Numerically stable coded matrix computations via circulant and rotation matrix embeddings

Several recent works have used coding-theoretic ideas for mitigating the effect of stragglers in distributed matrix computations (matrix-vector and matrix-matrix multiplication) over the reals. In particular, a polynomial code based approach distributes matrix-matrix multiplication among n worker nodes by means of polynomial evaluations. This allows for an ``optimal\u27\u27 recovery threshold whereby the intended result can be decoded as long as at least (n−s) worker nodes complete their tasks; s is the number of stragglers that the scheme can handle. However, a major issue with these approaches is the high condition number of the corresponding Vandermonde-structured recovery matrices. This presents serious numerical precision issues when decoding the desired result. It is well known that the condition number of real Vandermonde matrices grows exponentially in n. In contrast, the condition numbers of Vandermonde matrices with parameters on the unit circle are much better behaved. In this work we leverage the properties of circulant permutation matrices and rotation matrices to obtain coded computation schemes with significantly lower worst case condition numbers; these matrices have eigenvalues that lie on the unit circle. Our scheme is such that the associated recovery matrices have a condition number corresponding to Vandermonde matrices with parameters given by the eigenvalues of the corresponding circulant permutation and rotation matrices. We demonstrate an upper bound on the worst case condition number of these matrices which grows as ≈O(ns+6). In essence, we leverage the well-behaved conditioning of complex Vandermonde matrices with parameters on the unit circle, while still working with computation over the reals. Experimental results demonstrate that our proposed method has condition numbers that are several orders of magnitude better than prior work