Active MIMO Sensing With Exploration-Exploitation Tradeoff

1. Problem Statement

The paper addresses the challenge of active sensing in Multiple-Input Multiple-Output (MIMO) radar systems. The goal is to adaptively design transmit and receive beamformers across multiple sequential sensing stages to estimate unknown sensing parameters (specifically, the angles of arrival and reflection coefficients of multiple targets) with high accuracy.

Key constraints and challenges include:

• Hardware Limitations: The number of Radio Frequency (RF) chains ($M_T, M_R$) is strictly less than the number of antennas ($N_T, N_R$), limiting the number of simultaneous beamformers.
• Non-Convexity: When the number of beamformers is limited, minimizing the estimation error bound becomes a non-convex optimization problem.
• Exploration-Exploitation Tradeoff: The radar must balance between exploring the environment (probing multiple directions to gather new information, crucial in low Signal-to-Noise Ratio (SNR) regimes) and exploiting current knowledge (focusing energy on likely target locations to refine estimates, crucial in high SNR regimes).
• Adaptivity: Beamformers in stage $k$ must be designed based on measurements from stages $1$ to $k-1$.

2. Methodology

A. Performance Metric: Bayesian Cramér-Rao Bound (BCRB)

The authors use the Bayesian Cramér-Rao Bound (BCRB) as the optimization metric. Unlike the classical Cramér-Rao bound, the BCRB provides a lower bound on the Mean-Squared Error (MSE) averaged over the prior distribution of the unknown parameters. This is suitable for active sensing where prior statistical information (updated via posterior distributions) is available.

B. Two-Stage Optimization Framework

To address the exploration-exploitation tradeoff, the paper proposes a two-variant optimization strategy:

1. Exploitation-Centric Variant (P1): Minimizes the BCRB subject only to a total power constraint. This allows the optimizer to concentrate energy on a single (or few) optimal directions, ideal for refining estimates when targets are well-localized (High SNR).
2. Exploration-Centric Variant (P2): Minimizes the BCRB subject to a power constraint AND an orthogonality constraint on the transmit beamformers ($V_k^H V_k = \frac{P}{M_T} I$). This forces the radar to probe multiple orthogonal directions simultaneously, ensuring broad environmental coverage (Low/Moderate SNR).

The proposed active sensing strategy uses a hybrid approach: it employs the Exploration-Centric (P2) variant for an initial number of stages ($T_{explore}$) to gather broad data, followed by the Exploitation-Centric (P1) variant for the remaining stages to refine estimates.

C. Solution Algorithm: Alternating Optimization with Lagrangian Duality

The joint optimization of transmit ($V_k$) and receive ($W_k$) beamformers is non-convex. The authors propose an Alternating Optimization (AO) algorithm:

1. Fix $V_k$ and optimize $W_k$.
2. Fix $W_k$ and optimize $V_k$.
3. Repeat until convergence.

Key Theoretical Breakthrough:
While the sub-problems are non-convex (due to rank and orthogonality constraints), the authors analyze them in the Lagrangian dual domain.

• They derive the dual problem, which is unconstrained and convex.
• They prove that Global Optimality is guaranteed if specific sufficient conditions are met. These conditions relate to the multiplicity of the eigenvalues of a specific "direction matrix" derived from the optimal dual variables.
  • For the receive beamformer, global optimality holds if the $M_R$-th and $(M_R+1)$-th eigenvalues of the direction matrix are distinct.
  • For the transmit beamformer, global optimality holds if the largest eigenvalue is unique (multiplicity of 1).
• If these conditions hold, the Semidefinite Relaxation (SDR) of the problem is tight, meaning the solution to the relaxed convex problem yields the exact global optimum of the original non-convex problem.

D. Posterior Tracking

To implement the adaptive design, the system tracks the posterior distribution of the sensing parameters.

• Angles ($\phi$): Tracked via a discretized grid using Bayes' rule.
• Reflection Coefficients ($\alpha$): Tracked analytically using Kalman filtering, leveraging the linearity of the measurement model with respect to $\alpha$.

3. Key Contributions

1. Active Sensing Framework: Developed a comprehensive framework for jointly designing transmit and receive beamformers in MIMO radar using BCRB optimization, explicitly addressing the limited RF chain constraint.
2. Exploration-Exploitation Strategy: Introduced a two-phase strategy (Exploration-Centric vs. Exploitation-Centric) to manage the tradeoff. Simulations show that an initial exploration phase significantly improves performance in low-to-moderate SNR regimes.
3. Global Optimality Conditions: Provided a rigorous theoretical analysis showing that despite the non-convexity of the BCRB optimization, global optimality can be guaranteed via Lagrangian duality under specific eigenvalue multiplicity conditions. This provides a "certificate of optimality" for the proposed beamformers.
4. Insight on Beamformer Count: The analysis reveals that the minimum number of transmit beamformers required to achieve the optimal BCRB is at most equal to the multiplicity of the largest eigenvalue of the derived direction matrix. This reduces the search space for optimal solutions.
5. Tightness of SDR: Demonstrated that the Semidefinite Relaxation (SDR) of the optimization problems is tight under the derived conditions, offering an efficient computational path to the global optimum.

4. Results and Simulation Findings

• Performance vs. SNR:
  • Low/Moderate SNR: The Exploration-Centric strategy (or a hybrid with an initial exploration phase) outperforms pure exploitation strategies and state-of-the-art baselines (like random orthogonal beamforming or steering toward current estimates). This is because exploring multiple directions mitigates the risk of missing targets in noisy environments.
  • High SNR: The Exploitation-Centric strategy performs best, as the radar can confidently focus energy on the estimated target locations.
• Comparison with Deep Learning:
  • The proposed BCRB-based method outperforms Deep Learning (LSTM-based) approaches in scenarios where the number of sensing stages or channel conditions vary, as DL models require re-training for specific configurations.
  • While DL can perform well in low SNR with matched training, it lacks the generalizability and theoretical guarantees of the proposed analytical approach.
• Multiple Targets: The strategy successfully tracks and estimates parameters for multiple targets ($L=2$), with the beamforming patterns adapting to suppress interference from one target while estimating the other.
• Convergence: The alternating optimization algorithm converges to a stationary point, and increasing the number of iterations improves performance, confirming the benefit of joint transmit/receive design.

5. Significance

This paper makes a significant contribution to the field of cognitive radar and active sensing:

• Theoretical Rigor: It moves beyond heuristic or convex-relaxation-only approaches by providing conditions for global optimality in a non-convex setting, a rare achievement in beamforming literature.
• Practical Applicability: By addressing the hardware constraint of limited RF chains (where $M < N$), the solution is directly applicable to practical, cost-effective MIMO radar systems.
• Strategic Insight: It formally quantifies the "Exploration-Exploitation" tradeoff in radar sensing, demonstrating that a static "greedy" approach (always exploiting) is suboptimal in noisy environments. The proposed hybrid strategy offers a robust, model-based alternative to data-driven deep learning methods, ensuring adaptability without the need for extensive retraining.

In summary, the paper presents a mathematically sound, globally optimal (under conditions), and practically effective framework for adaptive MIMO radar sensing that balances the need to search for targets with the need to precisely locate them.