MatDot and PolyDot Codes for Straggler-Resilient Coded Matrix Multiplication
Algorithms AIMech:MatDotAndPolyDotCodesForStragglerResilientCodedMatrixMultiplication DRAFT
Aliases: MatDot Codes, PolyDot Codes, Systematic MatDot Codes, MatDot and PolyDot Coded Matrix Multiplication
A family of algebraic coded matrix-multiplication schemes proposed by Dutta, Fahim, Haddadpour, Jeong, Cadambe, and Grover (2017/2020) that reduces the straggler recovery threshold for distributed multiplication AB of N x N matrices under a 1/m per-worker storage constraint from m^2 (in Polynomial Codes) down to 2m - 1 in MatDot Codes—by partitioning A vertically into m column-blocks [A_0, ..., A_{m-1}] and B horizontally into m row-blocks [B_0^T, ..., B_{m-1}^T]^T, encoding p_A(x) = sum_{i=0}^{m-1} A_i x^i and p_B(x) = sum_{j=0}^{m-1} B_j x^{m-1-j}, and recovering AB = sum_{i=0}^{m-1} A_i B_i as the coefficient of x^{m-1} in p_A(x) p_B(x)—and unifies MatDot and Polynomial Codes via PolyDot(m, s, t) Codes (with st = m) to trade off worker-to-fusion communication cost O(N^2 / t^2) against recovery threshold k = t^2(2s - 1).
Identity Notes & Concept Scope
Distinct from Polynomial Codes (AIMech:PolynomialCodesForStragglerResilientHighDimensionalCodedMatrixMultiplication, which split A horizontally and B vertically into outer-product blocks requiring recovery threshold m^2 with O(N^2 / m^2) per-worker output communication). Dutta et al. (Allerton 2017 / IEEE Trans. Inf. Theory 2020 / arXiv:1801.10292, Constructions 1-5 & Theorems 1-6, pp. 4-17) introduce four coupled constructions: (1) MatDot Codes (Construction 1, Theorem 1), which encode column-blocks of A as p_A(x) = sum_{i=0}^{m-1} A_i x^i and row-blocks of B as p_B(x) = sum_{j=0}^{m-1} B_j x^{m-1-j} so that AB = sum_{i=0}^{m-1} A_i B_i is the coefficient of x^{m-1} in the degree-(2m-2) product polynomial p_C(x) = p_A(x) p_B(x), achieving recovery threshold k = 2m - 1; (2) Systematic MatDot Codes (Construction 2, Lemma 1, Theorem 2), which encode p_A(x) = sum_{i=1}^m A_{i-1} L_i(x) and p_B(x) = sum_{i=1}^m B_{i-1} L_i(x) using Lagrange basis polynomials L_i(x) so the first m workers output uncoded A_{r-1} B_{r-1} while any 2m - 1 workers suffice for interpolation; (3) PolyDot(m, s, t) Codes (Construction 3, Lemma 2, Theorem 3), which partition A into a t x s block grid and B into an s x t block grid (st = m) to achieve recovery threshold t^2(2s - 1) (improved to t^2 s + s - 1 under generalized substitution x = y^t, z = y^{st}, Table I); and (4) n-Matrix Multiplication Codes (Constructions 4-5, Theorems 4-6).
Claims & Grounded Evidence
4 Claims
Mechanism
#matdot-column-row-block-polynomial-encoding-and-systematic-lagrange-construction
MatDot splits A vertically into m column-blocks A_i in F^{N x (N/m)} and B horizontally into m row-blocks B_j in F^{(N/m) x N}, encodes p_A(x) = sum_{i=0}^{m-1} A_i x^i and p_B(x) = sum_{j=0}^{m-1} B_j x^{m-1-j}, and recovers AB = sum_{i=0}^{m-1} A_i B_i as the coefficient of x^{m-1} in the degree-(2m - 2) polynomial p_C(x) = p_A(x) p_B(x) from any 2m - 1 workers (Construction 1, Theorem 1); Systematic MatDot (Construction 2, Lemma 1, Theorem 2) uses Lagrange basis polynomials L_i(x) so the first m workers directly output A_{r-1} B_{r-1} while retaining the 2m - 1 worst-case recovery threshold.
Scope
#polydot-communication-recovery-threshold-interpolation-and-n-matrix-theorems
For any factorization st = m dividing N, PolyDot(m, s, t) splits A into a t x s grid and B into an s x t grid and uses the bijection f(alpha, beta, gamma) = alpha + t beta + t(2s - 1) gamma (Lemma 2, Equation 13) to achieve recovery threshold k = t^2(2s - 1) with per-worker communication O(N^2 / t^2) (Theorem 3), interpolating between MatDot (s = m, t = 1, k = 2m - 1) and Polynomial Codes (s = 1, t = m, k = m^2); Theorems 4-6 extend this construction to one-round coded multiplication of n >= 3 matrices with recovery threshold 2m^{n/2} - 1 for even n and (m + 1)m^{floor(n/2)} - 1 for odd n.
Limitation
#matdot-worker-to-fusion-communication-expansion-and-tensor-product-decoding-limitation
Reducing the recovery threshold from m^2 to 2m - 1 in MatDot increases the per-worker output dimension from (N/m) x (N/m) to a full N x N matrix (increasing total worker-to-fusion communication from O(N^2) to O(m N^2)), and when coded computing is applied to operations whose output size matches their O(N^4) computational complexity (such as Kronecker tensor products A otimes B), fusion-node decoding overhead Theta(N^4 log^2 m) becomes a bottleneck (Section VII-A).
Application
#matdot-polydot-fault-tolerant-distributed-linear-algebra-and-deep-network-application
Applied to straggler- and fault-tolerant distributed matrix-matrix multiplication, chained multi-matrix products D^(1) D^(2) ... D^(n), and coded deep neural network training on unreliable compute clusters.
⚠️ Curated Limitations & Negative Results
- MatDot requires each responding worker to transmit a full N x N matrix product to the fusion node, resulting in O(m N^2) total worker-to-fusion communication compared to O(N^2) for Polynomial Codes.
- Chained one-round n-matrix multiplication codes require a recovery threshold that grows exponentially with the chain length n as Theta(m^{ceil(n/2)}).
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:LinearRegression
Linear Regression
|
related_match | AIO:LinearRegression and distributed linear/deep learning rely on straggler-resilient coded matrix-matrix multiplication provided by MatDot and PolyDot Codes. |
Evidence Sources & Primary Literature
3 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
r541-primary-pdf
|
Sanghamitra Dutta, Mohammad Fahim, Farzin Haddadpour, Haewon Jeong, Viveck R. Cadambe, Pulkit Grover
|
primary paper
full_text
|
ID: https://dblp.org/rec/journals/tit/DuttaFHJCG20.xml
sha256:16643ac3cced…
|
r541-arxiv-abs
|
Sanghamitra Dutta, Mohammad Fahim, Farzin Haddadpour, Haewon Jeong, Viveck R. Cadambe, Pulkit Grover
|
primary paper
abstract_only
|
ID: https://arxiv.org/abs/1801.10292
sha256:0ed6854109ed…
|
r541-metadata-cr
|
Sanghamitra Dutta, Mohammad Fahim, Farzin Haddadpour, Haewon Jeong, Viveck R. Cadambe, Pulkit Grover
|
primary paper
abstract_only
|
ID: https://dblp.org/rec/journals/tit/DuttaFHJCG20.xml
sha256:a2bc60ebc732…
|
Ontology & Grounding Topology
Open in interactive network graph →Curation Provenance
Curator: Antigravity
Method: ai_assisted
Model: Gemini
Created: 2026-10-19T10:24:00Z
Updated: 2026-10-19T10:24:00Z
Explore across X-Mech
Related computational & biological Mechs:
Embedding Map
Network Graph
ProteinTraitsMech (ESM-2 pLM)
DUFMech (Domain pLM)
TraitMech
CultureMech
All 14 X-Mech Repositories