Adaptive Sensing beyond Non-Adaptive Information Limits: End-to-End Co-Design of Geometry, Policy, and Inference

This paper introduces "joint dynamic programming," a co-design framework that simultaneously optimizes continuous hardware geometry and adaptive measurement policies to significantly outperform traditional non-adaptive or separately optimized approaches in sensing tasks, as demonstrated by substantial error reductions across radar, quantum, and photonic sensor case studies.

The Gist

Imagine you are trying to find a hidden treasure in a large, foggy field. You have a metal detector (the sensor) and a map (the algorithm).

For a long time, engineers and scientists treated these two parts separately:

1. The Hardware Team built the best possible metal detector they could, hoping it would catch every signal.
2. The Software Team wrote a smart computer program to interpret the signals and guess where the treasure is.

The problem is that if the hardware misses a signal because it was poorly designed, no amount of smart software can fix it. The information is gone forever.

This paper proposes a radical new way to build sensors: Stop designing the hardware and the software separately. Design them together, at the same time.

Here is the breakdown of their idea using simple analogies:

1. The "Smart Detective" vs. The "Stiff Robot"

Imagine two detectives trying to find a suspect in a city.

• The Stiff Robot (Old Way): This detective has a fixed plan. "I will walk down Main Street, then Oak Street, then Elm Street," regardless of what they see. Even if they spot a clue on Main Street that proves the suspect is on Elm Street, they stick to the plan because their "hardware" (their legs) was built for a specific route.
• The Smart Detective (New Way): This detective adapts. If they see a clue on Main Street, they immediately change their plan to head to Elm Street.

The paper argues that you shouldn't just build a "Smart Detective" (an adaptive algorithm) and give them a "Stiff Robot's legs" (fixed hardware). Instead, you should design the legs specifically to help the detective change direction quickly. The shape of the legs should depend on the detective's strategy.

2. The "Co-Design" Secret Sauce

The authors created a mathematical method called Joint-DP (Joint Dynamic Programming). Think of this as a super-smart coach who trains both the detective and the legs simultaneously.

• The Coach's Job: The coach asks, "If I change the shape of the metal detector's antenna (the hardware), how does that change the best strategy for the detective?"
• The Loop: The coach tweaks the hardware, calculates the best new strategy for the detective, sees how well they do, and then tweaks the hardware again. They repeat this until the pair works perfectly together.

3. Why Old Methods Failed (The "Perfect Information" Trap)

In the past, scientists tried to guess the best hardware by asking: "What if we knew exactly where the treasure was? What hardware would be best then?" They called this the "Expected Value of Perfect Information."

The paper shows this is a trap.

• The Analogy: Imagine you are playing a game of "20 Questions." If you knew the answer was "A Cat," you would ask very specific questions. But since you don't know the answer, asking those specific questions is a waste of time. You need to ask broad questions first to narrow it down.
• The Result: The "Perfect Information" method designs hardware for a scenario that never happens (knowing the answer). The new "Joint-DP" method designs hardware for the real scenario (not knowing the answer), where the detective needs to adapt.

4. The Results: Big Wins in Three Scenarios

The paper tested this "Co-Design" method on three very different problems, and the results were massive:

• Scenario A: Radar Searching for a Target

  • The Setup: A radar trying to find a plane in a ring of 16 possible spots.
  • The Result: The old method (designing hardware first) was 2.8 times worse at finding the target than the new co-designed method. The new method learned to "zoom in" on the right spot much faster.

• Scenario B: Quantum Sensors (Superconducting Qubits)

  • The Setup: Measuring tiny magnetic fields using quantum particles.
  • The Result: The new method reduced the error by 11.3 times compared to the best previous method. It was like going from a blurry photo to a crystal-clear image.

• Scenario C: Photonic Metasensors (Light Sensors)

  • The Setup: A massive sensor with 90,000 tiny pixels designed to manipulate light.
  • The Result: This is the biggest win. The new method reduced the error by 123 times compared to a random design. It turned a sensor that was barely working into one that was incredibly precise.

5. How They Did It (The "Freeze" Trick)

You might wonder: "How do you mathematically optimize something that changes its mind every second?"

The authors used a clever math trick called the Envelope Theorem.

• The Analogy: Imagine you are climbing a mountain (optimizing the hardware). Usually, the path up the mountain changes as you move (the strategy changes). This makes it hard to calculate the slope.
• The Trick: The authors realized that at the very top of the hill (the best strategy), the path doesn't actually change because of your next step. So, they "froze" the strategy in place just long enough to calculate the slope of the mountain. This allowed them to use standard computer tools to find the perfect hardware shape without getting stuck in a math loop.

Summary

The paper's main message is simple: Don't build a tool and then teach it how to use itself. Build the tool for the way it will be used.

By designing the physical shape of the sensor and the adaptive strategy of the computer at the same time, they achieved results that were 10 to 100 times better than anything possible when the two were designed separately. This is a fundamental shift from "hardware first, software later" to "hardware and software as one team."

Technical Summary

1. Problem Statement

The paper addresses a fundamental gap in sensor design: the separation of hardware geometry optimization (inverse design) and adaptive measurement policy (sequential decision making).

• The Limitation: Traditional approaches optimize hardware (e.g., antenna patterns, metasurfaces, quantum circuits) assuming a fixed, non-adaptive measurement schedule. Conversely, adaptive sensing literature (e.g., POMDPs) assumes the hardware is immutable and optimizes only the policy.
• The Core Issue: In sensing, information lost at the analog-to-digital boundary cannot be recovered by downstream algorithms. If the hardware geometry is not designed to support the specific logic of an adaptive policy, the system fails to reach its theoretical performance limits.
• The Goal: To formulate and solve a joint optimization problem where the continuous physical geometry ($c$) and the Bellman-optimal adaptive measurement policy ($\pi$) are co-designed as a single objective.

2. Methodology: Joint Dynamic Programming (Joint-DP)

The authors propose a framework called Joint-DP, which treats the hardware parameters and the measurement policy as co-equal decision variables in a single optimization loop.

A. Mathematical Formulation

The problem is defined as maximizing a scalar objective $V(c, \pi)$:
$$(c^\star, \pi^\star) = \arg \max_{c \in \mathcal{C}, \pi \in \Pi} \mathbb{E}_{x, y} [\Phi(x, s_{1:K}, c)]$$
Where:

• $c$: Continuous physical geometry parameters.
• $\pi$: Adaptive policy mapping belief states to actions.
• $\Phi$: A pointwise scalar reward (e.g., Mutual Information, Bayesian MSE, or Fisher Information).
• The inner problem (finding $\pi^\star$ for a fixed $c$) is solved via Bellman Dynamic Programming (DP) on the belief state.
• The outer problem (finding $c^\star$) requires computing the gradient of the inner DP value with respect to $c$.

B. Differentiable Dynamic Programming with Sharp-Max

A major technical hurdle is differentiating through the non-smooth max operator in the Bellman recursion.

• The Solution: The authors utilize the Envelope Theorem. Instead of smoothing the max (e.g., via softmax), they keep the operator "sharp."
• Mechanism: At the optimal policy $\pi^\star(c)$, the sensitivity of the value function to changes in the policy vanishes (first-order optimality condition). Therefore, the gradient of the total value with respect to hardware $c$ is simply the partial derivative of the Bellman backup, holding the optimal action fixed.
• Implementation:
  1. Forward Pass: Solve the inner Bellman DP to find the optimal policy path (argmax) for a given $c$.
  2. Freeze: "Freeze" the argmax decisions (treat them as constants).
  3. Backward Pass: Apply standard reverse-mode automatic differentiation (AD) through the likelihood and belief update functions. This yields an exact, unbiased gradient without smoothing bias or temperature schedules.

C. Hierarchy of Relaxations

To scale from small discrete problems to high-dimensional continuous hardware (e.g., $10^5$ pixels), the authors introduce a "relaxation cascade" for the inner policy solver:

1. Rung 1 (Exact DP): Exact Bellman DP for discrete states or small continuous grids (used in Case Studies A & B).
2. Rung 2 (Certainty-Equivalent): Uses a point estimate of the state to plan a fixed schedule.
3. Rung 3 (Myopic): One-step lookahead optimization.
4. Rung 4 (Fisher-Trace + EKF): Replaces the full belief with a Gaussian approximation (Extended Kalman Filter) and optimizes the trace of the Fisher Information Matrix. This enables scaling to massive continuous hardware spaces (Case Study C).

3. Key Contributions

1. Unified Framework: The first formulation to co-optimize continuous physical hardware and a multi-step Bellman-optimal adaptive policy in a single optimization problem.
2. Sharp-Max Differentiable DP: A novel method for computing exact gradients through Bellman recursions using the Envelope Theorem, avoiding the bias inherent in softmax-regularized differentiable DP.
3. Scalar Reward Family: A unified treatment of three distinct reward families (Mutual Information, Bayesian MSE, and Fisher Information) under a single pointwise reward formalism $\Phi$.
4. Scalability: Demonstrated the ability to scale from small discrete POMDPs to $90,000$-dimensional photonic inverse design problems via the relaxation hierarchy.

4. Experimental Results

The paper validates Joint-DP across three distinct case studies, showing significant performance gains over non-adaptive baselines.

Case Study	System	Baseline Comparison	Result
A: Radar Beam Search	Discrete state (16 cells), 4 steps.	EVPI-guided geometry: Selects geometry maximizing Expected Value of Perfect Information (a common heuristic).	Joint-DP achieves 2.8$\times$ higher attainable adaptive value. EVPI picks the wrong geometry (extreme narrow/wide beams) because it ignores the policy's inability to replicate clairvoyance.
B: Superconducting Qubit	Continuous flux sensor, 4 steps.	Joint Bayesian Cramér-Rao (PCRB): Co-optimizes geometry and a fixed-schedule policy.	Joint-DP reduces deployed Mean Squared Error (MSE) by 11.3$\times$ (at $z=+132\sigma$). The fixed schedule fails to resolve aliasing; the adaptive policy uses short delays first to disambiguate, then long delays for precision.
C: Photonic Metasensor	90,000-pixel permittivity design, continuous state.	Randomized Baseline: Random geometry with EKF reconstruction.	Joint-DP reduces deployed MSE by 123$\times$. The optimized geometry acts as a task-aware filter, concentrating information in the least-constrained directions of the state space.

5. Significance and Implications

• Redefining Hardware Role: The paper argues that hardware should not be a passive conduit but an active participant in inference. Joint optimization shapes the physical medium to "pre-process" information in a way that is specifically tailored to the downstream adaptive policy.
• Failure of Heuristics: The results demonstrate that standard design heuristics (like maximizing EVPI or Fisher Information with a fixed schedule) are fundamentally flawed for adaptive systems. They often select "extreme" geometries that look good on paper but are unusable by a real-world adaptive agent.
• Bridging Communities: The work bridges the gap between Inverse Design/Photonics (hardware optimization) and Sequential Decision Theory/POMDPs (policy optimization), providing a mathematical and computational bridge for both fields.
• Practical Impact: For any sensor where the hardware is fixed at manufacturing but the software runs for the device's lifetime (e.g., cognitive radar, quantum sensors, programmable metasurfaces), Joint-DP represents the "minimum principled procedure" for design.

In summary, the paper proves that co-designing the sensor's body (geometry) and its brain (policy) yields orders-of-magnitude improvements in sensing performance, a feat impossible when optimizing either component in isolation.