Error Exponent Bounds for Optimal Short-Read Clustering

Abstract: Motivated by the operation of decoders for DNA storage, we consider the problem of unsupervised clustering of noisy short sequences, each generated from one of multiple possible unknown source sequences after passing through a memoryless channel. Focusing on the statistically optimal clustering rule, we derive upper and lower bounds on the probability of incorrect clustering as a function of the sequence length, the number of reads, and the channel statistics.

• The paper establishes sharp upper and lower bounds on the error exponent for unsupervised clustering of noisy short reads using MAP and large deviations techniques.
• It employs both finite-blocklength and asymptotic analyses, leveraging Bhattacharyya coefficients and concentration inequalities to bound clustering error probabilities.
• The findings provide practical design guidelines for DNA storage systems by quantifying how sequence length, channel noise, and source statistics affect clustering reliability.

Error Exponent Bounds for Optimal Short-Read Clustering

Problem Setting and Motivation

The paper investigates a fundamental statistical inference task encountered in DNA storage systems and related settings: the unsupervised clustering of $n$ noisy, short reads, each generated independently from one of $m$ unknown source sequences and passed through a memoryless channel. The aim is to optimally partition the $n$ output reads into $m$ groups so that each group contains reads originating from the same source sequence, while the original sequences and the correspondence between reads and sources are both unobserved.

This problem arises naturally in DNA storage, modeled by a noisy shuffling sampling channel, where a message is encoded as a multiset of $m$ DNA strings. Each molecule is sampled and sequenced with noise, generating a collection of output reads. Many practical decoders initiate with a clustering step to group reads by origin before decoding message content. Although practical clustering procedures are generally suboptimal, the focus here is exclusive on optimal clustering—specifically on determining sharp lower and upper bounds for its error exponent as a function of the relevant system parameters (notably, sequence length $\ell$, number of reads $n$, number of sources $m$, and channel statistics).

Optimal Clustering and Formalization

The model formalizes the clustering task as multiple hypothesis testing over patterns induced by sampling with replacement. The core decoding problem is to infer the sampling index vector (up to permutation), or equivalently, the induced pattern $\Psi(S_1^n)$, from the noisy observations $Y_1^n$. The optimal rule is Maximum A Posteriori (MAP) over the space of patterns, with likelihoods evaluated by marginalizing over the unknown (i.i.d.) source sequence realizations and the reading channel.

For a finite alphabet source $m$0, memoryless source distribution $m$1, and memoryless channel $m$2 (with $m$3 and $m$4), the error probability of any clusterer $m$5 is defined as the probability that $m$6 does not coincide with the true pattern, minimized over all decoders.

The analysis particularly focuses on the logarithmic regime $m$7 (with $m$8 fixed and $m$9 constant), as motivated by realistic design points in DNA storage, where sequences are much shorter than the number of reads. The principal asymptotic quantity of interest is the error exponent:

$n$0

Main Theoretical Results

Upper and Lower Bounds on the Error Exponent

The authors establish both finite-blocklength and asymptotic single-letter bounds on the error probability and the corresponding error exponent.

1. Finite-Length Error Probability Bound:

An explicit multi-letter upper bound on the MAP error probability is derived (see Eq. (16)), involving combinatorial enumeration over alternative clusterings and expectations over source realizations. The bound exploits concentration results on the empirical distribution of source assignments and is critically based on Bhattacharyya coefficients between possible source string realizations.

2. Asymptotic Lower Bound (Achievability):

By evaluating the typical set structure and employing large deviations via the method of types, a single-letter lower bound on $n$1 is established:

$n$2

where $n$3 and $n$4 are optimized large deviations exponents for atypically low Bhattacharyya distance and empirical probability fluctuations, respectively. For uniform distributions, $n$5 leads to a sharper characterization, and the dominating error event is a single erroneous assignment, controlled by pairwise Bhattacharyya distances.

3. Asymptotic Upper Bound (Converse):

Via a reduction to an assignment problem and leveraging a random-coding ensemble argument (Gallager’s exponent), a single-letter upper bound is proven (Theorem 2):

$n$6

where $n$7 is the expected Bhattacharyya coefficient between two independent $n$8 samples. The proof exploits the independence structure given the sources and applies a clipped union bound to tightly link the clustering and assignment error exponents.

Discussion of Tightness and Gap

The bounds coincide in their functional dependence on the system parameters but do not always match, reflecting a gap between the achievable and converse error exponents. Both are exponential in $n$9 and polynomial in $m$0 due to the logarithmic scaling, but the exact error exponent value depends on precise large deviations behavior of Bhattacharyya distances and the distribution $m$1.

Notably, for the uniform source case, the lower bound can be tight under proper parameterization, determined by an explicit equation involving $m$2. For non-uniform $m$3, especially for highly skewed distributions, the lower bound can be loose due to rare but high-probability source strings.

Numerical Illustration and Parameter Dependence

The theoretical bounds are illustrated for the canonical binary symmetric channel (BSC) and $m$4-ary symmetric channels with varying alphabet sizes and error rates. The exponent increases with alphabet size, reflecting greater distinguishability among source sequences. The bounds also elucidate sharp thresholds in the scaling parameter $m$5—below which clustering is fundamentally infeasible, regardless of algorithm.

Numerical analysis further reveals:

• The gap between lower and upper bounds narrows as the source alphabet grows and for moderate channel noise.
• In practical DNA (i.e., $m$6) regimes, the bounds are informative for the selection of system parameters to ensure a desired reliability level.
• The lower bound degrades for heavily biased sources, exposing the need for further refinement in these regimes.

Methodological Contributions

Key technical tools include:

• Type-based large deviations analysis for the joint distributions of source tuples.
• Concentration inequalities for the empirical frequencies of source assignment.
• Extension of Bhattacharyya-style pairwise error bounds to marginalized (over sources) likelihoods in the clustering context.
• Clipped union bounds and assignment reduction to connect the clustering hypothesis testing problem to classical random-coding arguments.
• Explicit single-letter characterizations of both the exponents and the threshold phenomena in the logarithmic regime.

Implications and Future Directions

From a theoretical standpoint, the results pin down the fundamental limits of clustering for unsupervised separation of noisy short reads, severely constraining or enabling practical algorithmic design in systems such as DNA storage and inference from mixed sources in sensor networks. The bounds apply directly to the regime where clustering is the critical computational bottleneck, exposing where the minimal error is dominated by inherent information-theoretic obfuscation rather than algorithmic suboptimality.

Practically, the work provides design guidelines on the required sequence length and number of reads for a given error target, dependent on the channel and sequence statistics. It quantifies the exponential improvement in reliability with increased sequence distinguishability, motivating the use of larger alphabets and careful codebook design.

Open Problems and Extensions:

• Closing the gap between upper and lower bounds, particularly for structured or unbalanced sources.
• Extending analysis to the regime where $m$7 scales with $m$8 (e.g., $m$9) and developing results for approximate clustering (allowing a prescribed number of mismatches).
• Considering more general channel models, such as synchronization errors (insertions, deletions), which are relevant in practical biomolecular or physical settings.
• Analyzing suboptimal, polynomial-time clustering algorithms to quantify their (likely modest) suboptimality as compared to the optimal rules analyzed here.

Conclusion

This paper establishes fundamental single-letter bounds on the capacity of statistically optimal short-read clustering under memoryless noise, sharply characterizing the exponential error decay as a function of sequence length, channel, and statistics. These results serve as a theoretical baseline for both evaluating practical algorithms and designing DNA storage and other systems subject to unsupervised mixture separation constraints. Continued work is indicated to extend these limits to larger and more complex regimes and to refine their tightness in the presence of source or channel structure.

Reference: "Error Exponent Bounds for Optimal Short-Read Clustering" (2606.21218)