---
title: 'RR-FBTC: Bayesian Tensor Completion'
url: https://www.emergentmind.com/topics/rank-revealing-functional-bayesian-tensor-completion-rr-fbtc
type: topic
---

# RR-FBTC: Bayesian Tensor Completion

Rank-Revealing Functional Bayesian Tensor Completion (RR-FBTC) is a variational Bayesian framework for low-rank tensor completion that enables rigorous automatic rank determination in both discrete and functional (continuous-indexed) tensor settings. RR-FBTC integrates functional tensor modeling with multi-output Gaussian process (MOGP) priors or hierarchical sparsity priors, providing a unified approach for rank-adaptive structure learning, probabilistic uncertainty quantification, and principled missing data recovery across discrete, continuous, and hybrid tensor domains [2512.21486][1505.02343][1301.7619].

## 1. Mathematical Model: Functional Tensor Completion

RR-FBTC models a $K$-mode tensor (potentially defined on continuous domains) via the functional CANDECOMP/PARAFAC (CP) decomposition:
$$
x_{\mathbf{i}} = \sum_{r=1}^{R} \prod_{k=1}^K u^k_r(i_k),
$$
where $i_k$ may be either discrete or real-valued, $u^k_r$ are latent mode-$k$ factor functions, and $R$ is the tensor rank to be inferred. Given noisy observations $\{(\mathbf{i}^n, y_n)\}_{n=1}^N$,
$$
y_n = x_{\mathbf{i}^n} + w_n, \quad w_n\sim\mathcal N(0,\tau^{-1}),
$$
the likelihood is Gaussian with respect to the observed entries.

Each mode-factor $U^k$ is governed by a multi-output Gaussian process prior:
$$
U^k(\cdot)\sim\mathcal{MGP}(0,\varsigma_k(\cdot,\cdot),\Gamma^{-1}),
$$
where $\varsigma_k$ is a specified kernel and $\Gamma = \text{diag}(\gamma_1, \ldots, \gamma_R)$ parameterizes an automatic relevance determination (ARD) shrinkage prior. This induces a matrix-normal prior distribution over the discretization $U^k\in\mathbb R^{N_k \times R}$. The ARD hyperpriors
$$
p(\gamma_r)=\text{Gamma}(\gamma_r\mid a_r,b_r)
$$
promote column-wise sparsity, which underpins the rank-revealing property of RR-FBTC [2512.21486].

## 2. Rank-Revealing Mechanism

The RR-FBTC framework exploits variational Bayesian inference with shrinkage-inducing (ARD) priors to reveal the effective tensor rank during learning. Posterior updates for the $\gamma_r$ hyperparameters ensure that components of negligible explanatory power are pruned automatically:
- If $\langle\gamma_r\rangle \to \infty$, the posterior variance of the corresponding mode-$k$ factors collapses, effectively removing component $r$ and decrementing the rank [2512.21486][1505.02343][1301.7619].
- The number of finite $\gamma_r$ determines the inferred rank, and iterative elimination is integrated into each variational inference step.

This form of semi-automatic model selection subsumes the use of fixed-rank methods and eliminates manual hyperparameter grid search or cross-validation for the rank parameter.

## 3. Theoretical Guarantees: Expressiveness and Approximation

RR-FBTC, when equipped with universal kernel families (e.g., product Matérn or RBF kernels), achieves a universal approximation property on compact domains. Specifically, for any continuous target function $g$ on a compact domain $\mathcal Z\subset\mathbb R^D$, there exists $R$ and mean functions $\bar u_r^k$ such that the CP-form expansion satisfies
$$
\|f-g\|_\infty<\epsilon,\quad f(\mathbf z)=\sum_{r=1}^R\prod_{k=1}^K \bar u_r^k(z_k).
$$
This ensures expressive capacity for both continuous and discretized multi-dimensional signals, and establishes a probabilistic functional generalization of classic low-rank tensor models [2512.21486].

## 4. Variational Inference and Algorithmic Framework

RR-FBTC adopts mean-field variational Bayesian inference for the joint posterior over factor functions, ARD hyperparameters, and noise precision. The updates admit closed-form expressions:
- $q(u^k_r) = \mathcal N(m^k_r, \Psi^k_r)$ for each factor function
- $q(\gamma_r) = \mathrm{Gam}(\hat a_r, \hat b_r)$
- $q(\tau) = \mathrm{Gam}(\hat a_0, \hat b_0)$

The key update equations for variational means and covariances involve mode-wise kernel inverses and products over non-pruned components. Pruning is performed at each iteration for components satisfying $\langle\gamma_r\rangle>\epsilon_\text{prune}$ [2512.21486][1505.02343]. Closed-form computational complexity is $O(R\max_k N_k^3)$ per iteration, dominated by kernel inversions.

The RR-FBTC optimization admits efficient implementations, "embarrassingly" parallel mode updates, and optional conjugate-gradient linear solvers for large-scale settings.

## 5. Empirical Results and Benchmarking

Evaluations on synthetic and real-world data, including MRI, climate, oceanography, and high-resolution image inpainting, demonstrate robust rank identification and state-of-the-art recovery accuracy:
- Synthetic continuous tensors: RRSE $< 0.2$ with 90% missing entries; rapid convergence to true rank [2512.21486][1505.02343].
- Image inpainting: PSNR and SSIM exceed fixed-rank and non-Bayesian competitors over a wide range of missing ratios and noise conditions.
- MRI inpainting: Achieved $-10.5$ dB NMSE at 50% missing with cross-slice kernel learning [1301.7619].
- Robustness: Rank converges from initial $R_{\text{init}}=\max_k N_k$ to the true value within a few iterations.
- Ablation studies: Recovery accuracy and rank-selection stability remain high across kernel hyperparameters.

A summary comparison of core features:

| Method (by arXiv ID)    | Tensor Model    | Rank Adaptation Mechanism       | Type of Prior              |
|-------------------------|-----------------|---------------------------------|----------------------------|
| [2512.21486]            | CP, functional  | ARD shrinkage via $\gamma_r$    | MOGP (kernel)              |
| [1505.02343]            | Tucker          | Group-sparsity on $\lambda^n_r$ | Student-t / Laplace        |
| [1301.7619]             | CP, PARAFAC     | Frobenius/ARD penalty           | Gaussian, kernel-enabled   |

## 6. Connections to Related Bayesian Tensor Models

RR-FBTC generalizes Bayesian low-rank tensor models for both canonical polyadic (CP), Tucker, and tensor ring decompositions:
- In Tucker models, group-sparsity priors on $\lambda^{(n)}_r$ enforce rank-revealing shrinkage in multilinear ranks [1505.02343].
- In tensor ring models, Student-t hierarchical priors on TR core tensors automatically drive redundant rank components to zero through precision variables $\lambda^{(n)}_r$ [2007.01055].
- In discrete PARAFAC formulations, Frobenius/ARD penalties on factor norms induces $\ell_{2/3}$ quasi-norm regularization for sparsity in component weights [1301.7619].

All these Bayesian models leverage variational inference schemes that yield posterior rank pruning, allow uncertainty quantification, and avoid tuning of rank or regularization hyperparameters.

## 7. Practical and Computational Considerations

Computational efficiency in RR-FBTC relies on kernel matrix inversion per factor and per component retained after pruning. Scalability is achieved when $R$ is substantially smaller than ambient tensor dimensions. The framework supports non-Gaussian observation models (e.g., Poisson), non-Euclidean mode kernels (RKHS-structured priors), and empirical estimation of covariance matrices from training data [1301.7619].

Open-source implementations, reproducibility scripts, data-generation procedures, and hyperparameter settings are detailed for various domains. Example code and datasets for the functional RR-FBTC implementation are available via https://github.com/OceanSTARLab/RR-FBTC [2512.21486].

---

RR-FBTC establishes a principled, theoretically expressive, and computationally scalable approach for Bayesian tensor completion with automatic rank determination, providing a bridge between functional data analysis, tensor decompositions, and probabilistic machine learning [2512.21486][1505.02343][1301.7619].

Source: https://www.emergentmind.com/topics/rank-revealing-functional-bayesian-tensor-completion-rr-fbtc