Laplacian WMSE: Graph Signal Recovery

• Laplacian WMSE is a quadratic cost function that generalizes MSE by incorporating the graph Laplacian to penalize high-frequency estimation errors.
• It employs Dirichlet energy and a Lehmann-unbiasedness condition to enforce smoothness and appropriately account for the graph's spectral structure.
• The framework defines a Cramér–Rao bound reflecting spectral penalties, which informs sensor placement and bandlimited sampling strategies.

The Laplacian weighted mean-squared error (Laplacian WMSE) is a quadratic cost function for signal recovery on graphs that generalizes the classical mean-squared error by incorporating the graph Laplacian. In this framework, estimation errors are penalized in accordance with both the topological connectivity and spectral structure of the underlying graph, allowing the error metric to enforce smoothness and penalize high graph-frequency components more heavily. The Laplacian WMSE supports a non-Bayesian estimation theory that includes a Dirichlet energy interpretation, a generalized notion of unbiasedness, and corresponding Cramér–Rao bounds with direct implications for the design of sensing and sampling strategies (Routtenberg, 2020).

1. Definition and Mathematical Formulation

Let $\theta \in \mathbb{R}^M$ denote the true graph signal and $\hat\theta$ its estimator, with the associated error $e = \hat\theta - \theta$. For a weighted undirected graph with Laplacian $L$, the Laplacian WMSE is defined as

$WMSE(\hat\theta;\theta) = \mathbb{E}\{ e^T L e \}.$

Explicitly, this can be written as

$e^T L e = \sum_{(i,j)\in E} w_{ij} (e_i - e_j)^2,$

where $w_{ij}$ denotes the weight of edge $(i,j)$ and $E$ is the edge set. The Laplacian WMSE thus penalizes differences in the estimation error across edges, weighting them by the connectivity and strength of the graph.

2. Dirichlet Energy and Spectral Structure

The Laplacian $L$ is symmetric and positive semidefinite and admits the eigendecomposition $L = U \Lambda U^T$, with $$\Lambda = \operatorname{diag}\{0, \lambda_2, \ldots, \lambda_M\}$$ and $U$ the orthonormal eigenvector basis. The cost

$$e^T L e = \tilde{e}^T \Lambda \tilde{e} = \sum_{m=1}^M \lambda_m \tilde{e}_m^2,$$

where $\tilde{e} = U^T e$ are the graph Fourier coefficients. This cost equals the Dirichlet energy of the error and acts as a spectral penalty: high graph-frequency errors (corresponding to large $\lambda_m$) are penalized more than smooth modes. The $\lambda_1 = 0$ term corresponds to the constant (DC) mode and is unpenalized, reflecting that constant offsets in the error vector are not “visible” to this metric.

3. Lehmann-Unbiasedness Relative to the Laplacian

Traditional unbiasedness, defined by a zero mean error, is inadequate because constant offsets are not penalized by Laplacian WMSE. Instead, one adopts a Lehmann-unbiasedness condition: an estimator $\hat\theta$ is “graph-unbiased” if

$$U^T \mathbb{E}[ \hat\theta - \theta ] = 0 \quad\Longleftrightarrow\quad \mathbb{E}[ \tilde{e}_m ] = 0 \text{ for } m=2,\ldots,M.$$

This means the expected error is required to be orthogonal to all non-constant eigenvectors of the Laplacian. For constrained estimation (e.g., bandlimitedness or anchor constraints), $U$ is replaced by a basis of the appropriate row space, and the orthogonality condition is imposed with respect to this basis.

4. Cramér–Rao Bound for Laplacian WMSE

Given a likelihood $f(x;\theta)$ with Fisher information matrix $$J(\theta) = \mathbb{E}[ \nabla_\theta \log f \nabla_\theta \log f^T ]$$, and under regularity and graph-unbiasedness, the covariance of any such estimator satisfies

$$\mathbb{E}[ e e^T ] \succeq U (U^T J U)^\dagger U^T.$$

Consequently, the Laplacian WMSE admits the lower bound

$$\mathbb{E}[ e^T L e ] \ge \operatorname{Tr} \left[ \Lambda^{1/2} (U^T J U)^\dagger \Lambda^{1/2} \right ] = \operatorname{Tr}\left[ (U^T J U)^\dagger \Lambda \right].$$

Equality is achieved when the “graph-frequency error” aligns with the score function, i.e.,

$$\Lambda^{1/2} U^T (\hat\theta - \theta) = \Lambda^{1/2} (U^T J U)^\dagger U^T \nabla_\theta \log f(x;\theta).$$

This strengthens the classical Cramér–Rao bound by incorporating the graph spectral structure and weighting errors accordingly.

5. Implications for Sensing, Sampling, and Design

The Laplacian WMSE framework has direct implications for sensor placement and measurement selection. The lower bound shows that increasing the Fisher information in high-$\lambda_m$ modes decreases the minimal achievable WMSE. Two principal cases from (Routtenberg, 2020) illustrate these principles:

• Relative-measurement networks: For observation models $h_{ij} = w_{ij}(\theta_i - \theta_j) + \nu_{ij}$, $J = (1/\sigma^2) L_{\mathrm{meas}}$. The bound becomes

$$\mathbb{E}[e^T L_{\mathrm{phys}} e] \ge \sigma^2 \operatorname{Tr}[ L_{\mathrm{meas}}^\dagger L_{\mathrm{phys}} L_{\mathrm{meas}}^\dagger ].$$

When measurements are restricted to a spanning tree, minimization reduces to finding a maximum-weight spanning tree with edge weights $w_{ij}^2$.

• Bandlimited sampling: If $\theta$ is confined to the span of the first $R$ eigenvectors $U_R$ and sampled at nodes $S$ via $\Phi$, the bound is

$$\mathbb{E}[e^T L e] \ge \sum_{m=2}^R \lambda_m \left[ (\Phi U_R)^T \Sigma^{-1} (\Phi U_R) \right]_{mm}^{-1},$$

where $\Sigma$ is the noise covariance. Greedy node selection to minimize this bound explicitly accounts for spectral penalties and can outperform standard A- and E-optimality criteria.

6. Theoretical Significance and Summary

Adopting Laplacian WMSE for signal recovery on graphs yields a unified and geometrically principled estimation framework. Key properties include:

• Enforces smoothness via Dirichlet energy of estimation errors.
• Weights graph-Fourier modes according to their respective graph-frequency penalties.
• Employs a Lehmann-unbiasedness condition suited to the nullspace inherent in graph Laplacians.
• Provides a closed-form expression for the Cramér–Rao bound reflecting the spectral structure.
• Guides node and edge selection through minimization of a spectral, $\lambda$-weighted inverse information criterion.

This framework naturally accommodates the intrinsic ambiguities and spectral structure of graph signal recovery problems, particularly where estimation is only defined up to a constant offset (Routtenberg, 2020).