Source-linked AI summary
Compressed Remote Sensing of Sparse Objects
Albert Fannjiang, Pengchong Yan, Thomas Strohmer
TL;DR
The paper studies how compressed sensing can recover sparse inverse-source and scattering images despite underdetermined measurements. It analyzes sensor and target ensembles, proves recovery guarantees, and identifies aperture and antenna-location effects, while restricting scattering analysis to a linear approximation and lattice-supported targets.
Problem
Inverse source and scattering imaging is nonunique, unstable, noisy, and underdetermined, motivating recovery methods that exploit sparse targets.
Method
The paper applies Basis Pursuit and compressed-sensing coherence analysis to linear inverse source and scattering models with sensor and target ensembles.
Results
Recoverable sparsity reaches O(n/(ln m)^2) for sources and O(n^2/(ln m)^2) for scatterers under product ensembles, with larger guarantees under sensor-only ensembles.
Takeaways & Limitations
The analysis identifies a threshold aperture and decoherence from random antenna locations as key factors governing sparse-target recovery.
Takeaways & Limitations
The scattering analysis uses a linear model under the Born approximation, and the target locations are assumed to lie on a square lattice.
Abstract
from arXiv · showhide
The linear inverse source and scattering problems are studied from the perspective of compressed sensing, in particular the idea that sufficient incoherence and sparsity guarantee uniqueness of the solution. By introducing the sensor as well as target ensembles, the maximum number of recoverable targets is proved to be at least proportional to the number of measurement data modulo a log-square factor with overwhelming probability. Important contributions of the analysis include the discoveries of the threshold aperture, consistent with the classical Rayleigh criterion, and the decoherence effect induced by random antenna locations. The prediction of theorems are confirmed by numerical simulations.
1. Introduction
The paper frames sparse inverse imaging as a compressed-sensing problem to address nonuniqueness, instability, noise, and limited target recovery. It replaces computationally infeasible ℓ0 optimization with Basis Pursuit under physical sensing constraints.
- Motivation: Inverse source and scattering imaging is challenging because it is nonunique, unstable, noisy, and underdetermined.Traditional methods can miss targets and produce artifacts that obscure target images.
- Motivation: Sparse targets motivate applying compressed sensing to recover signals from severely underdetermined measurements.The model uses Y = AX with n ≪ m, so unique recovery is generally impossible without exploiting sparsity.
- Method: Basis Pursuit provides a convex relaxation of the NP-hard ℓ0 recovery problem.It can be solved using linear and quadratic programming techniques.
- Method: Under suitable matrix and sparsity conditions, ℓ1 and ℓ0 minimization have the same unique solution.Restricted Isometry is one sufficient condition, with recovery possible for sparsity up to O(n/log(m)) in stated random-matrix settings.
- Physical constraints: Physical sensing matrices cannot be chosen arbitrarily because they are determined by wave propagation and controllable parameters such as wavelength, sensor locations, sensor count, and aperture.This physical constraint complicates practical compressed-sensing realization.
2. Problem formulations and main results
The paper formulates linear inverse source and scattering imaging in compressed-sensing terms using randomized sensor and target ensembles. It establishes sparsity-dependent recovery guarantees, identifies aperture thresholds, and confirms the predictions numerically.
- Problem formulation: Randomly placing sensors uniformly within a fixed square aperture constructs the sensor ensemble used in the recovery analysis.The sensor locations are independent and identically distributed in the array plane.
- Problem formulation: The analysis assumes targets lie on a transverse square lattice and models sparse target vectors with randomly selected supports and independent uniformly distributed phases.The sparsity is the number of nonzero target amplitudes.
- Inverse source results: For inverse source imaging, product ensembles recover sparsity up to O(n/(ln m)^2) by BP, while sensor-only ensembles recover up to O(√n) by BP and OMP.These guarantees hold with overwhelming probability.
- Aperture and validation: The threshold aperture is λz0/ℓ, consistent with the classical Rayleigh criterion, while synthetic aperture achieves comparable resolving performance with half the aperture.The numerical simulations indicate that the stated aperture condition is sufficient for the predicted inverse-source performance.
- Inverse Born scattering results: For response-matrix imaging, product ensembles recover scatterers up to O(n^2/(ln m)^2), while sensor-only ensembles recover up to O(n) by BP and OMP.The response matrix contains n^2 transmitter-receiver measurements.
- Inverse Born scattering results: For synthetic-aperture imaging, product ensembles recover up to O(n/(ln m)^2) and sensor-only ensembles up to O(√n), both with overwhelming probability.Synthetic aperture uses one physical antenna at multiple transmit-receive positions.
3. Source inversion
The source-inversion analysis bounds sensing-matrix coherence and spectral norm under randomized sensors, establishing exact sparse recovery by BP and OMP. It also identifies an aperture condition tied to the Rayleigh resolution criterion.
- Recovery guarantees: For the product ensemble of sources and sensors, BP exactly recovers sources of sparsity up to O(n/(ln m)2) with overwhelming probability.The guarantee follows from the coherence and spectral-norm estimates together with the target and sensor randomness.
- Coherence estimate: Random sensor locations yield coherence bounds that are optimal up to a constant factor.The analysis uses independent, uniformly distributed sensor coordinates and shows the coherence has the expected order under the stated aperture condition.
- Coherence estimate: The optimal aperture is defined by the condition corresponding to ρ = 1 and matches the classical Rayleigh resolution criterion.The paper interprets the resulting scale as the imaging-system resolution.
- Spectral norm bound: The sensing matrix has full rank and satisfies a spectral-norm bound under the random sensor ensemble.This norm estimate supplies the second ingredient used with coherence to establish sparse-recovery guarantees.
- Recovery guarantees: With only the sensor ensemble randomized, BP and OMP exactly recover all sources of sparsity up to O(√n) with overwhelming probability.The result applies to the source-inversion setting after the matrix estimates are combined with standard sparse-recovery implications.
4. Inverse Born scattering
For inverse Born scattering, the paper extends the randomized-sensor compressed-sensing analysis to response-matrix and synthetic-aperture imaging. The resulting theorems provide exact recovery guarantees whose sparsity scaling depends on the imaging setting and ensemble assumptions.
- Response matrix imaging: With only the sensor ensemble randomized, response-matrix scatterers of sparsity up to O(n) are exactly recoverable by BP and OMP with overwhelming probability.The guarantee follows from the response-matrix coherence and spectral-norm estimates.
- Response matrix imaging: For response-matrix imaging, product-ensemble scatterers of sparsity up to O(n2/(ln m)2) are exactly recoverable by BP with overwhelming probability.The theorem assumes randomly drawn targets and antenna arrays under the stated regime conditions.
- Synthetic aperture imaging: For synthetic-aperture imaging, product-ensemble scatterers of sparsity up to O(n/(ln m)2) are exactly recoverable by BP with overwhelming probability.The result follows from the correspondence between synthetic-aperture imaging and the inverse-source setting.
- Synthetic aperture imaging: With only the sensor ensemble randomized, synthetic-aperture scatterers of sparsity up to O(√n) are exactly recoverable by BP and OMP with overwhelming probability.The theorem applies under the corresponding aperture and regime conditions.
5. Numerical simulations
Numerical experiments test the theoretical aperture and sparsity predictions using exact and paraxial sensing matrices in source and scattering settings. The results show agreement with the predicted trends and demonstrate recovery beyond conventional matched field processing in one example.
- Simulation setup: A = 100 is used as the threshold, optimal aperture in the simulations, enforcing the relevant paraxial-regime conditions.The simulations set z0 = 10000 and mostly use λ = 0.1.
- Simulation assumptions: The simulations include model mismatch by propagating with the true Green function while inverting with its paraxial approximation, yet performance degradation remains manageable.The scattering simulations also include a Born-approximation mismatch through the Foldy–Lax formulation.
- Source inversion: The exact and paraxial coherence curves closely track one another and flatten near and beyond A = 100.This behavior agrees with the theoretical optimal-aperture prediction across the tested paraxial regime.
- Response matrix imaging: 35 scatterers are nearly exactly recovered with 20 antennas in the response-matrix setup.The sensing matrix has size 400 × 2500, while matched field processing is shown for comparison.
- Scattering simulations: The empirical maximum number of recoverable scatterers varies with antenna count consistently with the theoretical predictions for both RM and SA imaging.The comparison includes paraxial and exact sensing matrices at A = 100.
6. Conclusions
The paper studies inverse source and scattering imaging through compressed sensing, proving sparse-target recovery guarantees while identifying aperture and antenna-location effects. Its scope is limited for extended targets, which require a different approach.
- Conclusions: The analysis uses incoherence and sparsity to guarantee uniqueness, with sensor and target ensembles yielding recoverable-target counts proportional to measurements up to a log-square factor.The guarantee holds with overwhelming probability.
- Conclusions: The study covers inverse source, response-matrix scattering, and synthetic-aperture imaging, identifying antenna-location decoherence and threshold apertures ρ = 1 and ρ = 1/2, respectively.Here ρ = λz0/(Aℓ).
- Conclusions: The paper considers localization and amplitude estimation for point targets, not extended targets.Extended targets are identified as a scope boundary for the presented approach.
- Conclusions: Modeling extended targets as closely spaced point targets makes ρ ≫1, requiring unbounded aperture and antenna count, so a different approach is needed.The authors state that this is not feasible for extended-target imaging via compressed sensing.
Appendix A. Restricted isometry property (RIP)
The RIP appendix explains how restricted isometry supports unique recovery and applies this framework to randomly drawn sensor arrays for source and scattering problems. It also contrasts RIP’s broad guarantees with limitations for response-matrix imaging.
- RIP foundations: Under the stated RIP inequality, a target vector of sparsity at most s is the unique solution of basis pursuit.This is the fundamental RIP recovery result stated in Theorem 9.
- Source inversion: For randomly drawn sensor arrays, source amplitudes of sparsity less than s can be uniquely determined from basis pursuit with probability at least 1 − ϵ under the theorem’s condition.The result is stated for the source-inversion sensing matrix.
- Scattering: For randomly drawn sensor arrays, scatter amplitudes of sparsity less than s can likewise be uniquely determined from basis pursuit with probability at least 1 − ϵ under the theorem’s condition.The appendix states this result for the scattering setting.
- Discussion: RIP can guarantee uniqueness for all targets of sparsity at most s without introducing a target ensemble and also guarantees stability to noise, but it does not seem amenable to response-matrix imaging.The appendix identifies these as advantages and a limitation of the RIP approach.
Appendix B. Proof of Theorem 1
Appendix B proves Theorem 1 by combining random-column and target-ensemble propositions with spectral and singular-value estimates. The resulting probability bound establishes basis-pursuit uniqueness except on a controlled failure event.
- Proof strategy: Theorem 1 follows from two propositions due to Tropp concerning randomly selected columns and targets drawn from the target ensemble.The proof uses these propositions as its principal ingredients.
- Conclusion: The resulting argument states that X is the unique solution of basis pursuit except with probability 2ϵ.This conclusion follows from the target-ensemble proposition and associated bounds.
- Proof strategy: The proof combines the preceding estimates with α = 1/2 and a probability bound for the relevant event that the target is the unique basis-pursuit solution.The event estimate is the final substantive step of the proof.
Appendix C. Matched field processing
The appendix describes conventional matched field processing as a Bartlett ambiguity-surface method based on a matched-filter weight vector. It contrasts this linear processor with nonlinear compressed-sensing processors.
- Background: Matched field processing is used for source localization in underwater acoustics and is closely related to the matched filter.The appendix introduces it as conventional processing for source localization.
- Bartlett processor: The conventional method uses the Bartlett processor, formulated through an ambiguity-surface optimization over a normalized matched-filter weight vector.The weight vector is W = Y/∥Y∥2.
- Scattering setup: For inverse scattering in the response-matrix setup, the ambiguity surface is the sum of the ambiguity surfaces associated with the probe signals.There are n measurement vectors corresponding to n probe signals.
- Comparison: Compressed-sensing processors based on ℓ1-minimization or greedy algorithms are nonlinear, unlike the conventional matched field processor.The appendix makes this contrast explicitly.