sublinear-solver

Sublinear solver

Pick the entry point

• Library code: npm i @ruvnet/sublinear-core. It has 0 dependencies and provides ESM, CJS and types.
• Agent tools: the MCP server npx -y sublinear-time-solver serve. The tools to use are solve, pageRank, estimateEntry and analyzeMatrix.
• Rust: cargo add sublinear.

Recipes

import { solve, graphFromEdges, pageRank, personalizedPageRank, forwardPush, backwardPush, estimateEntry, topK } from '@ruvnet/sublinear-core';

// A x = b. A is COO {rows, cols, rowIndices, colIndices, values}, dense rows, or a CsrMatrix.
const { x, residual, converged, method } = solve(A, b, { tolerance: 1e-10 });

// Graph ranking with string ids; edges are [source, target, weight?].
const g = graphFromEdges(edges);
topK(pageRank(g).scores, 10, g);
topK(personalizedPageRank(g, seedId, { damping: 0.85 }).scores, 10, g);

// Large graph, local answer: the cost depends on epsilon, not on n.
topK(forwardPush(g, seedId, { epsilon: 1e-6 }).scores, 10, g);
backwardPush(g, targetId);              // ppr_s(target) for every source s

// One entry of x without solving for all of x (requires strict row diagonal dominance).
estimateEntry(A, b, i, { walks: 20000 }); // { estimate, standardError }

Rules

• damping is the probability of following an edge. A restart probability α means damping = 1 - α.
• Check converged and residual before relying on x. The residual is recomputed from the returned vector.
• Jacobi, Neumann and Gauss-Seidel need diagonal dominance. Conjugate gradient needs a symmetric positive definite matrix. method: 'auto' picks between them.
• A full solution vector costs time linear in nnz per iteration. Only forwardPush, backwardPush and estimateEntry have a cost independent of n.
• The MCP solveTrueSublinear tool is approximate. Check its verified_relative_residual, or use solve when you need accuracy.
• The consciousness, psycho-symbolic, emergence and temporal-lead MCP tools are experimental simulations, not solvers.