Regime Change: Bit-Depth versus Measurement-Rate in Compressive Sensing

1 Introduction

• Compressive sensing can reduce the number of acquired values below the signal dimension, benefiting settings such as time-consuming MRI and wideband sampling with a prohibitive Nyquist rate.The paper frames these as cases where each measurement is expensive or the required sampling rate is too high.Source: §1 · p. 2
• The acquisition model includes input noise, scalar quantization to a finite alphabet, and a fixed bit budget for transmitting or storing measurements.The paper models signal noise as a random vector and assumes zero measurement error in its analysis.Source: §1 · p. 2
• High input SNR favors fewer measurements with greater bit-depth (MC), while low input SNR favors more measurements with fewer bits each (QC).The SNR at which the transition occurs depends on the total bit budget.Source: §1 · p. 3
• An upper-bound analysis and simulations distinguish the regimes; simulations report that 1-bit quantization can outperform larger bit-depths at certain practical SNRs.The simulations concern sparse-signal acquisition from uniformly quantized measurements and use reconstruction techniques developed in.Source: §1 · p. 3
• In practical ADCs, decreasing bit-depth can yield an exponential increase in sampling rate, and 1-bit CS measurements reduce quantization to a comparator.Stable recovery with 1-bit measurements has been shown possible.Source: §1 · p. 3

2.1 CS Toolkit

• If the true signal support is known, least squares gives the optimal linear coefficient estimate, and this oracle-assisted reconstruction is the analysis’s best-case benchmark.The RIP provides stable oracle-assisted recovery under white measurement noise for sparse signals.Source: §2.1 · p. 4
• The RIP requires well-conditioned measurement matrices and supports robust sparse recovery; the noise bound assumes white, uncorrelated measurement noise.The measured noise is white under the stated tight-frame condition, which requires M ≤ N.Source: §2.1 · pp. 4–5

2.2 1-bit CS

• In 1-bit CS, measurements retain only signs, so reconstruction seeks a sparse unit-norm signal consistent with those signs because quantization loses scale.The consistency problem is generally non-convex, though approximate algorithms and convex programs have been proposed.Source: §2.2 · p. 5
• The binary ε-stable embedding relates normalized Hamming distance between sign measurements to angular distance between unit-norm sparse signals, and Gaussian measurement matrices satisfy it with high probability.This provides a robustness analogue to the RIP for consistent 1-bit reconstruction.Source: §2.2 · p. 5
• The simulations use BIHT and BIHT-ℓ2, which minimize one-sided ℓ1 and ℓ2 objectives respectively; BIHT favors low-noise settings, while BIHT-ℓ2 favors high-noise settings.A one-sided norm sets positive vector elements to zero before computing the norm.Source: §2.2 · p. 6

• For strictly sparse signals with white noise and an RIP matrix, Theorem 1 bounds oracle-assisted reconstruction MSE under a fixed bit budget using an optimal scalar quantizer with B > 1.The theorem assumes M < N and a specified sparse-signal and measurement-noise model.Source: §3 · p. 6
• The bound captures quantization error, signal-noise effects, and correlation among quantized measurements, with the correlation term expected to be near zero in many CS scenarios.The analysis therefore primarily balances the terms inside the bound’s parentheses when choosing bit-depth.Source: §3 · p. 6
• For compressible signals, tail coefficients also undergo noise folding, and less-compressible signals are expected to induce the same regime as low-input-SNR signals.The bound can be extended by inflating its second term to account for additional correlation between quantized measurements.Source: §3 · pp. 6–7
• At 35dB ISNR, the bound is minimized at approximately 7 bits per measurement; at 10dB ISNR, it is minimized at approximately 2 bits per measurement.The error follows a convex tradeoff with bit-depth, so using more bits per measurement is not always optimal.Source: Figure 1 · p. 7, §3 · p. 8
• The bound is pessimistic because it omits benefits from additional measurements, may not reflect behavior when quantization error dominates measurement noise, and does not apply to 1-bit quantization or M > N.Source: §3 · p. 7
• The plotted evaluation uses only the parenthetical term of the bound, omitting effects of the RIP constant, sparsity, quantization-error correlation, and measurement count on RIP constants.Figure 1 evaluates this term across bit-depth and ISNR; black dots mark each curve’s minimum.Source: §3 · p. 7, Figure 1 · p. 7

4.1 Setup

• The simulations use sparse signals, Gaussian sensing and noise, uniform quantization, and algorithm-specific reconstruction, averaging results over 100 trials per parameter tuple.The reported results did not change when signals were DCT-sparse.Source: §4.1 · p. 8

4.2 Oracle-assisted reconstruction

• Oracle-assisted reconstruction generally follows the analytical bound's trends, with the minimum error shifting to smaller bit-depths as ISNR decreases.The simulations compare bit-depths 2–12 because the bound does not cover lower bit-depths, and use a uniform rather than optimal quantizer.Source: §4.2 · p. 9, Figure 2 · p. 9

4.3 Reconstruction performance as a function of B

• 1-bit CS outperforms multibit methods at low ISNR and is competitive at moderate ISNR; at ISNRs of 20, 10, and 5 dB, 1-bit BIHT-ℓ2 achieves the best performance.The reported pattern supports the QC regime: with lower ISNR, smaller bit-depths and more measurements are beneficial.Source: §4.3 · pp. 10–11, Figure 3 · p. 10

4.4 Reconstruction performance as a function of ISNR

• Transition centers rise from 19 dB at B = N to 23 dB at B = 2N and 38 dB at B = 5N, placing the QC-to-MC transition at higher ISNR for larger bit-budgets.At B = 2N, the recommended bit-depth is approximately 7 at 30 dB (M/N = 0.29) and 1 at 15 dB (M/N = 2).Source: §4.4 · pp. 11–12, Figure 4 · p. 11

5 Discussion

• High-ISNR measurement compression favors fewer measurements, while low-ISNR quantization compression favors fewer bits per measurement.The low-ISNR regime is suited to cheap measurements or expensive large-bit-depth quantizers; the high-ISNR regime is suited to expensive measurements.Source: §5 · p. 12
• The QC tradeoff is between amplified input noise from underdetermined measurements and bits allocated per measurement under a fixed budget.Source: §5 · p. 12
• Unlike the oversampled ADC architectures discussed, CS uses sparse signal models to reduce total acquired bits and does not adapt measurements to the input.The passage notes that almost all previous oversampled ADCs require feedback during quantization for stable representations.Source: §5 · p. 12

A Proof of Theorem 1

• The oracle-assisted estimate has an error upper bound for a K-sparse signal when the noise is zero-mean with covariance Σ and Φ satisfies RIP.The proof begins by extending an existing oracle-assisted bound to account for correlated measurement noise.Source: A Proof of Theorem 1 · p. 12
• For a fixed support, RIP makes the selected measurement submatrix full rank, enabling the oracle estimate and its error analysis.The proof uses eigenvalue bounds for the covariance transformed through the pseudoinverse.Source: A Proof of Theorem 1 · p. 13
• Under the random-support sparse-signal model, Φx is a mixture of Gaussians whose measurement components are zero-mean and uncorrelated.For fixed support, each measurement is Gaussian; averaging across possible supports yields the mixture.Source: A Proof of Theorem 1 · p. 13
• The proof bounds measurement-error variance by combining quantization error and input-noise contributions, then bounds the covariance matrix's largest eigenvalue.The scalar quantization bound is applied to the sum of signal and input-noise measurements.Source: A Proof of Theorem 1 · pp. 12, 14