Robust Berrut-Approximated Coded Computing via Discrete Cosine Transforms

What the math gives to ML

The paper provides a concrete rational-coding mechanism for distributing nonlinear computations: data vectors are embedded at Chebyshev nodes into a Berrut rational interpolant, and workers evaluate this interpolant at separate points. This structure can be transferred to distributed neural inference or training to tolerate stragglers while reducing the need to replicate every data shard. The most practical initial experiment is to use Berrut-coded evaluation of a smooth MLP or activation-rich layer, then reconstruct outputs from any sufficiently large subset of workers using the same rational interpolant. The extracted mathematics does not include the paper's explicit DCT error-correction theorem, so Byzantine handling should initially be implemented as a residual-based decoder and evaluated separately from the interpolation approximation error.