Source-linked AI summary

A typical reconstruction limit of compressed sensing based on Lp-norm minimization

Y. Kabashima, T. Wadayama, T. Tanaka

arXiv:0907.0914v2cs.ITcond-mat.dis-nnmath.ST

TL;DR

The paper studies when a sparse N-dimensional signal can be reconstructed from P linear constraints by minimizing an Lp-norm. Using the replica method and replica-symmetric stability analysis, it derives typical-case critical relations αc(ρ) for p = 0, 1, and 2, including an L1 threshold expressed by αc(ρ) = 2(1−ρ)H(bχ−1/2)+ρ.

  • Problem

    The paper asks under what conditions an N-dimensional sparse signal can be correctly reconstructed from its compressed vector when the compression matrix is known.

  • Method

    The paper uses the replica method under the replica-symmetric ansatz and assesses the stability of successful-reconstruction solutions in the limit N, P →∞ with α = P/N finite.

  • Results

    The analysis yields critical relations αc(ρ) for p = 0, 1, and 2; for p = 1, αc(ρ) = 2(1−ρ)H(bχ−1/2)+ρ, while the p = 0 solution achieves the best possible performance and p = 2 has no compressed-sensing capability.

  • Takeaways & Limitations

    The findings extend to relatively wide classes of compression matrices and signals, since the critical values do not depend on details of the non-zero signal distribution under finite mean and variance.

  • Takeaways & Limitations

    Replica-symmetric analysis is invalid when the AT instability condition holds, requiring replica-symmetry-breaking analysis; for p = 0, that further analysis is beyond the Letter's scope.

Abstract

from arXiv · show

We consider the problem of reconstructing an $N$-dimensional continuous vector $\bx$ from $P$ constraints which are generated by its linear transformation under the assumption that the number of non-zero elements of $\bx$ is typically limited to $ρN$ ($0\le ρ\le 1$). Problems of this type can be solved by minimizing a cost function with respect to the $L_p$-norm $||\bx||_p=\lim_{ε\to +0}\sum_{i=1}^N |x_i|^{p+ε}$, subject to the constraints under an appropriate condition. For several $p$, we assess a typical case limit $α_c(ρ)$, which represents a critical relation between $α=P/N$ and $ρ$ for successfully reconstructing the original vector by minimization for typical situations in the limit $N,P \to \infty$ with keeping $α$ finite, utilizing the replica method. For $p=1$, $α_c(ρ)$ is considerably smaller than its worst case counterpart, which has been rigorously derived by existing literature of information theory.

1. Introduction

Compressed sensing reconstructs high-dimensional signals from lower-dimensional measurements by exploiting sparsity. This paper focuses on typical-case reconstruction thresholds and contrasts them with worst-case guarantees.

  • Compressed sensing reconstructs high-dimensional signals from lower-dimensional data using prior knowledge that the signal is sparse.
  • The problem asks when an N-dimensional sparse signal can be correctly recovered from P<N linear measurements using a known compression matrix.
  • Lp-norm minimization is studied as a sparsity-exploiting reconstruction scheme subject to the measurement constraints.
  • Worst-case L1 guarantees provide sufficient conditions for arbitrary signals, but typical-case thresholds can differ considerably and may better reflect practical situations.
  • The paper assesses typical reconstruction conditions in the limit N,P→∞ with finite α=P/N using statistical-mechanical methods, agreeing with prior numerical indications.

2. Problem setting

The problem setting models a sparse random signal, Gaussian measurements, and reconstruction through general Lp-cost minimization. Performance is evaluated at finite compression rate in the large-system limit.

  • Each component of x0 is independently drawn from a distribution with probability 1−ρ at zero and Gaussian nonzero values, where ρ is the signal density.
  • Measurements are generated as y=Fx0, with y and F known while the original signal x0 remains hidden.
  • The measurement matrix entries are i.i.d. Gaussian with mean zero and variance N^-1.
  • The analysis considers general Lp-reconstruction and examines p=0, 1, and 2 as N,P→∞ with finite compression rate α=P/N.
  • Unlike the noiseless setting studied here, finite-variance measurement noise prevents exact reconstruction of x0 in the related noisy problem.

3. Analysis

The analysis converts constrained minimization into a statistical-mechanical posterior and evaluates its typical behavior with the replica method. The resulting order parameters identify reconstruction success and its stability.

  • A posterior distribution with inverse temperature β represents the constrained minimization, approaching the solution set as β→∞.
  • Averaging over quenched randomness in F and x0 under the replica-symmetric ansatz yields a typical free-energy expression.
  • The extremum variables encode signal overlaps and norms, allowing the typical mean-square error and minimized Lp cost to be evaluated.
  • The single-site optimization profile x∗p(h; bQ) can support an approximation algorithm for solving the reconstruction problem.
  • Replica-symmetry breaking must be considered when the de Almeida–Thouless instability condition holds, because the RS treatment is then invalid.

4. Results

The analysis derives typical reconstruction limits for p = 0, 1, and 2, identifies RS-stability boundaries, and validates the p = 1 prediction numerically at ρ = 0.5.

  • Typical transitions: The numerical RS extremization found a stable success solution at sufficiently large α and a transition to failure as α decreased.The success solution is characterized by Q = m = ρ, whereas failure has Q ≠ m ≠ ρ.
  • p = 0: For p = 0, the success solution is stable when α > ρ, giving αc(ρ) = ρ, but the RS estimate is physically invalid because of AT instability.Further RSB analysis is required for an accurate p = 0 assessment.
  • p = 1: For p = 1, the typical limit is αc(ρ) = 2(1−ρ)H(bχ−1/2)+ρ, and the RS success solution is locally stable under the stated condition.The p = 1 criticality coincides with the relevant AT-stability condition.
  • p = 2: For p = 2, the success solution is stable only when α ≥ 1, so L2 minimization cannot reconstruct compressed expressions.At α ≥ 1, the constraints themselves suffice for perfect reconstruction.
  • Typical versus worst-case performance: The typical L1 limit differs greatly from the worst-case critical compression rate, while L1 balances relatively high reconstruction capability with practical computational feasibility.The cited comparison notes O(N3) cost for interior-point methods and NP-hardness for L0 reconstruction.
  • Numerical validation: For L1 reconstruction at ρ = 0.5, extrapolation gave αc(0.5) ≃ 0.83165, close to the theoretical value 0.83129.The estimate used trials for N = 10, 12, . . ., 30 and quadratic extrapolation in 1/N.

5. Summary and discussion

The paper assesses typical compressed-sensing performance for L0-, L1-, and L2-norm reconstruction using the replica method under the replica-symmetric ansatz. It finds strong practical promise for L1 reconstruction, while identifying instability for L0 and no compressed-sensing capability for L2.

  • The study evaluates typical reconstruction performance for p = 0, 1, and 2 using the replica method under the replica-symmetric ansatz.
  • L0 reconstruction achieves the best possible performance in the RS analysis but is unstable against replica-symmetry-breaking perturbations.
  • L2 reconstruction has no capability for compressed sensing, whereas L1 reconstruction has considerably high reconstruction ability.
  • L1 reconstruction is practically attractive because it can be solved by linear programming with feasible computational cost.
  • For i.i.d. zero-mean fixed-variance compression matrices, the results remain identical, and α_c(ρ) does not depend on the non-zero signal distribution when its mean and variance are finite.
  • Further work is needed on replica-symmetry-breaking analysis for L0 reconstruction and lower-cost mean-field algorithms.

Appendix A. Derivation of eq. (7)

Appendix A derives the replica-symmetric free-energy expression by averaging replicated Gaussian constraints and applying saddle-point methods. The derivation uses order parameters and zero-temperature scaling to obtain the main equation.

  • The replica calculation averages over Gaussian compression variables whose covariances are expressed through replica overlaps Q_ab.
  • Under the replica-symmetric ansatz, diagonal, off-diagonal, signal-overlap, and signal-density order parameters organize the replicated system.
  • The Gaussian constraint variables are represented using independent standard Gaussian variables before being substituted into the replicated partition function.
  • A saddle-point evaluation provides the volume associated with the replica-symmetric order parameters.
  • Analytic continuation from integer replica number to real n, together with zero-temperature scaling of the order parameters, yields eq. (7).

Appendix B. Treatment of rotationally-invariant matrix ensembles

Appendix B extends the analysis to rotationally invariant compression-matrix ensembles. Under the stated spectral assumptions, the zero-temperature free energy agrees with the main result, so the critical relation is unchanged.

  • The rotationally invariant ensemble uses a diagonal matrix with an asymptotically fixed eigenvalue distribution and a uniformly sampled orthogonal matrix.
  • The appendix applies the technique associated with the rotationally invariant matrix representation within the replica-symmetric framework.
  • For large x, the extremization yields Λ ≃ (1−α)/x, providing the asymptotic form needed in the free-energy comparison.
  • In the β →∞ limit, the rotationally invariant ensemble's free energy differs from the main expression only by a β-independent constant, so the result is identical.
Loading 0907.0914v2…