CRB-Constrained Utility Maximization

• CRB-Constrained Utility Maximization is a framework that maximizes system utility while enforcing a lower bound on estimation error as defined by the Cramér-Rao bound.
• It employs advanced techniques such as sequential linear-proximal methods and SVD-based designs to transform nonconvex CRB constraints into efficiently solvable problems.
• The framework is applied in domains like MIMO radar, integrated sensing and communication, and quantitative MRI, balancing performance tradeoffs between utility and estimation precision.

Cramér-Rao Bound (CRB)-Constrained Utility Maximization addresses the problem of maximizing a system utility function—such as communication rate, signal fidelity, or other task-oriented metrics—while explicitly controlling or minimizing the estimation error lower bound as prescribed by the Cramér-Rao bound (CRB). The CRB imposes a fundamental limit on the variance of unbiased estimators, thus serving as a rigorous surrogate for estimation accuracy. This framework is central in advanced signal processing domains such as radar waveform design, integrated sensing and communication (ISAC), and quantitative MRI, enabling a principled optimization of resources to achieve simultaneous performance in multiple objectives under fundamental statistical constraints.

1. Mathematical Formulations and Problem Classes

The general CRB-constrained utility maximization problem is formally posed as

$$\begin{aligned} &\max_{x} \quad U(x) \ &\text{subject to} \quad \operatorname{tr}\left[F(x)^{-1}\right] \leq \Gamma, \ \|x\|^2 \leq P_t, \end{aligned}$$

where $U(x)$ is the utility to maximize, $F(x)$ is the Fisher information matrix (FIM) (function of design variable $x$), and $\Gamma$ is a prescribed upper bound on the (scalarized) CRB—typically the trace of $F(x)^{-1}$, though other scalarizations (max eigenvalue, determinant, or partial traces) may also be employed (Hua et al., 2022).

In MIMO radar and ISAC contexts, the constraint is frequently combined with a transmit-power constraint $\|x\|^2 \leq P_t$, and the design variable $x$ encapsulates transmit beamformer vectors, covariance matrices, waveform coefficients, or other system resources (Zhou et al., 2024, Hua et al., 2022, Zhou et al., 17 Feb 2025).

Extensions include:

• CRB constraints as matrix inequalities, e.g., $B^T F(x)^{-1} B \preceq C$ (partial-parameter sub-blocks) (Tune, 2012).
• Scenarios where the CRB is a function of auxiliary resources, such as antenna position (movable antennas) (Guo et al., 29 Oct 2025).
• Simultaneous scalarization of multiple CRB criteria (e.g., joint angle and distance estimation) in multi-target scenarios (Zhou et al., 17 Feb 2025).

2. Structural Techniques and Algorithmic Solutions

CRB-constrained problems typically involve nonconvexity due to matrix inversion in the CRB term. Recent advances provide efficient, application-adapted algorithmic strategies:

• Sequential Linear-Proximal Methods: Transform the nonconvex CRB objective/constraint to a sequence of convex surrogates using first-order Taylor approximations and proximal regularization. Each subproblem admits a closed-form solution via the Karush-Kuhn-Tucker (KKT) conditions, leading to efficient updates with provable sublinear convergence (Zhou et al., 2024).
• Majorization-Minimization and Penalty-Dual Decomposition (PDD): For highly nonconvex problems—e.g., joint beamformer, receiver, and antenna position optimization—alternating block-wise updates are performed, with surrogate convex subproblems solved iteratively. The PDD framework enables strict CRB constraint enforcement via augmented Lagrangians, with many subproblems solvable in closed form or by fast convex quadratic solvers (Guo et al., 29 Oct 2025, Zhou et al., 17 Feb 2025).
• SVD-Based Diagonalization and Semi-Closed-Form Power Allocation: In MIMO ISAC, transmit covariance design under CRB constraints is greatly simplified via singular value decomposition (SVD) of the communication channel, reducing the optimization to scalar power allocation across decoupled subchannels. The constrained optimization is solved using Lagrange multipliers and admits water-filling-like closed-form solutions (Hua et al., 2022, Hua et al., 2022).
• Quadratic Matrix Programming: For constrained submatrices of the CRB, the computation is reformulated as a quadratic matrix program, which admits highly efficient solution via majorization-minimization, conjugate gradient, or projected block coordinate descent (Tune, 2012).

3. CRB-Constrained Utility Maximization in Key Domains

MIMO Radar and Waveform Design

Optimization of waveform or beamforming vector $p$ for MIMO radar under the trace-CRB criterion leads to: $$\min_p \ \operatorname{tr}\left[F(p)^{-1}\right], \quad \text{subject to } \|p\|_2^2 \leq P_t,$$ addressed via sequential linear approximation, proximal terms, and KKT-based updates that drastically reduce computational complexity versus SDP-based routines (Zhou et al., 2024).

Integrated Sensing and Communication (ISAC)

The canonical problem for ISAC is: $$\max_Q R(Q) \quad \text{s.t.} \ \mathrm{tr}(Q^{-1}) \leq \Gamma, \ \mathrm{tr}(Q) \leq P,$$ where $Q$ is the transmit covariance and $R(Q)$ the communication rate. The joint admissible rate–CRB region is characterized by solving this problem as $\Gamma$ varies, tracing out the Pareto frontier (Hua et al., 2022, Hua et al., 2022, Zhou et al., 17 Feb 2025).

Subspace Optimization for Quantitative MRI

The subspace matrix $U$ is selected to minimize the signal compression error while constraining the parameter CRB: $$\min_U \Delta E(U) \quad \text{s.t.} \ B_{\text{ec}}(\theta_i; U) \leq \epsilon_i \ \forall i,$$ where $\Delta E$ is preserved signal energy and $B_{\text{ec}}$ is the compressed-domain approximate CRB. The optimal subspace basis is efficiently computed via SVD on a joint data–derivative matrix, enabling joint energy/CRB minimization in a single step (Mao et al., 2023).

4. Pareto-Optimal Characterization and Tradeoff Analysis

CRB-constrained utility maximization enables rigorous exploration of the fundamental performance tradeoff between estimation accuracy (as quantified by the chosen CRB scalarization) and utility metrics (e.g., rate, signal energy). The Pareto boundary is traced by parameterizing the CRB constraint, e.g., by varying the allowed upper bound $\Gamma$, and solving the constrained maximization for each value.

In ISAC, for instance:

• Loose CRB constraint ($\Gamma$ large): The problem reduces to unconstrained utility maximization, often yielding water-filling solutions.
• Tight CRB constraint ($\Gamma$ small): Pure sensing-optimal designs dominate, typically resulting in isotropic or equal-power allocations.
• Intermediate regimes require nuanced power allocation across subspaces, quantifiably sacrificing rate for estimation accuracy (Hua et al., 2022, Hua et al., 2022).

This rigorous approach reveals nontrivial behaviors, such as the necessity of full-rank transmit covariances even when unconstrained utility maximization would prefer rank-deficient solutions. ISAC designs derived via this methodology consistently outperform benchmark methods such as time-switching or naive power-splitting (Hua et al., 2022).

5. Practical Considerations and Computational Efficiency

Implementation efficiency is critical, given the nonconvexity and dimensionality of typical CRB-constrained problems:

• Analytic and toolbox-free updates: Closed-form KKT updates, trust-region methods, and block-coordinate alternating minimization (with majorization where needed) can outperform general-purpose solvers by several orders of magnitude (Zhou et al., 2024, Guo et al., 29 Oct 2025).
• Warm-start and preconditioning: Iterative methods exploit previous solutions effectively under slow parameter drift or in batch contexts (Tune, 2012).
• SVD-based designs: Precompute key matrix decompositions to decouple variables and enable analytic or one-dimensional search for optimal resource allocation (Hua et al., 2022, Hua et al., 2022, Mao et al., 2023).

Computational experiments consistently demonstrate sublinear or near-linear convergence rates and robustness to initialization, provided proper parameter tuning and penalty/dual parameter selection (Zhou et al., 2024, Guo et al., 29 Oct 2025).

6. Generalizations, Extensions, and Emerging Directions

The CRB-constrained utility maximization framework generalizes across domains and is extensible to:

• Constrained estimation with equality/inequality parameter constraints: The CRB is modified via projection onto the feasible set using nullspace representations (Tune, 2012).
• Partial or block constraints: Only selected estimation variances (submatrix CRBs) are constrained, allowing targeted tradeoff (Tune, 2012).
• Complex CRB scalarizations: determinant, max-eigenvalue, or user-weighted traces enable risk-sensitive or fairness-driven designs (Hua et al., 2022).
• Hybrid architectures and hardware constraints: CRB–rate tradeoff with discrete analog beamforming, partially-connected architectures, and antenna position optimization are efficiently addressed via penalty-based decomposition (Zhou et al., 17 Feb 2025, Guo et al., 29 Oct 2025).
• Parametric subspace refinement: In MRI and similar applications, inclusion of derivative information (Fisher information directions) in subspace optimization achieves superior downstream estimation performance at little computational cost (Mao et al., 2023).

In all cases, the CRB constraint acts as an information-theoretic regularizer, shaping system designs to deliver guaranteed estimation fidelity under utility-oriented resource allocation.