SC—001 Sparse by design

Intelligence lives in the nonzeros.

Sparse matrices reveal a simple truth: structure matters more than density. We’re exploring AI built on selective computation—activating useful work and leaving the rest at zero.

Read the sparse premise
Illustrative matrix
Shape
A ∈ ℝ128×128
Nonzeros
655 / 16,384
Density
4.00%
Format
CSR
Explore

The next leap may come from computing only what matters.

ρ(A) = nnz(A) m × n

In a sparse matrix, a small number of nonzero entries carry the structure. We’re investigating how that principle can shape AI: structured sparsity, selective activation, and less data movement—with the aim of reducing memory, bandwidth, and energy spent on unnecessary work.

Compare notes

Combinatorial matrix theory

The zeros need a quorum.

Frobenius–König says a square sparsity pattern is structurally singular exactly when its zeros can organize an empty r × s rectangle with r + s = n + 1.

sprank(A) < n A[I,J] = 0 |I| + |J| = n + 1

That empty rectangle blocks every perfect row–column matching. In less formal terms: the zeros unionized.

Structural statement only. Particular values can still conspire through cancellation.

Read the theorem
Sparsity pattern A 6 × 6

3 rows + 4 columns = 7 = n + 1

Row–column graph Maximum matching
r1 r2 r3 r4 r5 r6 c1 c2 c3 c4 c5 c6
Compatibility is structural. Chemistry is numerical.