR and FoSR together with high recall of active predictors.

Significance. If the claimed gains hold under broader conditions, DFR supplies a practical, geometry-preserving tool for a data structure that is increasingly common in additive manufacturing, environmental monitoring and reliability engineering—namely, static covariates driving smooth trajectories of entire distributions. The two-stage sparse-metric procedure is a genuine contribution: it performs variable selection without Euclidean coefficients and is shown to recover known active sets with perfect recall while improving prediction. The efficient quantile-space implementation (PAVA) and the real DED case study that recovers physically plausible laser-power effects further increase the work’s applied value. These strengths are empirical and algorithmic; a clearer theoretical foundation for the weight transfer would raise the paper’s lasting methodological impact.

major comments (3)

[§3.3–3.4, Property 1, Eqs. (2)–(4)] The central construction equates the DFR estimator (Eq. 2) with a weighted Fréchet mean whose weights are taken directly from the Euclidean FoSR smoother (Property 1 / Eq. 4). The transfer is motivated solely by formal analogy; the manuscript supplies neither consistency of the resulting weighted Fréchet mean for the true conditional Fréchet mean under any dynamic distributional model, nor an analysis of how the spline roughness parameter continues to control bias–variance once the response geometry is non-Euclidean. Because both the reported MSE gains and the subsequent feature-selection stage rest on these weights, a theoretical justification (or at least a precise statement of the conditions under which the analogy remains valid) is load-bearing.
[§3.5, Eqs. (5)–(9)] Feature selection deliberately freezes λ = 0 (Eqs. 5–6) so that the sparse metric M is learned under the unpenalized weight form. If the true cross-index dependence requires non-zero roughness, the recovered active set (Eq. 9) and the refitted predictions can be systematically biased. The paper should either (i) provide evidence that selection under λ = 0 remains consistent for the active set under the three simulation regimes, or (ii) explore joint optimization of M and λ, or at minimum report sensitivity of the selected sets to small positive λ.
[§4.1, Table 3] All simulation responses are univariate Gaussians whose means and variances are linear in the same six active predictors; the three evolution scenarios therefore probe only location-scale dynamics. The empirical superiority over GFR and FoSR may not extend to multimodal, heavy-tailed or multivariate distributions, nor to settings in which the active predictors affect higher-order shape features. At least one non-Gaussian or shape-varying scenario is needed to support the claim of general geometry-aware superiority.

minor comments (4)

[§2.3, §3.4–3.5] Several typographical errors appear in the text (e.g., “cenrtain”, “determins”, “d i a g(p)”, “∑̂ −𝟏”). A careful proof-reading pass is required.
[Figure 1] Figure 1 captions and axis labels are sparse; adding the numerical ranges of βℓ, γℓ and a brief statement of the two displayed samples would improve readability.
[Eq. (4)] The notation for the roughness penalty matrix R and the Kronecker products in Eq. (4) is dense; a short expanded derivation or an algorithmic box would help readers implement the weights.
[§2.2] References to recent Wasserstein time-series and dynamical barycenter work (Zhang et al. 2020, Cheng et al. 2022) are present but the discussion of why those methods cannot incorporate scalar covariates is brief; one additional sentence would clarify the gap DFR fills.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: DFR is a proposed estimator (weighted Fréchet means with FoSR-derived index-aware weights plus sparse metric learning) evaluated by out-of-sample Wasserstein MSE and known-active-set recovery; nothing reduces by construction to its own inputs.

full rationale

The paper defines the DFR predictor (Eq. 2) by replacing the Euclidean squared distance in the FoSR smoother representation (Property 1 / Eq. 4) with a general metric distance, then learns a sparse metric M under the unpenalized form (Eqs. 5–8) and refits with tuned λ. This is an explicit methodological proposal, not a derivation that forces the reported quantities. Prediction accuracy is measured by mean squared 2-Wasserstein error on held-out Monte Carlo test sets (Section 4) and by leave-one-out cross-validation on the 25 DED builds (Section 5); feature recovery is scored against the known active set S of six predictors in the simulations (Table 2). Tuning of μ and λ is ordinary LOOCV regularization. No equation equates a claimed prediction or recovered feature set to a quantity defined solely by the fitted weights or metric; no uniqueness theorem or load-bearing self-citation is invoked to forbid alternatives; the FoSR-to-metric transfer is an unproved analogy (a correctness/assumption issue) but does not create definitional circularity. The empirical superiority over GFR and FoSR is therefore an independent, falsifiable claim rather than a tautology.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The central claim rests on the transfer of Euclidean FoSR weights to metric spaces, the validity of the 2-Wasserstein geometry for the responses, the group-lasso sparse-metric formulation, and a handful of tuning constants chosen by cross-validation. No new physical entities are postulated; the free parameters are ordinary regularization and thresholding constants.

free parameters (4)

roughness parameter λ
Controls cross-index smoothing; set to 0 during feature selection and later tuned by LOOCV; directly affects the index-aware weights.
sparsity parameter μ
Strength of the group-lasso penalty on rows of the metric matrix M; selected by LOOCV and determines which predictors survive.
selection threshold τ
Row-norm cutoff that declares a predictor active after metric learning; chosen by the authors and affects the reported active sets.
learning rate η and convergence tolerance ε
Gradient-descent hyperparameters for the sparse-metric optimization (Table 1); influence the recovered metric and therefore feature selection.

assumptions (5)

standard math Prediction of a metric-space response can be expressed as a weighted Fréchet mean (Eq. 1–2).
Standard definition of Fréchet regression (Petersen & Müller 2019); used throughout §3.
ad hoc to paper The roughness-penalized FoSR smoother induces index-aware weights that remain valid when Euclidean distance is replaced by a general metric (Property 1).
The paper motivates the transfer by analogy but does not prove that the same weights are optimal or consistent under a non-Euclidean metric.
domain assumption Univariate distributions are adequately represented by quantile functions on a fixed grid and the 2-Wasserstein distance reduces to L2 on those quantiles.
Standard for univariate Wasserstein geometry; invoked in §3.6 to obtain the isotonic regression form.
ad hoc to paper A learnable positive-semidefinite metric M with group-lasso row penalties correctly identifies predictors that drive the similarity weights.
Extension of Huang & Sun (2013) sparse metric learning; no consistency proof is given for the dynamic Fréchet setting.
domain assumption Distributional trajectories evolve smoothly enough across the ordered index that neighboring indices share information usefully.
Stated in the problem formulation (§3.1) and required for the roughness penalty to be beneficial.

invented entities (2)

Dynamic Fréchet Regression (DFR) estimator with index-aware weights
purpose: Defines the predicted distribution at a new predictor and target index as a weighted Fréchet mean whose weights depend jointly on predictor similarity and index proximity.
The estimator itself is the paper’s central construction; it has no independent existence outside the proposed method.
Geometry-aware sparse metric M for DFR feature selection
purpose: Replaces the fixed precision matrix so that predictor relevance is learned by shrinking entire rows of M, enabling coefficient-free variable selection.
Introduced in §3.5; its utility is demonstrated only inside the paper’s simulations and case study.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Dynamic Frechet Regression with Feature Selection for Distributional Data." pith.science (2026). https://pith.science/paper/I4ZUT6UY

@misc{pith2026260709613,
  author       = {Pith},
  title        = {Pith review of: Dynamic Frechet Regression with Feature Selection for Distributional Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I4ZUT6UY}},
  note         = {Machine review of arXiv:2607.09613}
}

read the original abstract

Many scientific and engineering applications generate responses that are not scalars or vectors, but statistical objects whose form evolves over an ordered index such as time, depth. Probability distributions are a prominent example, capturing variability and uncertainty that cannot be summarized by low-dimensional statistics. When such responses are observed sequentially, the resulting dynamic distributional trajectories pose significant challenges for regression, particularly in relating scalar predictors to both within-index variability and cross-index evolution. We propose Dynamic Fr\'echet Regression (DFR), a framework for modeling index-dependent trajectories of distribution-valued responses. DFR extends Global Fr\'echet Regression by introducing an index-aware weighting mechanism. At each index, predictions are defined as weighted Fr\'echet means in a metric space of distributions (e.g., Wasserstein space), preserving the intrinsic geometry of the response. The weights depend jointly on predictor similarity and index proximity, enabling index-specific prediction while borrowing strength across neighboring indices. To improve interpretability in high-dimensional settings, DFR incorporates a geometry-aware feature selection approach based on sparse metric learning, which identifies predictors driving distributional dynamics without relying on Euclidean coefficients. Simulation studies show improved predictive accuracy and feature recovery over existing methods. An application to additive manufacturing data demonstrates its ability to produce interpretable, index-specific distributional predictions.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

4 extracted references

[1]

MeltpoolNet: Melt pool characteristic prediction in Metal Additive Manufacturing using machine learning,

Akbari, P., Ogoke, F., Kao, N.-Y., Meidani, K., Yeh, C.-Y., Lee, W., and Barati Farimani, A. (2022), “MeltpoolNet: Melt pool characteristic prediction in Metal Additive Manufacturing using machine learning,” Additive Manufacturing, 55, 102817. Bauer, A., Scheipl, F., Küchenhoff, H., and Gabriel, A.-A. (2018), “An introduction to semiparametric function-on...

2022
[2]

Kernel regression with sparse metric learning,

Huang, R., and Sun, S. (2013), “Kernel regression with sparse metric learning,” Journal of Intelligent & Fuzzy Systems, 24, 775–787. 37 Iao, S. I., Zhou, Y., and Müller, H.-G. (2025), “Deep Fréchet Regression,” Journal of the American Statistical Association, Taylor & Francis, 120, 1437–1448. Karthikeyan, A., Balhara, H., Hanchate, A., Lianos, A. K., and ...

2013
[3]

Prediction of lack-of-fusion porosity for powder bed fusion,

Tang, M., Pistorius, P. C., and Beuth, J. L. (2017), “Prediction of lack-of-fusion porosity for powder bed fusion,” Additive Manufacturing, 14, 39–48. Tian, B., Kang, J., and Zhong, W. (2025), “Feature screening for metric space-valued responses based on Fréchet regression with its applications,” Biometrics, 81, ujaf007. Tucker, D. C., Wu, Y., and Müller,...

2017
[4]

Dimension Reduction for Fréchet Regression,

Zhang, Q., Xue, L., and Li, B. (2024), “Dimension Reduction for Fréchet Regression,” Journal of the American Statistical Association, Taylor & Francis, 119, 2733–2747. Zhang, Z., Liu, Z., and Wu, D. (2021), “Prediction of melt pool temperature in directed energy deposition using machine learning,” Additive Manufacturing, 37, 101692. Zhu, X., Jiang, F., Gu...

2024

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.

[1] [1]

MeltpoolNet: Melt pool characteristic prediction in Metal Additive Manufacturing using machine learning,

Akbari, P., Ogoke, F., Kao, N.-Y., Meidani, K., Yeh, C.-Y., Lee, W., and Barati Farimani, A. (2022), “MeltpoolNet: Melt pool characteristic prediction in Metal Additive Manufacturing using machine learning,” Additive Manufacturing, 55, 102817. Bauer, A., Scheipl, F., Küchenhoff, H., and Gabriel, A.-A. (2018), “An introduction to semiparametric function-on...

2022

[2] [2]

Kernel regression with sparse metric learning,

Huang, R., and Sun, S. (2013), “Kernel regression with sparse metric learning,” Journal of Intelligent & Fuzzy Systems, 24, 775–787. 37 Iao, S. I., Zhou, Y., and Müller, H.-G. (2025), “Deep Fréchet Regression,” Journal of the American Statistical Association, Taylor & Francis, 120, 1437–1448. Karthikeyan, A., Balhara, H., Hanchate, A., Lianos, A. K., and ...

2013

[3] [3]

Prediction of lack-of-fusion porosity for powder bed fusion,

Tang, M., Pistorius, P. C., and Beuth, J. L. (2017), “Prediction of lack-of-fusion porosity for powder bed fusion,” Additive Manufacturing, 14, 39–48. Tian, B., Kang, J., and Zhong, W. (2025), “Feature screening for metric space-valued responses based on Fréchet regression with its applications,” Biometrics, 81, ujaf007. Tucker, D. C., Wu, Y., and Müller,...

2017

[4] [4]

Dimension Reduction for Fréchet Regression,

Zhang, Q., Xue, L., and Li, B. (2024), “Dimension Reduction for Fréchet Regression,” Journal of the American Statistical Association, Taylor & Francis, 119, 2733–2747. Zhang, Z., Liu, Z., and Wu, D. (2021), “Prediction of melt pool temperature in directed energy deposition using machine learning,” Additive Manufacturing, 37, 101692. Zhu, X., Jiang, F., Gu...

2024