Source-linked AI summary

Super-Resolution from Noisy Data

Emmanuel Candes, Carlos Fernandez-Granda

arXiv:1211.0290v3cs.ITmath.NA

TL;DR

The paper asks how to recover a finer-scale spectrum of multiple point sources from noisy low-frequency observations. It uses convex programming for total-variation recovery and proves stable super-resolution when sources are sufficiently separated, with error scaling quadratically in SRF and proportionally to noise. The analysis is worst-case for bounded, otherwise arbitrary noise and focuses on a particular recovery problem and perturbation model.

  • Problem

    The problem is to estimate a wider spectral band and finer-resolution signal from noisy observations restricted to a low-pass band.

  • Method

    The paper uses a convex program that minimizes the total-variation norm subject to constraints from the noisy bandlimited data.

  • Results

    The smoothed reconstruction error is proportional to the noise level and scales quadratically with the super-resolution factor under a minimum-separation condition.

  • Takeaways & Limitations

    Convex optimization provides a tractable and stable way to approximate the higher-resolution spectrum of sufficiently separated point sources.

  • Takeaways & Limitations

    The analysis studies one recovery problem with bounded L1-norm perturbations, although the techniques may extend to other schemes and noise models.

Abstract

from arXiv · show

This paper studies the recovery of a superposition of point sources from noisy bandlimited data. In the fewest possible words, we only have information about the spectrum of an object in a low-frequency band bounded by a certain cut-off frequency and seek to obtain a higher resolution estimate by extrapolating the spectrum up to a higher frequency. We show that as long as the sources are separated by twice the inverse of the cut-off frequency, solving a simple convex program produces a stable estimate in the sense that the approximation error between the higher-resolution reconstruction and the truth is proportional to the noise level times the square of the super-resolution factor (SRF), which is the ratio between the desired high frequency and the cut-off frequency of the data.

1 Introduction

The paper addresses stable super-resolution of multiple point sources from noisy low-frequency data, estimating a wider spectral band than measurements provide. Under source-separation conditions, convex recovery yields error proportional to noise and SRF^2.

  • Super-resolution seeks high-resolution information from coarse-scale measurements, with applications including imaging, astronomy, medical imaging, and microscopy.
  • The paper focuses on superpositions of point sources, for which precise recovery of many source locations is less analyzed than distinguishing two blurred sources.
  • Noisy observations provide the spectrum only in the low-pass band [−f_lo, f_lo], while the goal is estimating it over the wider band [−f_hi, f_hi].The target resolution satisfies λ_hi ≪ λ_lo, with f_hi > f_lo.
  • The super-resolution factor compares desired high resolution with measured low resolution, so larger SRF represents a greater resolution increase.An SRF of two doubles resolution, while an SRF of four quadruples it.
  • The recovery method minimizes the total-variation norm subject to data constraints, using convex programming and remaining agnostic to the target resolution.The total-variation norm generalizes the ℓ1 norm to measures, and the program can be implemented with off-the-shelf software.
  • The resulting approximation is stable: under the minimum-separation condition, its smoothed error scales proportionally to noise level times SRF^2.The analysis treats bounded-norm noise that may otherwise be arbitrary; related higher-dimensional guarantees are also described.

2 Proof of Theorem 1.2

The proof bounds the smoothed recovery error by decomposing the estimation error near and away from the source support, then controlling each part with low-frequency polynomial certificates and local kernel estimates.

  • Error decomposition: The proof partitions the support neighborhoods at the low- and high-resolution scales and analyzes the measure error on these regions.The sets associated with λlo and λhi organize the subsequent local and off-support bounds.
  • Dual-polynomial control: A low-frequency dual polynomial controls the error away from the support and underpins the stability argument.The construction extends the dual certificate used for exact noiseless recovery.
  • Local approximation: Near each source, a Taylor expansion of the super-resolution kernel reduces the local error to constant, linear, and quadratic terms.The quadratic remainder is controlled through bounds on the second derivative of the kernel.
  • Conclusion: The resulting inequalities establish the main theorem’s stability estimate for total-variation minimization under the source-separation condition.The proof combines the regional estimates with minimal total variation of the feasible solution.

3 Numerical implementation

The convex recovery problem can be solved through a semidefinite-programming formulation, with dual solutions helping identify the support of the primal estimate.

  • Semidefinite formulation: The dual of problem (1.9) can be recast as a semidefinite program involving an n × n Hermitian matrix.This formulation leverages a known semidefinite representation and usually permits accurate numerical solution.
  • Implementation: A feasible primal solution can be recovered from a dual solution using the stated primal-dual sign condition.The implementation is described as usable with off-the-shelf software.
  • Primal-dual relation: At nonzero points of the primal solution, the dual polynomial equals the sign of the recovered measure.This relation supports extracting the primal support from the dual solution.
  • Support recovery: The primal support is usually located by finding points where the dual polynomial q(t) has modulus one.After support estimation, the primal coefficients can be obtained from a discrete problem.

4 Discussion

The paper establishes non-asymptotic stability guarantees for convex super-resolution of multiple point sources, while identifying broader noise models and related recovery problems as extensions.

  • Contribution: The framework provides non-asymptotic stability guarantees for tractable super-resolution of multiple point sources in a continuous domain.Its analysis treats bounded-norm noise without requiring a stochastic distribution.
  • Main result: The spectrum can be extrapolated by convex programming with error scaling quadratically in the super-resolution factor.The result is a worst-case guarantee for arbitrary bounded-norm noise.
  • Scope and extensions: The analysis is limited to a bounded L1 perturbation and one recovery scheme, although the techniques may adapt to other settings.The discussion specifically points to stochastic noise, line-spectrum denoising, and random Fourier sampling as extensions.

A Proof of Lemma 2.5

The proof of Lemma 2.5 constructs and locally bounds a low-frequency polynomial that interpolates prescribed phases on the source support.

  • Polynomial construction: The construction reuses the low-frequency polynomial framework from the companion work.The coefficients are selected so the polynomial matches prescribed values at the support points.
  • Local analysis: Writing q(t) − vj as real and imaginary parts enables separate local bounds near each support point.The proof uses prior kernel and polynomial estimates to control both components.
  • Lemma conclusion: The resulting local estimate establishes the additional property required by Lemma 2.5.This conclusion follows after combining the componentwise bounds in the interval of interest.

B Proof of Lemma 2.7

The lemma constructs and bounds low-frequency interpolation polynomials under source separation, establishing local and global controls needed for the proof.

  • The constructed polynomial q1 is constrained to approximate vj(t − tj) around every support point tj.
  • The matrix system uses kernel values G, G(1), and G(2) evaluated at pairwise support differences.
  • Minimum separation makes the interpolation system invertible, so the coefficient vectors α and β are well defined.
  • The proof bounds α and β through a Neumann-series argument applied when ||I − M||∞ < 1.
  • For t near its nearest support point ti, all other support points lie at least Δ/2 away, yielding the required local bound.
  • The local interpolation error satisfies |w(t)| ≤ C f_lo t^2 within the stated neighborhood.

C Proof of Corollary 1.3

The corollary follows by controlling the noise norm probabilistically, making the true signal feasible for the convex program and establishing the required bounds.

  • The noise energy ||ϵ||_2^2 follows a χ^2 distribution with 4f_lo + 2 degrees of freedom.
  • A concentration inequality and Parseval imply that ||Q_lo x − y||_L2 = ||ϵ||_2 ≤ δ with high probability.
  • The resulting feasibility of x_est implies condition (C.1), while Cauchy–Schwarz establishes condition (C.2).

D Extension to multiple dimensions

The multidimensional extension requires low-frequency polynomials with controlled directional derivatives near well-separated support points.

  • Multidimensional proofs construct bounded low-frequency polynomials interpolating sign patterns on well-separated point sets.
  • Bounding directional derivatives in every direction is ensured by controlling the Hessian eigenvalues near the support.

D.1 Proof of an extension of Lemma 2.1 to two dimensions

The two-dimensional extension uses a low-frequency polynomial with controlled Hessian, then transfers the one-dimensional proof through polar decomposition and norm inequalities.

  • The proof relies on a low-frequency polynomial satisfying the required interpolation identities.
  • The support points must be separated by 2.38 λ_lo in infinity norm, although the norm choice can be changed to ℓ2.
  • The polynomial’s Hessian eigenvalues at the support are of order f_lo^2, providing the needed second-derivative control.
  • The proof performs a polar decomposition of P_T h and uses vj = e^−iφ(tj) before applying the low-frequency polynomial bounds.
  • Hölder’s inequality and the triangle inequality combine the local polynomial bounds to establish the extension.
Loading 1211.0290v3…