Skip to content

python-prescribed-transport.md

Latest commit

 

History

History
56 lines (49 loc) · 2.95 KB

python-prescribed-transport.md

File metadata and controls

56 lines (49 loc) · 2.95 KB
Raw

Prescribed physical transport from Python

ContinuationPrescription binds exact physical threshold declarations to a KinematicTransport. It lives in symbolica.community.hep.integration alongside the existing transport and boundary-cache classes. Pass an ordered sequence of (Symbolica expression, "+i0" or "-i0") pairs, a homotopy-domain description, and an optional unprescribed_side. Exact expressions never pass through Python floats. The native PhysicalContinuation owner checks allowed kinematic variables, polynomial form, and original denominator domains when the declaration is bound. Epsilon-dependent, root-dependent and nonpolynomial declarations are rejected. Declaration order and the domain are preserved in the cache identity.

Both the ordinary constructor and KinematicTransport.canonical accept continuation=.... For example, given native expressions s, epsilon, Y and one (the exact constant 1):

from symbolica.community.hep.integration import (
    ContinuationPrescription, KinematicTransport,
)

continuation = ContinuationPrescription(
    [(s, "+i0")], domain="upper s-plane approach", unprescribed_side="+i0"
)
transport = KinematicTransport(
    epsilon, {s: [[epsilon / s]]}, [Y], one,
    branch_domain="logarithm normalized at s=1", continuation=continuation,
)
# `boundaries` contains values and evidence inserted for this transport identity.
result = transport.evaluate(
    boundaries, {s: -one}, 0, 2, admit_prescribed_path=True
)

The existing native planner generates and checks the detours. Polynomial sides do not establish global monodromy: admit_prescribed_path=True is the caller's assertion that this declared homotopy applies to the integral. It is separate from admit_straight_path, which cannot authorize a prescribed detour. Exact compatible cache hits need no new route admission. Opposite prescriptions or different homotopy domains have different identities and cannot share numerical evidence merely because coordinates coincide. Registered roots retain explicit source and destination sheets through the same native continuation owner.

For singular endpoints, a bound continuation also selects the native prescribed matching route in evaluate_endpoint. admit_matching_path admits that route; admit_endpoint separately admits the final local chart. A cached terminal still requires endpoint admission. Terminal limits remain separate from regular ODE starting values, and binary save/load and cache merging preserve both the terminal evidence and its regular anchor.

Computation releases the GIL and accepts the existing ComputationControl. Typed native failures preserve unsupported inputs, missing evidence, inconsistent sheets, exhausted limits and cancellation. The Python layer adds no contour, Frobenius, graph or tensor algorithm, and no cache format. Local and accumulated accuracy retain the native solver's documented scope; a declared contour is not an accuracy certificate.