Constituent Prior Matrix for Matrix Completion

• Constituent Prior Matrix is a rank-r matrix built from estimated row and column subspaces that encodes structural priors for matrix completion.
• Its incorporation in optimization frameworks augments nuclear-norm minimization to improve correlation with prior estimates and reduce sample complexity.
• Empirical results on both synthetic and real-world data demonstrate improved recovery performance and lower reconstruction errors with accurate priors.

A constituent prior matrix refers to a matrix constructed from estimated row and column subspaces to encode prior structural information for use in matrix completion problems. Specifically, when the task is to recover a low-rank matrix from a subset of its entries, incorporating subspace priors via a constituent prior matrix provides a mechanism to reduce the sample complexity of recovery. This approach integrates prior subspace estimates into the objective function by maximizing the correlation between the candidate solution and the prior, enabling improvements in both theoretical guarantees and empirical performance (Zhang et al., 2020).

1. Matrix Completion and the Role of Priors

Matrix completion is the problem of reconstructing an unknown low-rank matrix $X^\star \in \mathbb{R}^{n \times n}$ given only a subset of its entries. Under standard assumptions—such as incoherence and uniformly random sampling—the canonical approach is nuclear-norm minimization: $$\min_{X\in\mathbb{R}^{n\times n}} \|X\|_{*} \quad\text{s.t.}\quad Y = \mathcal{R}_p(X),$$ where $\mathcal{R}_p$ is the Bernoulli-sampling operator and $Y$ contains the observed entries. This formulation leverages only the observed data, omitting any auxiliary structural information.

However, in many applications, approximate knowledge about the subspace structure of $X^\star$ can be obtained from historical data, domain knowledge, or side information. The constituent prior matrix $M_0$ formalizes the incorporation of such prior subspace information, offering a principled modification to the standard matrix completion pipeline.

2. Construction of the Constituent Prior Matrix

Given estimates $\widetilde U_{r} \in \mathbb{R}^{n\times r}$ and $\widetilde V_{r} \in \mathbb{R}^{n\times r}$ for the true row and column subspaces $U_{r}, V_{r}$ of $X^\star$, the constituent prior matrix is defined as

$M_0 = \widetilde U_{r} \widetilde V_{r}^T.$

This construction encodes the estimated subspace structure as a rank-$r$ matrix. In symmetric scenarios, $M_0$ is taken as $M_0 = \widetilde U_{r} \widetilde U_{r}^T$, ensuring symmetry. The quality of $M_0$ is quantified via principal angles $\{\gamma_{i}\}$ between the estimated and true subspaces; subspace errors are small when $\sin\gamma_{i} \ll 1$ (Zhang et al., 2020).

3. Optimization Formulation Incorporating the Constituent Prior

The prior is incorporated into the matrix completion problem by augmenting the objective function with a correlation-maximizing term: $$\min_{X\in\mathbb{R}^{n\times n}} \|X\|_{*} - \lambda \langle X, M_0\rangle \;\;\text{s.t.}\;\; Y = \mathcal{R}_p(X),$$ or

$\|X\|_{*} - \lambda\,\mathrm{tr}(M_{0}^{T}X),$

where $\lambda \geq 0$ is a regularization parameter trading off between low-rank structure and alignment with the prior. This convex formulation encourages the recovered matrix to both be low-rank and to correlate with the constituent prior matrix, integrating data-driven and prior-driven information (Zhang et al., 2020).

4. Sample Complexity and Performance Guarantees

The theoretical contribution of the constituent prior matrix lies in its reduction of the sample complexity required for exact recovery. Let $X^\star$ be a rank-$r$ incoherent matrix with leverage scores $\{\mu_i\}, \{\nu_j\}$. Four alignment measures quantify how well $M_0$ matches the true support $\mathcal{T}$:

• $$\alpha_1 = \| U_r V_r^T - \lambda \mathcal{P}_{\mathcal{T}}(M_0) \|_F$$
• $$\alpha_2 = \lambda \| \mathcal{P}_{\mathcal{T}^\perp}(M_0) \|$$
• $$\xi_1 = \| U_r V_r^T - \lambda \mathcal{P}_{\mathcal{T}}(M_0) \|_{\mu(\infty)}$$
• $$\xi_2 = \| U_r V_r^T - \lambda \mathcal{P}_{\mathcal{T}}(M_0) \|_{\mu(\infty,2)}$$ (see Theorem 1 in (Zhang et al., 2020)).

A typical sampling probability threshold is

$$p_{ij} \gtrsim \max\Big\{\log\big(\frac{\alpha_1^2 n}{r\log n}\big), 1\Big\} \cdot \frac{(\mu_i+\nu_j) r \log n}{n} \cdot \max\{(2\xi_1+\xi_2)^2,1\}.$$

In the absence of a prior ($\lambda=0$), $m \sim O(rn\log^2 n)$ samples suffice (recovering conventional bounds). In the presence of a highly accurate prior ($\alpha_1\to 0$, $\xi_1,\xi_2\to 0$, $\lambda=1$), sample complexity reduces to $m \sim O(rn\log n)$, a full log-factor improvement. These bounds hold for both noiseless and noisy settings (Zhang et al., 2020).

5. Empirical Demonstrations

Synthetic and real-world experimental results substantiate the theoretical improvements offered by constituent prior matrices (Zhang et al., 2020). In synthetic examples with $n=32$, $r=4$, perturbed priors ($\widehat X = X^\star + \sigma Z$ with $\sigma=0.01$ or $0.1$) yield constituent prior matrices via rank-$r$ SVD. Across varying sampling fractions, the proposed max-correlation method ($\lambda$ optimally selected) consistently outperforms baseline matrix completion, pushing the $50\%$-success threshold from approximately $0.45$ down to $0.30$ for accurate priors.

Real-data experiments on the Wine and Iris datasets further confirm lower relative reconstruction errors (measured as $\|X^\star - \widehat X\|_F / \|X^\star\|_F$) versus both standard and weighted matrix completion, especially under low sampling ratios.

6. Summary, Scope, and Extensions

The constituent prior matrix $M_0 = \widetilde U_{r}\widetilde V_{r}^T$ encapsulates estimated subspace information, enabling its integration into matrix completion by augmenting the nuclear-norm objective with a linear alignment term. Sufficiently accurate priors yield provable sample complexity reduction from $O(rn\log^2 n)$ to $O(rn\log n)$, approaching optimality under standard incoherence/Bernoulli sampling assumptions. These findings generalize to both symmetric/asymmetric cases and to additive noise (Zhang et al., 2020).

A plausible implication is that the use of constituent prior matrices provides a flexible mechanism for improvement when statistical or empirical subspace estimates are available, but its efficacy depends critically on the prior's accuracy as quantified by principal angles. Empirical and theoretical results jointly demonstrate substantial gains in efficiency and reconstruction accuracy across synthetic and real-world scenarios.