DeepDIVE: Optimizing Input-Constrained Distributions for Composite DNA Storage via Multinomial Channel

TL;DR

The paper tackles capacity-achieving input design for a multinomial channel under a finite input-support constraint, motivated by composite DNA storage. It introduces a Blahut-Arimoto–driven alternating optimization where a Variational Autoencoder selects the $d$ mass-point locations on the input simplex and the BA algorithm tunes their weights, using a differentiable Gumbel-Softmax surrogate for channel sampling. The results show that simplex vertices are not universally optimal, with learned constellations and composite alphabets yielding higher mutual information than baselines, especially for larger $n$ and constrained supports. This approach enables more efficient data encoding in DNA storage by optimizing input distributions under strict support constraints and high-dimensional alphabets.

Abstract

We address the challenge of optimizing the capacity-achieving input distribution for a multinomial channel under the constraint of limited input support size, which is a crucial aspect in the design of DNA storage systems. We propose an algorithm that further elaborates the Multidimensional Dynamic Assignment Blahut-Arimoto (M-DAB) algorithm. Our proposed algorithm integrates variational autoencoder for determining the optimal locations of input distribution, into the alternating optimization of the input distribution locations and weights.

Figures (7)

• Figure 1: End-to-end autoencoder model.
• Figure 2: Loss decomposition during the training process.
• Figure 3: DeepDIVE's geometric and probabilistic shaping results compared to previous methods.
• Figure 4: DeepDIVE's only geometric shaping results compared to previous methods.
• Figure 5: Five-symbols constellations on two-dimensional simplex; corners configuration (a) and middle configuration (b).