Samples and objective sense

A sample pairs x with shape (n, d) and y with shape (n,). Bounds have shape (d,). The NumPy backend stores floating-point arrays as float64.

LandscapeSample validates its inputs on construction: X must be two-dimensional and y one-dimensional with matching length, all values must be finite, every lower bound must be strictly below its upper bound, and all observations must lie inside the box. The sample is immutable and its arrays are read-only, so a sample can be reused across many compute calls. The arrays are copied from the caller's inputs, so later changes to the original handles never affect the sample. Because NumPy arrays that own their storage can have the read-only flag re-enabled (directly or through an alias), each compute call first invokes sample.validate_unchanged(), which recomputes an integrity digest and raises RuntimeError if any exposed array was mutated in place — so a stale fingerprint can never be used as a provenance or cache key. The Torch TensorLandscapeSample offers the same guarantee via tensor version counters.

Maximization problems

Declare the objective sense instead of negating y by hand. Features that claim invariance under sense reversal return identical values either way.

from orivex import ObjectiveSense

maximizing = LandscapeSample(x, -y, lower=lower, upper=upper, sense=ObjectiveSense.MAXIMIZE)
minimizing = LandscapeSample(x, y, lower=lower, upper=upper)

compute(maximizing, "nbc.*")  # same values as compute(minimizing, "nbc.*")

sample.minimization_y exposes the objective in minimization convention if you need it directly.

The examples above use x, y, lower, and upper from the quick start. See the sample API for properties and constructor signatures.