[Paper Review] New upper bounds for trace reconstruction

Zachary Chase|arXiv (Cornell University)|Sep 7, 2020

Digital Image Processing Techniques25 references10 citations

TL;DR

This paper improves the upper bound for trace reconstruction from exp(O(n^{1/3})) to exp(Õ(n^{1/5})) traces, using a novel analysis of sparse polynomials derived from subsequence appearances in random traces. The method leverages complex analysis and concentration bounds on weighted subsequence counts to reconstruct any n-bit string with high probability using exponentially fewer traces than previous methods.

ABSTRACT

We show that any $n$-bit string can be recovered with high probability from $\exp(\widetilde{O}(n^{1/5}))$ independent random subsequences.

Motivation & Objective

• To improve the upper bound on the number of i.i.d. traces needed to reconstruct an unknown n-bit string with high probability.
• To extend prior results from q ≤ 1/2 to all q ∈ (0,1) in the trace reconstruction problem.
• To develop tighter concentration bounds for weighted subsequence counts using complex analysis on the unit circle.
• To achieve a sub-cubic root dependence on n, improving upon the previous exp(O(n^{1/3})) bound.

Proposed method

• Analyzes the polynomial ∑ₖ[1_{x_{k+i}=w_i} − 1_{y_{k+i}=w_i}]z^k for a carefully chosen substring w to exploit sparsity.
• Uses a generalized trace identity to relate expected values of trace statistics to weighted counts of substring appearances in the original string.
• Applies complex analysis techniques to lower-bound the magnitude of the polynomial near z=1 on a small arc of the unit circle.
• Employs a compact family of Lipschitz functions and integral comparison arguments to control approximation error in trace sampling.
• Leverages a probabilistic method with concentration inequalities to ensure high-probability reconstruction from T = exp(Cn^{1/5} log⁵n) traces.
• Extends the method to all deletion probabilities q ∈ (0,1) using a short argument from Shyam Narayanan.

Experimental results

Research questions

• RQ1 Can the trace reconstruction problem be solved with significantly fewer than exp(O(n^{1/3})) traces?
• RQ2 Can the upper bound be extended to all deletion probabilities q ∈ (0,1), not just q ≤ 1/2?
• RQ3 Can sparse polynomial analysis on the unit circle yield better concentration bounds than general trace statistics?
• RQ4 What is the optimal trade-off between trace complexity and reconstruction accuracy in the trace reconstruction model?

Key findings

• The paper establishes that any n-bit string can be reconstructed with probability at least 1−δ using T = exp(Cn^{1/5} log⁵n) i.i.d. traces for some C>0.
• The new upper bound of exp(Õ(n^{1/5})) improves upon the previous best-known bound of exp(O(n^{1/3})).
• The method achieves better concentration by focusing on sparse polynomials derived from substring appearances rather than bit-level differences.
• The analysis is extended to all q ∈ (0,1) using a short argument from Shyam Narayanan, removing prior restrictions on deletion probability.
• The proof relies on complex analysis and integral comparison techniques to bound the magnitude of the polynomial on a small arc near z=1.
• The result demonstrates that trace reconstruction is possible with sub-exponential but still super-polynomially many traces, narrowing the gap to the known Ω̃(n^{3/2}) lower bound.