Daubechies and DeVore (Annals of Mathematics, 2003) constructed the first provably stable family of one-bit sigma-delta (ΣΔ) quantizers of arbitrary order for bandlimited functions and proved super-polynomial error decay C′λ−γlog⁡λC'\lambda^{-\gamma\log\lambda} in the oversampling ratio λ\lambda, closing their paper with the question of what the information-theoretic constraints on such coarse quantization of redundant information are: how fast can the reconstruction error actually decay, and can high accuracy be reconciled with the stability and robustness required in practice. The qualitative part was resolved essentially simultaneously with the source: Güntürk (CPAM, 2003) built one-bit ΣΔ schemes with feedback filters achieving error O(2−0.07λ)O(2^{-0.07\lambda}), and Deift, Krahmer and Güntürk (CPAM, 2011) improved the exponent to r≈0.102r\approx 0.102; however, information-entropy arguments force r<1r<1, it is not known whether this ceiling is attainable by a scheme, the best constructions only apply to small-amplitude inputs and are not robust, and Krahmer and Ward (ACHA, 2012) proved the first coarse-quantization-specific lower bounds tying the best possible exponent to the signal amplitude. The surviving open problem is to determine the optimal error-decay rate of one-bit ΣΔ quantization of bandlimited signals with convolutional reconstruction — including whether the entropy ceiling is attainable by a scheme and how it depends on the amplitude class — and to decide whether exponential accuracy is achievable by a scheme that is simultaneously stable (uniformly bounded internal state variables) and robust (tolerant to imprecise quantizer thresholds and gain factors), together with the source's still-open structural sub-questions: the limitation of schemes without feedback filters, the fast growth of the state-variable bounds in the arbitrary-order family, and the lack of constructions guaranteeing that internal states automatically fall back to a small dynamic range during low-amplitude input stretches.

Background

An analog signal is modeled as a bandlimited function ff, i.e. a function whose Fourier transform vanishes outside a fixed finite frequency band. By the Shannon–Nyquist sampling theorem such a function is determined by its samples taken at a rate above twice its bandwidth, and it can be reconstructed by the sampling formula f(t)=∑nf(n/λ) g(t−n/λ)f(t)=\sum_n f(n/\lambda)\,g(t-n/\lambda), where λ>1\lambda>1 is the oversampling ratio (sampling rate divided by the Nyquist rate) and gg is a suitable reconstruction kernel whose Fourier transform approximates the characteristic function of the frequency support (an ideal low-pass filter). Analog-to-digital (A/D) conversion replaces the real-valued samples f(n/λ)f(n/\lambda) by values qnq_n from a finite alphabet; the quantized signal f~(t)=1λ∑nqn g(t−n/λ)\tilde f(t)=\frac{1}{\lambda}\sum_n q_n\,g(t-n/\lambda) is then used as an approximation of ff. With pulse code modulation (PCM) one samples near the Nyquist rate and replaces each sample by a truncated binary expansion of mm bits, giving error O(2−m)O(2^{-m}); spending λ\lambda bits per Nyquist interval on binary expansions of samples taken slightly above the Nyquist rate therefore yields exponential accuracy O(2−cλ)O(2^{-c\lambda}) in λ\lambda. Coarse quantization instead keeps a fixed, small alphabet — in the extreme case the one-bit alphabet {−1,+1}\{-1,+1\} — and gains accuracy by increasing λ\lambda, which is cheaper in hardware because circuit complexity and power grow rapidly with the number of quantization levels.

Sigma-delta (ΣΔ, also called delta-sigma) modulation is the dominant coarse-quantization architecture. It is a feedback algorithm: the output bits qn∈{−1,+1}q_n\in\{-1,+1\} are chosen sequentially, typically by the greedy rule that minimizes the magnitude of an internal state variable unu_n that accumulates past quantization errors; a scheme of order kk is built from kk-th order differences of the error sequence. The reconstruction error is then controlled by summation by parts against a smooth kernel gg: if the relevant kk-th order difference sequence is uniformly bounded, the error decays like λ−k\lambda^{-k}. Two properties beyond accuracy are essential in practice. Stability means the internal state variables remain uniformly bounded over the whole signal class, so the hardware does not saturate. Robustness, in the precise sense of Daubechies and DeVore, means the accuracy guarantees survive implementation imperfections: the sign function used to pick the bits may toggle within a tolerance τ\tau of the true threshold, and the multiplicative factors in the recursion may be perturbed by a known relative tolerance, without degrading the error bound. Practical schemes often also insert causal feedback filters hh into the state recursion, un=(h∗u)n+yn−qnu_n=(h*u)_n+y_n-q_n with yn=f(n/λ)y_n=f(n/\lambda), whose mathematical role was left unexplained by the source paper.

Daubechies and DeVore (2003) constructed, for every order JJ, a stable family of one-bit JJ-th order ΣΔ quantizers (their recursion (36)) that tolerates imperfect quantizers and gain perturbations; the price is that the uniform bound on the internal states grows extremely fast with the order, like ν(J−1)2\nu^{(J-1)^2} for a constant ν>1\nu>1, and the constants in the error bound λ−k\lambda^{-k} for the kk-th order scheme grow like kγkνkk\gamma^k\nu^k. Optimizing the order against λ\lambda yields their headline estimate ∥f−f~∥L∞≤C′λ−γlog⁡λ\|f-\tilde f\|_{L^\infty}\le C'\lambda^{-\gamma\log\lambda} for f∈C1f\in C^1 with ∥f∥L∞≤a<1\|f\|_{L^\infty}\le a<1: better than any inverse polynomial in λ\lambda, but far from the O(2−cλ)O(2^{-c\lambda}) of binary quantization. Their concluding section asks how much of this discrepancy is an artifact of the proof, a limitation of their specific stable family, or a fundamental constraint of one-bit ΣΔ quantization (with or without filters), and calls for an understanding of the information-theoretic constraints on ΣΔ schemes and other practical quantizers for redundant information.

Two types of limits were known around that time. First, a simple counting (entropy) argument shows that any one-bit scheme spending λ\lambda bits per Nyquist interval cannot beat error decay O(2−λ)O(2^{-\lambda}), i.e. any exponent rr in an error bound O(2−rλ)O(2^{-r\lambda}) must satisfy r<1r<1. Second, Calderbank and Daubechies (2002) proved that 'democratic' coarse representations — in which all bits enter the reconstruction with equal weight, as in convolutional reconstruction — can never achieve the optimal accuracy of optimal nondemocratic schemes. Subsequent work sharpened both sides: Güntürk (2003) achieved genuine exponential accuracy O(2−0.07λ)O(2^{-0.07\lambda}) with filter-based one-bit ΣΔ schemes (not robust in the Daubechies–DeVore sense); Deift, Krahmer and Güntürk (2011) optimized the class of 'minimally supported' feedback filters under the greedy rule — the optimal filters are related to zero sets of Chebyshev polynomials of the second kind — improving the exponent to r≈0.102r\approx 0.102; and Krahmer and Ward (2012) used large-deviation theory to prove the first lower bound specific to coarse quantization, showing that for a KK-bit alphabet and inputs bounded by μ<1\mu<1 the error decay is capped, up to constants, at 2−αm2^{-\alpha m} with α=1−K−1(1−h((1+μ)/2))\alpha=1-K^{-1}(1-h((1+\mu)/2)) and m=Kλm=K\lambda the number of bits spent per Nyquist interval, where hh is the binary entropy function — so for one-bit quantization the best possible exponent tends to 00 as the input amplitude approaches its maximum, while it may approach 11 for very small amplitudes. In parallel, Daubechies, DeVore, Güntürk and Vaishampayan (2006) showed that classical ΣΔ modulation self-corrects quantizer threshold bias and constructed non-classical (β-encoder-type) encoders with exponential accuracy that retain this self-correction, and Daubechies and Yılmaz (2006) made β-encoders robust also to uncertainty in the expansion base — so robustness together with exponential precision is achievable, but outside the classical one-bit ΣΔ framework. For schemes without feedback filters, Krahmer, Saab and Ward (2012) showed that greedy ΣΔ quantization of tailored finite frame expansions achieves root-exponential accuracy e−cλe^{-c\sqrt{\lambda}}; Ward (2010) constructed 'quiet' second-order schemes whose output and state fall back to zero when the input vanishes; and stability analysis of second-order schemes beyond the conservative sufficient condition ∥h∥ℓ1+∥f∥∞≤2\|h\|_{\ell^1}+\|f\|_\infty\le 2 remains an active open topic (Joy, Krahmer and Lupoli, 2026).

Problem Statement

For one-bit (±1) sigma-delta quantization of bandlimited functions with convolutional reconstruction at oversampling ratio λ\lambda, determine the fundamental information-theoretic constraints on the achievable reconstruction accuracy — the question posed at the end of Daubechies and DeVore (2003) — as it survives the partial resolutions of 2003–2012. The three parts below are independent sub-questions. Concretely:

(i) Optimal decay rate. Determine the optimal error-decay exponent r∗r^{*}, as a function of the input amplitude bound μ<1\mu<1, for one-bit ΣΔ (and more generally KK-bit coarse) quantization of bandlimited signals reconstructed by convolution with a kernel: prove matching achievability and impossibility results, and in particular decide whether the information-entropy ceiling r<1r<1 is tight, in the sense of being attained in the small-amplitude limit — that is, whether the optimal exponent r∗(μ)r^{*}(\mu) tends to 11 as μ→0\mu\to0, rather than remaining bounded away from 11. These are separate targets: the small-amplitude tightness question and the determination of r∗(μ)r^{*}(\mu) for all μ<1\mu<1, with matching achievability and impossibility results. The current gap is between the achievable r≈0.102r\approx 0.102 (greedy rule, minimally supported feedback filters, inputs of amplitude μ≤0.05\mu\le 0.05) and the known constraints r<1r<1 (entropy) and the amplitude-dependent cap of Krahmer and Ward; the cases of arbitrary (not necessarily minimally supported) feedback filters and of quantization rules beyond the greedy rule are also open.

(ii) Compatibility with stability and robustness. Decide whether exponential accuracy O(2−rλ)O(2^{-r\lambda}) with r>0r>0 can be achieved by a one-bit ΣΔ scheme that is simultaneously stable (internal state variables uniformly bounded over the signal class) and robust in the Daubechies–DeVore sense (accuracy preserved under tolerance-τ\tau threshold imprecision of the sign quantizers and bounded multiplicative perturbations of the recursion coefficients). All known exponentially accurate one-bit ΣΔ schemes sacrifice robustness; robust exponential-precision encoders exist only outside the classical ΣΔ framework.

(iii) Structural sub-questions raised by the same source and still open: whether one-bit ΣΔ without feedback filters can achieve better than super-polynomial (e.g. root-exponential) accuracy for bandlimited signals — that is, whether the discrepancy between the source's C′λ−γlog⁡λC'\lambda^{-\gamma\log\lambda} and exponential decay is an artifact of the proof, a limitation of the stable family, or a fundamental constraint of filterless schemes; whether the fast, roughly ν(J−1)2\nu^{(J-1)^2}, growth with order JJ of the uniform state-variable bounds in the arbitrary-order stable family can be substantially reduced; and whether there exist constructions (beyond Ward's special second-order 'quiet' schemes) guaranteeing that the internal state variables automatically fall back, after a transition time, to a bounded range much smaller than their full dynamic range when the input signal runs through a low-amplitude stretch, as desired for audio signals.

References

  1. Ingrid Daubechies and Ron DeVore, "Approximating a bandlimited function using very coarsely quantized data: A family of stable sigma-delta modulators of arbitrary order", Annals of Mathematics 158 (2003) 679–710, DOI 10.4007/annals.2003.158.679, https://doi.org/10.4007/annals.2003.158.679
  2. C. Sinan Güntürk, "One-bit sigma-delta quantization with exponential accuracy", Communications on Pure and Applied Mathematics 56 (2003), no. 11, 1608–1630, DOI 10.1002/cpa.3044, https://doi.org/10.1002/cpa.3044
  3. C. Sinan Güntürk, "Approximating a bandlimited function using very coarsely quantized data: Improved error estimates in sigma-delta modulation", Journal of the American Mathematical Society 17 (2004) 229–242, DOI 10.1090/s0894-0347-03-00436-3, https://doi.org/10.1090/s0894-0347-03-00436-3
  4. A. R. Calderbank and Ingrid Daubechies, "The pros and cons of democracy", IEEE Transactions on Information Theory 48 (2002) 1721–1725, DOI 10.1109/TIT.2002.1003852, https://doi.org/10.1109/TIT.2002.1003852
  5. Ingrid Daubechies, Ronald A. DeVore, C. Sinan Güntürk, and Vinay A. Vaishampayan, "A/D conversion with imperfect quantizers", IEEE Transactions on Information Theory 52 (2006) 874–885, DOI 10.1109/TIT.2005.864430, https://doi.org/10.1109/TIT.2005.864430
  6. Ingrid Daubechies and Özgür Yılmaz, "Robust and practical analog-to-digital conversion with exponential precision", IEEE Transactions on Information Theory 52 (2006) 3533–3545, DOI 10.1109/TIT.2006.878220, https://doi.org/10.1109/TIT.2006.878220
  7. Percy Deift, Felix Krahmer, and C. Sinan Güntürk, "An optimal family of exponentially accurate one-bit Sigma-Delta quantization schemes", Communications on Pure and Applied Mathematics 64 (2011) 883–919, DOI 10.1002/cpa.20367, https://doi.org/10.1002/cpa.20367 (preprint arXiv:1001.5079, https://arxiv.org/abs/1001.5079)
  8. Felix Krahmer and Rachel Ward, "Lower bounds for the error decay incurred by coarse quantization schemes", Applied and Computational Harmonic Analysis 32 (2012) 131–138, DOI 10.1016/j.acha.2011.06.003, https://doi.org/10.1016/j.acha.2011.06.003 (preprint arXiv:1004.3517, https://arxiv.org/abs/1004.3517)
  9. Felix Krahmer, Rayan Saab, and Rachel Ward, "Root-exponential accuracy for coarse quantization of finite frame expansions", IEEE Transactions on Information Theory 58 (2012) 1069–1079, DOI 10.1109/TIT.2011.2168942, https://doi.org/10.1109/TIT.2011.2168942
  10. Rachel Ward, "Quiet sigma delta quantization, and global convergence for a class of asymmetric piecewise-affine maps", Nonlinearity 23 (2010) 2165–2182, DOI 10.1088/0951-7715/23/9/007, https://doi.org/10.1088/0951-7715/23/9/007
  11. C. Sinan Güntürk, "Mathematics of analog-to-digital conversion", Communications on Pure and Applied Mathematics 65 (2012) 1671–1696, DOI 10.1002/cpa.21425, https://doi.org/10.1002/cpa.21425
  12. Özgür Yılmaz, "Stability analysis for several second-order Sigma-Delta methods of coarse quantization of bandlimited functions", Constructive Approximation 18 (2002) 599–623, DOI 10.1007/s00365-002-0505-2, https://doi.org/10.1007/s00365-002-0505-2
  13. Rohan Joy, Felix Krahmer, and Alessandro Lupoli, "On trajectory-based stability analysis for 1-bit Sigma-Delta quantization and its application to the second-order case", arXiv:2605.17127 (2026), https://arxiv.org/abs/2605.17127
  14. Theodore Faust, Mark Iwen, Rayan Saab, and Rongrong Wang, "On the ℓ∞-norms of the singular vectors of arbitrary powers of a difference matrix with applications to sigma-delta quantization", Linear Algebra and its Applications 626 (2021) 79–151, DOI 10.1016/j.laa.2021.05.015, https://doi.org/10.1016/j.laa.2021.05.015