pith. sign in

arxiv: 2205.12361 · v4 · submitted 2022-05-24 · 📊 stat.ME

Bayesian modeling of nearly mutually orthogonal processes

James Matuk , Amy H. Herring , David B. Dunson This is my paper

Pith reviewed 2026-05-24 11:11 UTC · model grok-4.3

classification 📊 stat.ME
keywords nearly mutually orthogonal processesfunctional factor analysisBayesian modelingorthogonalityfunctional datadimension reductionlatent factorslongitudinal data
0
0 comments X

The pith

Nearly mutually orthogonal processes enforce mutual orthogonality of functional factor loadings while preserving computational simplicity.

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

The paper introduces nearly mutually orthogonal processes as a way to constrain factor loadings in functional factor analysis to be nearly orthogonal. This matters for functional and longitudinal data because orthogonality improves model parsimony and interpretability of variability patterns, yet exact constraints are computationally difficult. The approach defines a joint distribution controlled by a penalty parameter that trades off the degree of orthogonality against ease of posterior sampling. In an application to early childhood weight data, the method yields flexible inference on effects of breastfeeding, illness, and demographics.

Core claim

Nearly mutually orthogonal processes are stochastic processes whose joint distribution is governed by a penalty parameter that determines the degree to which the processes are mutually orthogonal; this construction allows the processes to be used as factor loadings in a functional factor model while maintaining computational simplicity and efficiency for Bayesian posterior inference.

What carries the argument

Nearly mutually orthogonal processes, stochastic processes whose joint distribution incorporates a penalty parameter that encourages mutual orthogonality.

If this is right

  • The penalty parameter directly controls the achievable level of orthogonality in the factor loadings.
  • Posterior computation remains straightforward because the processes avoid hard constraints.
  • The resulting model supports interpretable inference on covariate effects in longitudinal functional data.
  • The approach extends the range of functional factor models that can be fit in Bayesian settings without prohibitive cost.

Where Pith is reading between the lines

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

  • The construction could be adapted to enforce other linear constraints on loadings in non-functional factor models.
  • In very high-dimensional settings the penalty might need calibration to avoid under- or over-constraining the factors.
  • Simulations with known orthogonal loadings could quantify how close the posterior loadings come to exact orthogonality for different penalty values.

Load-bearing premise

The penalty parameter trades off orthogonality against computational tractability without introducing substantial bias or loss of flexibility in the posterior.

What would settle it

Compare posterior samples from the nearly orthogonal model against samples from an exactly orthogonal constrained model on the same dataset and check whether the loadings deviate substantially from orthogonality or whether predictive performance degrades.

Figures

Figures reproduced from arXiv: 2205.12361 by Amy H. Herring, David B. Dunson, James Matuk.

Figure 1
Figure 1. Figure 1: (a) Value of λ1 on which λ2 is conditioned. Realizations of λ2 | λ1 ∼ GP{0, Cνλ 2 (·, ·)} with l 2 2 = .1, τ 2 2 = 1 fixed, while (b) νλ = 0.0001, (c) νλ = 0.01, (d) νλ = 1, (e) νλ = 100. (f) Boxplots of 1000 realizations of hλ1, λ2i, with λ2 | λ1 ∼ GP{0, Cνλ 2 (·, ·)} with fixed l 2 2 = 0.1, τ 2 2 = 1 and νλ = 0.0001, 0.01, 1, 100. While an orthogonal Gaussian process is a prior for a single function orth… view at source ↗
Figure 2
Figure 2. Figure 2: (a) & (b) Mean integrated squared error for estimated parameters at different [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (a) Weight by age for n = 2898 children in the Cebu Longitudinal Health and Nutrition Survey. (b) Proportion of missing weight data by age. (c)-(f) Posterior samples of fitted weight trajectories, µ( #»t )+{λ1( #»t ), . . . , λ5( #»t )} >ηi+ P4 j=1 R T βj (s, #»t )zi,j (s)ds overlaid on yi( #»t i) for i = 421, 1626, 2601. The longitudinal recordings of weight of the i th child are denoted by yi( #»t i), wh… view at source ↗
Figure 4
Figure 4. Figure 4: (a)-(d) Posterior samples of (λ1( #»t ), . . . , λ5( #»t ))>Θq for q = 1, . . . , 4, representing the effect of the scalar covariates (height of the mother, sex of the child, area (urban or rural) where the family lives, season of birth (rainy or dry)) on weight. Mother’s height and sex of child are clearly associated with weight dynamics. Strata and season of birth appear to have near zero effects with re… view at source ↗
Figure 5
Figure 5. Figure 5: (a)-(d) Posterior samples of R T β11s≤s 0ds for s 0 = 0, 0.5, 1, 2 years, representing the effect of breastfeeding for different durations on weight. (e) Boxplots of posterior samples of k R T β11s≤s 0dsk2 for different values of s 0 , representing the magnitude of the effect of breastfeeding for different durations. Children that are fed breast milk tend to weigh less than non-breast milk fed counterparts… view at source ↗
Figure 6
Figure 6. Figure 6: (a) The value of λ1 on which λ2 is conditioned. Boxplots of 1000 realizations of hλ1, λ2i, with λ2 | λ1 ∼ GP(0, Cνλ 2 (·, ·)) (b) τ 2 2 = 1, νλ = 0.0001 are fixed as l2 = 0.001, 0.01, 0.1, 1 varies, (c) l2 = 0.01, νλ = 0.0001 are fixed as τ 2 2 = 0.25, 0.5, 1, 2 varies. Realizations of λ2 | λ1 ∼ GP(0, Cνλ 2 (·, ·)) with (d)-(g) τ 2 2 = 1, νλ = 0.0001 fixed, while l 2 2 = 0.001, l 2 2 = 0.01, l 2 2 = 0.1, l… view at source ↗
Figure 7
Figure 7. Figure 7: (a) Simulated observations. (b) Functional factor loadings that underlie the [PITH_FULL_IMAGE:figures/full_fig_p035_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: (a) & (b) Mean integrated squared error for estimated parameters at different [PITH_FULL_IMAGE:figures/full_fig_p041_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: (a) Observed growth curves for n = 39 male subjects in the Berkeley growth study. (b) Mean integrated squared error results of estimated model parameters. reasonable coverage, except for that of Yao et al. (2005) in the high sparsity setting. 41 [PITH_FULL_IMAGE:figures/full_fig_p041_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Scree plots that display boxplots of posterior samples of norms of factor loadings [PITH_FULL_IMAGE:figures/full_fig_p045_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: (a) Posterior samples of the mean process, »» [PITH_FULL_IMAGE:figures/full_fig_p046_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: (a)-(d) Proportion of missing data by age for each of the binary longitudinal [PITH_FULL_IMAGE:figures/full_fig_p047_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: (a)-(c) Posterior samples of Φ(µ z1 ( #»t )) (black) and Φ(µ z1 ( #»t ) ± sd(η z ·k )λ z k ( #»t )) (blue and red) for k = 1, 2, 3, representing the variability described by the loadings of breastfeading trajectories. (d)-(f) Posterior samples fitted breastfeading trajectories, Φ(µ z1 ( #»t ) + (λ z 1 ( #»t ), λz 2 ( #»t ), λz 3 ( #»t ))>η z i ) overlaid on zi,1( #»t i) for i = 250, 1079, 2087. not to hav… view at source ↗
Figure 14
Figure 14. Figure 14: Scatter plot matrices of posterior means of latent factors for (a) [PITH_FULL_IMAGE:figures/full_fig_p060_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: (a)-(f) Posterior samples of fitted atypical breastfeeding trajectories, Φ( »»»» [PITH_FULL_IMAGE:figures/full_fig_p061_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: (a)-(c) Posterior samples of Φ(µ zj ( #»t )) for j = 2, 3, 4, representing the proportion of children in the study that experienced an illness by age.(d)-(f) Posterior samples of βj ( #»t ), j = 2, 3, 4, representing the effect of experiencing an illness on weight. Children tend to experience cough more often than fever and fever more often than diarrhea. Experiencing diarrhea or fever tend to have a larg… view at source ↗
Figure 17
Figure 17. Figure 17: Inferred parameters related to functional factor analysis of weight and [PITH_FULL_IMAGE:figures/full_fig_p063_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Inference for parameters related to modeling breastfeeding status on weight using [PITH_FULL_IMAGE:figures/full_fig_p063_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Inferred functions related to modelling the effects of illness on weight using the [PITH_FULL_IMAGE:figures/full_fig_p064_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Inferred weight trajectories for different subjects using the Crainiceanu & [PITH_FULL_IMAGE:figures/full_fig_p064_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Boxplots of residuals by age. 64 [PITH_FULL_IMAGE:figures/full_fig_p064_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Tmag,yi based on posterior predictive samples (histogram) and observation (red lines) for random subjects (a)-(c) and subjects whose predicted values were furthest from the observed values (d)-(f). (g) the absolute difference between the mean of Tmag,yi based on posterior predictive samples and the observed value for all subjects. 65 [PITH_FULL_IMAGE:figures/full_fig_p065_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Tmean,(y1,...,yn) based on posterior predictive samples (histogram) and observation (red lines) for (a) t = 0 years, (b) t = 2/24 years, ..., (m) t = 2 years. 66 [PITH_FULL_IMAGE:figures/full_fig_p066_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Tvar,(y1,...,yn) based on posterior predictive samples (histogram) and observation (red lines) for (a) k = 1, (b) k = 2, ..., (e) k = 5. 67 [PITH_FULL_IMAGE:figures/full_fig_p067_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Trace plots related to the functional factor analysis results presented Figures [PITH_FULL_IMAGE:figures/full_fig_p068_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: Trace plots related to latent factor regression onto scalar covariates presented [PITH_FULL_IMAGE:figures/full_fig_p069_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Trace plots related to breastfeeding status interpolation and historical regression [PITH_FULL_IMAGE:figures/full_fig_p070_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: Trace plots related to illness status interpolation and concurrent regression [PITH_FULL_IMAGE:figures/full_fig_p071_28.png] view at source ↗
read the original abstract

Functional factor analysis is an important dimension reduction method for functional and longitudinal data. Factor loadings give insight into patterns of variability of the observations, while latent factors provide a low-dimensional representation of the data that is useful for inferential tasks. Constraining the functional factor loadings to be mutually orthogonal is desirable for model parsimony but is computationally challenging. In this work, we introduce nearly mutually orthogonal processes, which can be used to effectively enforce mutual orthogonality of factor loadings while maintaining computational simplicity and efficiency. The joint distribution is governed by a penalty parameter that determines the degree to which the processes are mutually orthogonal and is related to ease of posterior computation. We demonstrate that our approach can be used for flexible and interpretable inference in an application to studying the effects of breastfeeding status, illness, and demographic factors on weight dynamics in early childhood. Code is available on GitHub: https://github.com/jamesmatuk/NeMO-FFA

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript introduces nearly mutually orthogonal processes (NeMO) as a device for Bayesian functional factor analysis. These processes are defined via a joint distribution controlled by an explicit penalty parameter that encourages approximate mutual orthogonality of the functional factor loadings while preserving computational tractability for posterior sampling. The approach is illustrated through an application to longitudinal childhood weight data examining effects of breastfeeding status, illness, and demographic factors.

Significance. If the construction holds, the method supplies a practical route to interpretable factor loadings in functional data models without the computational burden of hard orthogonality constraints. The public GitHub code is a clear strength that aids reproducibility and adoption.

major comments (1)
  1. [Model definition and simulation study] The central modeling claim rests on the penalty parameter successfully trading off orthogonality against bias and flexibility in the posterior (reader's weakest assumption). The manuscript should include a targeted simulation study or sensitivity analysis quantifying how posterior inferences on loadings and factors change as the penalty varies across a range that spans near-orthogonality to near-independence.
minor comments (2)
  1. [Abstract] The abstract states that the method maintains efficiency but provides no quantitative metrics (e.g., runtime, effective sample size, or comparison to unconstrained FFA); adding one sentence summarizing such evidence would strengthen the claim.
  2. [Methods] Notation for the processes, the penalty, and the resulting covariance structure should be introduced with a short table or diagram in the methods section to aid readers new to functional factor models.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive review and positive recommendation for minor revision. We address the single major comment below and will incorporate the suggested addition in the revised manuscript.

read point-by-point responses

  1. Referee: The central modeling claim rests on the penalty parameter successfully trading off orthogonality against bias and flexibility in the posterior (reader's weakest assumption). The manuscript should include a targeted simulation study or sensitivity analysis quantifying how posterior inferences on loadings and factors change as the penalty varies across a range that spans near-orthogonality to near-independence.

    Authors: We agree that explicitly quantifying the effect of the penalty parameter via simulation would strengthen the presentation of the central modeling claim. In the revised manuscript we will add a targeted simulation study in which we generate data under the NeMO construction, vary the penalty across a grid spanning near-orthogonality to near-independence, and report the resulting changes in posterior summaries of the loadings and latent factors. This will directly illustrate the bias-flexibility trade-off and complement the existing real-data application. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper proposes nearly mutually orthogonal processes as an explicit modeling device controlled by a tunable penalty parameter that directly trades off orthogonality against posterior sampling ease. This construction is introduced as a deliberate modeling choice for functional factor analysis rather than a result derived from first principles or external theorems. No load-bearing self-citations, fitted inputs renamed as predictions, or self-definitional reductions appear in the provided abstract or described claims; the joint distribution is defined directly in terms of the penalty, and the application to childhood weight data serves as an illustration of the method rather than a validation that collapses into the modeling assumptions themselves. The derivation chain is therefore self-contained.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The penalty parameter is the primary free parameter introduced to control orthogonality. The new entity 'nearly mutually orthogonal processes' is postulated to achieve the desired computational properties. No external benchmarks or machine-checked results are referenced in the abstract.

free parameters (1)
  • penalty parameter
    Governs the degree to which the processes are mutually orthogonal and is tied to ease of posterior computation.
axioms (1)
  • domain assumption A joint distribution can be constructed such that a single penalty parameter controls mutual orthogonality while preserving tractable posterior sampling.
    This is the core modeling assumption stated in the abstract.
invented entities (1)
  • nearly mutually orthogonal processes no independent evidence
    purpose: To enforce approximate mutual orthogonality of functional factor loadings with computational simplicity.
    New modeling construct introduced by the authors.

pith-pipeline@v0.9.0 · 5686 in / 1228 out tokens · 23541 ms · 2026-05-24T11:11:16.736191+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

8 extracted references · 8 canonical work pages

  1. [1]

    M., VanDerslice, J., Akin, J., Guilkey, D., Black, R., Briscoe, J

    Adair, L., Popkin, B. M., VanDerslice, J., Akin, J., Guilkey, D., Black, R., Briscoe, J. & Flieger, W. (1993), ‘Growth dynamics during the first two years of life: A prospective study in the Philippines’, European Journal of Clinical Nutrition 47(1), 42–51. Adair, L. S. & Guilkey, D. K. (1997), ‘Age-specific determinants of stunting in Filipino children’, T...

    work page 1993

  2. [2]

    & Thompson, W

    55 Jiang, L., Zhong, Y., Elrod, C., Natarajan, L., Knight, R. & Thompson, W. K. (2020), ‘Bayestime: Bayesian functional principal components for sparse longitudinal data’, arXiv:2012.00579. Khan, K. & Calder, C. A. (2022), ‘Restricted spatial regression methods: Implications for inference’, Journal of the American Statistical Association 117(537), 482–494...

    work page arXiv 2020

  3. [3]

    URL: https://doi.org/10.1214/20-BA1213 Kowal, D. R. & Canale, A. (2021), ‘Semiparametric functional factor models with Bayesian rank selection’, arXiv:108.02151. Kowal, D. R., Matteson, D. S. & Ruppert, D. (2017), ‘A Gaussian multivariate functional dynamic linear model’, Journal of the American Statistical Association 112(518), 733–

    work page doi:10.1214/20-ba1213 2021

  4. [4]

    Lenk, P. J. & Choi, T. (2017), ‘Bayesian analysis of shape-restricted functions using Gaus- sian process priors’, Statistica Sinica pp. 43–69. Lin, L. & Dunson, D. B. (2014), ‘Bayesian monotone regression using Gaussian process projection’, Biometrika 101(2), 303–317. Malfait, N. & Ramsay, J. O. (2003), ‘The historical functional linear model’, Canadian J...

    work page doi:10.1093/biomet/asab056 2017

  5. [5]

    Tipping, M. E. & Bishop, C. M. (1999), ‘Probabilistic principal component analysis’, Jour- nal of the Royal Statistical Society: Series B 61(3), 611–622. van der Linde, A. (2008), ‘Variational Bayesian functional PCA’, Computational Statistics & Data Analysis 53, 517–533. van der Linde, A. (2009), ‘A Bayesian latent variable approach to functional princip...

    work page arXiv 1999

  6. [6]

    Children tend to experience cough more often than fever and fever more often than diarrhea

    61 (a) (b) (c) (d) (e) (f) Figure 16: (a)-(c) Posterior samples of Φ(µzj( # »t )) for j = 2, 3, 4, representing the proportion of children in the study that experienced an illness by age.(d)-(f) Posterior samples ofβj( # »t ), j = 2 , 3, 4, representing the effect of experiencing an illness on weight. Children tend to experience cough more often than fever...

    work page 2010

  7. [7]

    63 (a) (b) (c) (d) Figure 19: Inferred functions related to modelling the effects of illness on weight using the Crainiceanu & Goldsmith (2010) model: (a) logit −1(ˆµzj(t))

    (f) Boxplots of posterior samples of ∥ ∫ T ˆβ11s≤s′ds∥2 for different values of s′. 63 (a) (b) (c) (d) Figure 19: Inferred functions related to modelling the effects of illness on weight using the Crainiceanu & Goldsmith (2010) model: (a) logit −1(ˆµzj(t)). (b)-(d) Posterior samples of βj(t) for j = 2, 3,

    work page 2010

  8. [8]

    ,ˆλK( # »t i))⊤ˆηi + Oi ∑4 j=1 ∫ T βj(s, # »t )zi,j(s)ds, overlaid on yi( # »t i), i = 421, 1626,

    (a) (b) (c) Figure 20: Inferred weight trajectories for different subjects using the Crainiceanu & Goldsmith (2010) model: (a)-(c) Posterior samples of ˆ µ( # »t i) + (ˆλ1( # »t i), . . . ,ˆλK( # »t i))⊤ˆηi + Oi ∑4 j=1 ∫ T βj(s, # »t )zi,j(s)ds, overlaid on yi( # »t i), i = 421, 1626,

    work page 2010