Dual Functional Factor Model

• Dual Functional Factor Model is a framework that splits latent structures into complementary dimensions to capture both shared and unique variation in high-dimensional functional data.
• It encompasses various formulations such as static sparse factor curves, shared versus group-specific components, and global versus local covariance decompositions.
• Estimation leverages Bayesian and variational methods to ensure computational tractability and identifiability in complex, longitudinal, and multi-group settings.

Searching arXiv for papers on dual/functional factor models and the specific cited work. Dual Functional Factor Model denotes a family of factor-analytic constructions for functional, longitudinal, and high-dimensional functional data in which latent structure is split across two complementary dimensions rather than a single factorization axis. In the recent literature, this duality appears in several forms: static sparse loadings paired with subject-specific latent factor curves, shared and group-specific functional components across populations, smooth signal plus latent factor residual structure, cross-sectional basis functions combined with temporal functional loadings, and globally smooth covariance separated from locally banded covariance (Jaoua et al., 21 Mar 2026, Dai et al., 1 Apr 2026, Tang et al., 2021, Stanley et al., 2024). The common objective is parsimonious representation of dependence, smoothness, and heterogeneity while preserving interpretability and computational tractability in regimes where conventional longitudinal models, standard FPCA, or purely multivariate factor models are inadequate.

1. Conceptual scope

Across the literature, the term does not denote a single invariant parametrization. Instead, it refers to models that impose two coordinated latent structures on functional observations. In high-dimensional longitudinal molecular profiling, the primary construction uses static sparse variable loadings and subject-specific latent factor curves, while also admitting a column-functional alternative with functional loadings and static subject scores (Jaoua et al., 21 Mar 2026). In multi-group functional analysis, duality means a decomposition into components shared across groups and components specific to each group (Dai et al., 1 Apr 2026). In semiparametric settings, it means an orthogonal split between a parametric template and a nonparametric factor expansion (Kowal et al., 2021). In brain connectivity, the same phrase denotes a decomposition into global and local covariance components (Stanley et al., 2024).

Variant	Structural split	Representative source
Row-/column-functional	Static loadings with functional factors, or functional loadings with static scores	(Jaoua et al., 21 Mar 2026)
Shared/group-specific	Common latent functions plus group-specific latent functions	(Dai et al., 1 Apr 2026)
Smooth/factor-augmented	Smooth signal plus high-dimensional factor residuals	(Gao et al., 2021)
Cross-sectional/temporal	Population-specific bases plus common temporal factor dynamics	(Tang et al., 2021)
Global/local covariance	Smooth finite-rank global covariance plus banded local covariance	(Stanley et al., 2024)

This heterogeneity is substantive rather than terminological noise. It indicates that “dual” is used to describe two coordinated latent mechanisms: one typically captures broad, structured, or shared variation, and the other captures heterogeneity that is subject-specific, group-specific, local, parametric, or temporally dynamic. A recurring theme is that the dual decomposition is imposed to obtain identifiable scientific axes such as severity versus recovery, shared versus idiosyncratic neural activity, or long-range versus short-range dependence (Jaoua et al., 21 Mar 2026, Dai et al., 1 Apr 2026).

2. Core mathematical formulations

A central formulation for longitudinal molecular data is the sparse functional factor model

$$y_i^{(j)}(t)=\mu^{(j)}(t)+\sum_{q=1}^{Q} b_{jq}\, h_i^{(q)}(t)+\varepsilon_i^{(j)}(t),$$

with variable-specific mean $\mu^{(j)}(t)$, sparse static loadings $b_{jq}$, and subject-specific latent factor curves $h_i^{(q)}(t)$. Each factor curve is further decomposed by FPCA,

$$h_i^{(q)}(t)=\sum_{\ell=1}^{L}\zeta_{i\ell}^{(q)}\psi_\ell^{(q)}(t),$$

yielding the row-functional representation

$Y(t)\approx B\,H(t)+E(t).$

The same framework admits a column-functional alternative,

$$y_i^{(j)}(t)=\mu^{(j)}(t)+\sum_{q=1}^{Q} a_{jq}(t)\,s_{iq}+\varepsilon_i^{(j)}(t),$$

or, more generally, a fully dual functional form with both functional loadings and functional subject factors (Jaoua et al., 21 Mar 2026).

A distinct multi-group formulation writes the latent trajectory for subject $i$ in group $s$ as

$$f_{is}(t)=\mu_s(t)+\sum_{l=1}^{L}\eta_{ils}\widetilde{\lambda}_l(t)+\sum_{k=1}^{K_s}\rho_{iks}\widetilde{\phi}_k^s(t),$$

where $\mu^{(j)}(t)$0 are shared latent functional components and $\mu^{(j)}(t)$1 are group-specific latent functional components. Here the dual structure is explicit: common variation is encoded by the shared block, and group-level deviations are encoded by the group-specific block (Dai et al., 1 Apr 2026).

In factor-augmented smoothing, the duality is between a smooth functional component and a high-dimensional factor component:

$\mu^{(j)}(t)$2

or, in continuous-time notation,

$\mu^{(j)}(t)$3

This construction is designed for settings in which smoothing alone is inadequate because residuals still contain pervasive structure arising from measurement error, basis misspecification, or step changes (Gao et al., 2021, Gao et al., 2022).

For high-dimensional functional time series with multiple populations, the dual-factor structure is staged. First,

$\mu^{(j)}(t)$4

where $\mu^{(j)}(t)$5 captures cross-sectional heterogeneity. Second,

$\mu^{(j)}(t)$6

so the low-dimensional common functional time series is further decomposed into common functional loadings and a matrix-valued time series carrying temporal dynamics. The resulting model separates population-specific age profiles from shared temporal evolution (Tang et al., 2021).

Other variants adopt duality at the covariance level. In functional connectivity,

$\mu^{(j)}(t)$7

with a smooth finite-rank global covariance and a banded local covariance. In high-dimensional covariance estimation, duality appears as either functional factors with scalar loadings or scalar factors with functional loadings, followed by conditionally sparse idiosyncratic covariance estimation (Stanley et al., 2024, Li et al., 2023).

3. Identifiability, orthogonality, and rank determination

Identifiability is a defining issue because factor models are intrinsically rotation-nonidentifiable. In the longitudinal Bayesian functional factor model, components are identifiable only up to orthogonal transformations and sign changes. The proposed remedy is post-processing: estimate dense-grid latent curves, eigendecompose their empirical covariance, rotate the scores, and order components by explained variance, thereby restoring orthonormality of the FPCA basis within each factor (Jaoua et al., 21 Mar 2026).

Multi-group Bayesian functional factor analysis adds another layer of nonidentifiability because the loading parameterization uses parameter expansion. The model therefore rescales auxiliary and global scale parameters at each MCMC iteration, then applies a rotation-sign-permutation algorithm, and finally interprets the latent structure in covariance space through spectral decomposition of the shared and residual covariance operators. The resulting shared and group-specific subspaces are orthogonal and uniquely defined up to sign (Dai et al., 1 Apr 2026).

In semiparametric functional factor models, orthogonality is not merely a normalization device but a safeguard against functional confounding. The nonparametric basis is constrained to satisfy

$\mu^{(j)}(t)$8

so the parametric template and nonparametric factors are distinct in the discrete observation inner product. This yields likelihood factorization and posterior independence between parametric and nonparametric components under prior independence, preventing the nonparametric term from absorbing template-driven variation (Kowal et al., 2021).

The global/local covariance model supplies a stronger identifiability theorem. If the global covariance is smooth and finite rank with real-analytic eigenfunctions, and the local covariance is banded, then the decomposition $\mu^{(j)}(t)$9 is identifiable. At finite grid resolution, identifiability holds under the rank constraint

$b_{jq}$0

which links recoverable global rank to grid resolution and local bandwidth (Stanley et al., 2024).

Rank selection is handled differently across formulations. The Bayesian molecular model uses overspecification, spike-and-slab shrinkage on the loading matrix, shrinkage of unnecessary FPCA components, and factor-level posterior inclusion probabilities, retaining factors with $b_{jq}$1 above a threshold such as $b_{jq}$2 (Jaoua et al., 21 Mar 2026). The multi-group model estimates active shared and group-specific dimensions through parameter-expanded cumulative shrinkage priors (Dai et al., 1 Apr 2026). The semiparametric model uses an ordered spike-and-slab prior with effective rank

$b_{jq}$3

and proves posterior control of spurious activations in simplified settings (Kowal et al., 2021). Operator-based and covariance-based frequentist models typically rely on eigenvalue-ratio rules or information criteria (Tavakoli et al., 2019, Li et al., 2023).

4. Estimation and computation

The most explicitly scalable Bayesian instance employs mean-field variational Bayes with an annealed objective

$b_{jq}$4

using heated coordinate updates and a geometric temperature schedule decreasing from $b_{jq}$5 to $b_{jq}$6 over a default grid of $b_{jq}$7 levels. All updates are closed-form, avoid Gaussian-process covariance inversion, and scale to $b_{jq}$8 variables, with per-iteration cost near-linear in $b_{jq}$9 and approximately linear in $h_i^{(q)}(t)$0 and $h_i^{(q)}(t)$1. The implementation is provided in the R package bayesSYNC (Jaoua et al., 21 Mar 2026).

The multi-group Bayesian model instead uses a Gibbs sampler with separate shared and group-specific blocks. In practice, chains were initialized with $h_i^{(q)}(t)$2 and $h_i^{(q)}(t)$3, run for $h_i^{(q)}(t)$4 iterations with $h_i^{(q)}(t)$5 burn-in, and convergence was assessed with Geweke diagnostics and traceplots. A single chain took about $h_i^{(q)}(t)$6 hour for $h_i^{(q)}(t)$7 and $h_i^{(q)}(t)$8–$h_i^{(q)}(t)$9 hours for $$h_i^{(q)}(t)=\sum_{\ell=1}^{L}\zeta_{i\ell}^{(q)}\psi_\ell^{(q)}(t),$$0 because repeated high-dimensional matrix operations dominate the cost (Dai et al., 1 Apr 2026).

Factor-augmented smoothing uses an alternating ridge-PCA scheme. Conditional on the factor loading space, smoothing coefficients are updated through projection,

$$h_i^{(q)}(t)=\sum_{\ell=1}^{L}\zeta_{i\ell}^{(q)}\psi_\ell^{(q)}(t),$$1

and then factor loadings are re-estimated by PCA on residual covariance. The smoothing parameter is selected by mean generalized cross-validation, and the algorithm alternates until coefficient updates stabilize (Gao et al., 2021, Gao et al., 2022).

High-dimensional functional time-series forecasting uses a two-stage estimator. Stage 1 applies dynamic FPCA populationwise to estimate cross-sectional bases from long-run covariance operators and then estimates the common functional time series by functional concurrent regression. Stage 2 builds an auto-cross-covariance operator of the reduced common series, extracts common temporal loading functions, computes matrix-valued factors by inner products, and finally forecasts the vectorized factors with a VAR model (Tang et al., 2021).

Other computational paradigms emphasize covariance recovery rather than direct latent-score inference. The brain-connectivity model treats off-band sample covariance entries as observed and reconstructs the smooth low-rank global covariance by rank-constrained matrix completion with a roughness penalty, followed by varimax rotation and adaptive soft-thresholding for interpretability (Stanley et al., 2024). High-dimensional covariance estimators such as DIGIT and FPOET combine eigenanalysis for the common component with adaptive functional thresholding of residual covariance under conditional functional sparsity (Li et al., 2023, Leng et al., 2024).

5. Empirical domains and substantive findings

In longitudinal COVID-19 biomarker data, the Bayesian functional factor model was fit to symptomatic PCR-positive subjects over seven weeks post onset with irregular sampling and $$h_i^{(q)}(t)=\sum_{\ell=1}^{L}\zeta_{i\ell}^{(q)}\psi_\ell^{(q)}(t),$$2 biomarkers. Using annealed VB with $$h_i^{(q)}(t)=\sum_{\ell=1}^{L}\zeta_{i\ell}^{(q)}\psi_\ell^{(q)}(t),$$3, $$h_i^{(q)}(t)=\sum_{\ell=1}^{L}\zeta_{i\ell}^{(q)}\psi_\ell^{(q)}(t),$$4, and factor-level $$h_i^{(q)}(t)=\sum_{\ell=1}^{L}\zeta_{i\ell}^{(q)}\psi_\ell^{(q)}(t),$$5, two factors were selected. Factor 1 loaded on inflammatory proteins and kynurenine-pathway metabolites, with opposite signs for tryptophan and serotonin; its first eigenfunction was positive over time and acted as a severity axis, while its second crossed zero and acted as a recovery axis. Factor 2 was dominated by amino acid metabolism. Subject-level FPCA scores captured clinically meaningful heterogeneity in severity and recovery dynamics (Jaoua et al., 21 Mar 2026).

In multi-group EEG analysis of alcoholic and healthy subjects, the posterior modal configuration was $$h_i^{(q)}(t)=\sum_{\ell=1}^{L}\zeta_{i\ell}^{(q)}\psi_\ell^{(q)}(t),$$6. The shared covariance-derived loading captured canonical ERP components common to both groups, specifically a negative deflection near $$h_i^{(q)}(t)=\sum_{\ell=1}^{L}\zeta_{i\ell}^{(q)}\psi_\ell^{(q)}(t),$$7 ms and a positive deflection around $$h_i^{(q)}(t)=\sum_{\ell=1}^{L}\zeta_{i\ell}^{(q)}\psi_\ell^{(q)}(t),$$8–$$h_i^{(q)}(t)=\sum_{\ell=1}^{L}\zeta_{i\ell}^{(q)}\psi_\ell^{(q)}(t),$$9 ms. Group-specific loadings then revealed distinct neural activity patterns, with alcoholic-specific components showing more pronounced positive fluctuations between $Y(t)\approx B\,H(t)+E(t).$0–$Y(t)\approx B\,H(t)+E(t).$1 ms (Dai et al., 1 Apr 2026).

In factor-augmented smoothing, simulations and real data were used to demonstrate when the dual smooth-plus-factor structure matters. For example, with $Y(t)\approx B\,H(t)+E(t).$2, $Y(t)\approx B\,H(t)+E(t).$3, and $Y(t)\approx B\,H(t)+E(t).$4, the factor-augmented smoothing model achieved function-estimation aMSE $Y(t)\approx B\,H(t)+E(t).$5 versus $Y(t)\approx B\,H(t)+E(t).$6 for smoothing alone. In a step-jump design with $Y(t)\approx B\,H(t)+E(t).$7, it achieved RMSE $Y(t)\approx B\,H(t)+E(t).$8 versus $Y(t)\approx B\,H(t)+E(t).$9 and degrees of freedom $$y_i^{(j)}(t)=\mu^{(j)}(t)+\sum_{q=1}^{Q} a_{jq}(t)\,s_{iq}+\varepsilon_i^{(j)}(t),$$0 versus $$y_i^{(j)}(t)=\mu^{(j)}(t)+\sum_{q=1}^{Q} a_{jq}(t)\,s_{iq}+\varepsilon_i^{(j)}(t),$$1. Applications to Canadian weather, Australian temperature, and Adelaide Airport Friday temperatures showed that factor adjustment removes residual spikes and improves fit when residual covariance is structured (Gao et al., 2021, Gao et al., 2022).

For high-dimensional functional time series, the dual-factor forecasting model was applied to Japanese subnational age-specific mortality for $$y_i^{(j)}(t)=\mu^{(j)}(t)+\sum_{q=1}^{Q} a_{jq}(t)\,s_{iq}+\varepsilon_i^{(j)}(t),$$2 prefectures over 1975–2022. In expanding-window evaluation, it achieved the smallest RMSFE for both sexes across most horizons and produced competitive or superior interval performance. The same forecasts were translated into annuity pricing; for female age $$y_i^{(j)}(t)=\mu^{(j)}(t)+\sum_{q=1}^{Q} a_{jq}(t)\,s_{iq}+\varepsilon_i^{(j)}(t),$$3 in year $$y_i^{(j)}(t)=\mu^{(j)}(t)+\sum_{q=1}^{Q} a_{jq}(t)\,s_{iq}+\varepsilon_i^{(j)}(t),$$4 at $$y_i^{(j)}(t)=\mu^{(j)}(t)+\sum_{q=1}^{Q} a_{jq}(t)\,s_{iq}+\varepsilon_i^{(j)}(t),$$5, the pricing error per $y_i^{(j)}(t)=\mu^{(j)}(t)+\sum_{q=1}^{Q} a_{jq}(t)\,s_{iq}+\varepsilon_i^{(j)}(t),$60.011 for the dual-factor method versus $y_i^{(j)}(t)=\mu^{(j)}(t)+\sum_{q=1}^{Q} a_{jq}(t)\,s_{iq}+\varepsilon_i^{(j)}(t),$70.152 for alternatives (Tang et al., 2021). Related Hilbert-space factor models for mortality curves also reported uniformly better MAFE and MSFE than competing approaches across horizons $y_i^{(j)}(t)=\mu^{(j)}(t)+\sum_{q=1}^{Q} a_{jq}(t)\,s_{iq}+\varepsilon_i^{(j)}(t),$8 (Tavakoli et al., 2019).

Applications outside biomedical and demographic settings reinforce the breadth of the framework. Functional dynamic factor methods for monthly yield curves recovered level, slope, and curvature structure and outperformed FPCA-based estimators in simulation and reconstruction of yield data (Martínez-Hernández et al., 2020). The earlier functional dynamic factor model with spline loading curves showed lower RMSFE than Diebold–Li at multiple horizons and supported profit-based trading exercises on zero-coupon yields (Hays et al., 2012). In resting-state fMRI, matrix-completion-based dual functional factor modeling of one axial slice from 210 subjects selected $y_i^{(j)}(t)=\mu^{(j)}(t)+\sum_{q=1}^{Q} a_{jq}(t)\,s_{iq}+\varepsilon_i^{(j)}(t),$9 and recovered anatomical segmentation and known resting-state networks, while post-processing improved interpretability and reduced fragmentation relative to ICA under higher model orders (Stanley et al., 2024).

6. Relation to adjacent methods, limitations, and extensions

A persistent comparison point is FPCA. Standard FPCA treats variables independently or yields dense multivariate components, whereas several dual functional factor models impose additional sparse or structured factorization across variables, groups, or covariance blocks. In the longitudinal molecular setting, this produces sparse loading patterns and subject-level functional scores that are directly interpretable; in time-series settings, lagged covariance or long-run covariance information is used to recover factors that reflect temporal dynamics rather than only static variance (Jaoua et al., 21 Mar 2026, Martínez-Hernández et al., 2020).

Several papers position dual functional factor models against dynamic factor models, Gaussian-process formulations, matrix factorizations, and ICA. The molecular Bayesian model emphasizes that FPCA-based factor functions provide orthogonal, interpretable within-factor variability and subject-specific heterogeneity without large Gaussian-process covariance operations, contrasting specifically with GP-based models such as MEFISTO (Jaoua et al., 21 Mar 2026). The multi-group model argues that explicit shared-versus-specific factorization improves reconstruction when both kinds of structure are present, while single-group BFPCA can dominate when only one kind of variation is present (Dai et al., 1 Apr 2026). The brain-connectivity model contrasts its banded-local plus smooth-global covariance split with classical orthogonal factor models, whose idiosyncratic covariance is diagonal, and with ICA, whose components can fragment as model order increases (Stanley et al., 2024).

The main limitations are similarly consistent across the literature. Rotation non-identifiability remains fundamental and is usually addressed by orthogonality constraints, post hoc rotations, or covariance-space interpretation rather than by model elimination (Jaoua et al., 21 Mar 2026, Dai et al., 1 Apr 2026). Structured mean-field VB may understate uncertainty, although annealing improves local-optimum behavior and simulations in the Bayesian molecular model showed close-to-nominal predictive coverage in moderate regimes (Jaoua et al., 21 Mar 2026). Many formulations assume Gaussian noise, weak dependence, stationarity, or independent Gaussian scores; several papers list non-Gaussian noise, covariate-dependent means, irregular sampling, partial sharing across subsets of groups, nonlinear factor dynamics, and full 3D or spatiotemporal scalability as open extensions (Dai et al., 1 Apr 2026, Tang et al., 2021, Stanley et al., 2024, Leng et al., 2024).

A common misconception is that a Dual Functional Factor Model must place functional objects on both sides of the factorization. The literature does not support such a narrow definition. Some models are dual because they combine static sparse loadings with latent factor curves, some because they separate shared and group-specific functions, some because they augment smoothing with factor-structured residuals, and some because they decompose covariance into global and local components. What unifies them is not a single algebraic template, but the deliberate use of two coordinated latent structures to capture complementary forms of variation in functional data (Jaoua et al., 21 Mar 2026, Li et al., 2023).