Bespoke algorithms

A bespoke algorithm is a new solution method: it changes how a problem is solved rather than what is solved. Where a problem specification hands one monolithic model to Ipopt, an algorithm wraps the staged API (or the engine's admittance-matrix primitives) in a custom loop.

Available

• HELM power flow — the Holomorphic Embedding Load-flow Method: a non-iterative 4-wire power flow that expands each voltage as a power series in a load-scaling parameter and evaluates it by Padé analytic continuation, with physical-residual, Padé-spread, coefficient-tail, and heuristic singularity diagnostics exposed for independent validation.

Question-driven candidates not yet built

The next algorithms should support the research program, rather than form an unconnected catalogue:

• Continuation and branch discovery — pseudo-arclength continuation, singular-point handling, deflation, and systematic multistart to identify which nonlinear solution a sensitivity or decision uses.
• Sensitivity validation — KKT regularity, active-set transition, and finite-difference/continuation comparisons for DiffOpt and, where useful, direct sensitivities through the HELM coefficient recursion.
• Counterexample generation — adversarial utilization/model searches for falsifying operating-envelope candidates, with explicit local-search stopping semantics.
• Relaxations and valid bounds — multiconductor convex relaxations when a valid bound can strengthen a forensics or robust-decision claim.
• Decomposition — spatial, temporal, or hypothesis decomposition only when the integrated inference or verification studies require it for scale.

When one lands, add it under src/algorithms/, give it a page here, and document the validity of its claims. See Contributing for the pattern.