On Unlimited Sampling

I. INTRODUCTION

• Compact Fourier-domain support allows a function to be uniquely represented by equidistant samples.The spacing between samples depends on the function’s bandwidth.Source: §I · p. 3
• Sampling theory extensions have addressed time-domain attributes such as sparsity, smoothness, and uniform versus non-uniform grids.The passage lists extensions including shift-invariant and union-of-subspaces sampling, sparse signals, phase-space representations, spectrum-blind sampling, and operator sampling.Source: §I · p. 3
• This paper instead considers a sampling hypothesis based on the amplitude dimension.Source: §I · p. 3

A. Shannon’s Sampling Theory and a Practical Bottleneck

• An ADC’s dynamic range is the maximum recordable voltage range [−λ, λ], and signals crossing the threshold are clipped.Source: A. Shannon’s Sampling Theory and a… · p. 3
• Clipping a bandlimited signal creates discontinuities and high-frequency distortion, and can produce aliasing artifacts.Clipping also manifests as nonlinear artifacts in audio-visual and physiological or biomedical data.Source: A. Shannon’s Sampling Theory and a… · pp. 3–4
• A conventional ADC saturates at λ, whereas a self-reset ADC folds inputs into [−λ, λ], avoiding clipping but introducing discontinuities.Source: Figure 1 · p. 3
• Fig. 2 depicts the modulo signal and its uniform samples alongside the continuous residual and its sampled version.Source: Figure 2 · p. 4

B. A Solution via Modular Arithmetic

• Self-reset ADCs reset at a threshold to capture subsequent changes beyond saturation; implementations include folding-ADC and self-reset-ADC designs.The self-reset architecture is used in natural-image, life-science, and bio-imaging applications.Source: B. A Solution via Modular Arithmetic · p. 4
• The shared ADC model is a memoryless nonlinear mapping that folds amplitudes into the range [−λ, λ] and can be interpreted as a centered modulo operation.Source: B. A Solution via Modular Arithmetic · p. 4
• The paper’s sampling-and-reconstruction conditions and constructive recovery algorithm address the largely unexplored identifiability and recovery of modulo samples.Earlier approaches required additional information, such as how often amplitudes had been folded.Source: B. A Solution via Modular Arithmetic · pp. 4–5
• The paper’s first contribution is a formal step toward sufficient conditions for sampling and reconstructing bandlimited functions from self-reset ADC modulo samples.Source: B. A Solution via Modular Arithmetic · p. 5
• The work targets unlimited sampling, where signal amplitude can far exceed the usual ADC range, by trading sampling rate for dynamic range.Source: B. A Solution via Modular Arithmetic · p. 5

• When f is not bandlimited, filtering with φ yields g = f ∗ φ, which the paper assumes has normalized bandwidth π and belongs to Bπ.Source: §II · p. 5
• The folded output Mλ(g(t)) is sampled uniformly at interval T to produce modulo samples.Source: §II · p. 5
• The self-reset ADC turns a smooth function into a discontinuous one through simple functions arising from the folding operation.Source: §II · p. 5
• For zero-mean g ∈ Bπ and fixed λ, modular decomposition reduces recovery from z to finding εg.Source: §II · p. 5

• The first question is which conditions allow recovery of g from its modulo samples yk.Source: §III · p. 6
• The second question asks whether a constructive sampling-theorem-style algorithm can map yk to g.Source: §III · p. 6
• The authors state that both recovery questions are answered affirmatively.Source: §III · p. 6

A. A Sufficiency Condition for Unlimited Sampling

• Applying higher-order finite differences to the modulo samples enables recovery when the corresponding differences of the original samples are small enough that folding has no effect.Finite differences commute with the modulo operation, and the resulting differences can be computed from the folded samples.Source: A. A Sufficiency Condition for… · p. 6
• For T < 1/(πe), choosing N logarithmically in ||g||∞ can make the relevant finite differences smaller than λ; the paper uses T ≤ 1/(2πe).A known upper bound β_g on ||g||∞ is assumed for choosing N.Source: A. A Sufficiency Condition for… · pp. 6–7
• Repeated summation with rounding to the nearest multiple of 2λ is stable, while a constant-sequence ambiguity during inversion is resolved for n > 1 by summing a second time.For n = 1, adding multiples of 2λ cannot be distinguished from the modulo samples.Source: A. A Sufficiency Condition for… · p. 7
• With a known β_g ≥ ||g||∞ and a suitable N satisfying (Tπe)^N||g||∞ < λ, Algorithm 1 recovers g from modulo samples up to additive multiples of 2λ.The theorem gives this as a sufficient condition for recovery.Source: A. A Sufficiency Condition for… · p. 8
• Because T is set by circuit design while N matters only for reconstruction, a fixed architecture can capture signals of arbitrary amplitude when a norm bound is available for recovery.The circuit architecture itself imposes no limit on β_g.Source: A. A Sufficiency Condition for… · pp. 7–8
• Increasing the oversampling rate with fixed N may suit some computational settings, but lowering T requires changing the circuit architecture.The paper therefore focuses on fixed T and variable N.Source: A. A Sufficiency Condition for… · p. 9

B. Numerical Demonstration

• For λ = 1/20, T ≈ 1/(2πe), and N = 3, the authors reconstructed samples and ε_g with errors on the order of Matlab’s numerical precision.Source: B. Numerical Demonstration · p. 9

IV. CONCLUSIONS

• The paper establishes a sufficient condition for recovering bandlimited signals from folded samples and provides a recovery method.This is presented as a first step towards a sampling theory for self-reset ADCs.Source: §IV · p. 9
• The sufficient recovery condition typically does not seem empirically optimal, motivating investigation of sharper conditions.Source: §IV · p. 9
• Sampling in shift-invariant spaces and sampling parametric signals in imaging are identified as unexplored extensions of modulo sampling theory.Source: §IV · p. 9

A. Proof of Proposition 1

• The proof obtains the desired result by writing h as its integer and fractional parts.Source: A. Proof of Proposition 1 · p. 9

B. Proof of Lemma 1

• Taylor’s theorem expresses g(t) around an anchor point τ as a degree-(N−1) polynomial plus a remainder involving g^(N)(ξ_l).The intermediate value ξ_l lies between lT and τ.Source: B. Proof of Lemma 1 · p. 9
• The Nth finite difference annihilates polynomials of degree N−1, so only the Taylor remainder contributes.Source: B. Proof of Lemma 1 · pp. 9–10
• The proof bounds the finite difference using the remainder vector and the inequality ∥Δv∥∞≤∥v∥∞, then applies Stirling’s approximation.Source: B. Proof of Lemma 1 · p. 10

C. Proof of Proposition 2

• Writing a ≜ Δ^N a, the proof decomposes it as M_λ(a)+ε_a, where ε_a is a simple function, yielding Δ^N M_λ(a)=a−ε_a.Source: C. Proof of Proposition 2 · p. 10
• Because ε_a takes values in multiples of 2λ, its Nth finite difference lies in the kernel of M_λ, establishing the result in (10).Source: C. Proof of Proposition 2 · p. 10