xinterp

This package enables index to value mapping both in a forward and backward way. Inverse retrieval of indices from given values can be done with different matching rules (None, nearest, forward-fill, backward-fill). Results are exact even when using big integer values (e.g., nanosecond datetime64), and dtypes that are integer, floating, datetime64 or timedelta64 are all supported for fp.

method="nearest" clamps without limit: a value beyond the first or last point resolves to that boundary's index, however far outside the range it falls. Where rounding to an integer is needed, ties round to even (banker's rounding), matching the behaviour of round() and numpy.round.

The two families

An axis can be described in two ways, and the package mirrors the same operations across both.

Points -- the knots are listed, and each piece implies its own slope. Use forward_points / inverse_points when the axis is irregular and every knot matters.

Step -- the knots are generated from a starting value at a fixed exact rate num / den. Use forward_step / inverse_step when the axis is regular: a long run of samples costs two integers rather than a knot each, and the rate stays exact because it is never collapsed to a float.

Guarantees

Values live on the tick grid of their dtype and every result is the exactly rounded one, so the following hold at every sample rather than on average. They hold identically for the points and the step family.

Round trip, index to value and back. inverse(forward(x)) == x whenever the mapping is injective -- that is, whenever values advance by at least one tick per index step (|num| >= den for the step family). Below that rate several indices share a value, and inverse returns one of them.

Round trip, value to index and back. forward(inverse(f)) == f whenever inverse returns. Under method=None a value the curve never attains raises KeyError rather than resolving to a neighbour; the other methods resolve to a neighbouring index by design, so the law does not apply to them.

Simplification is lossless at zero. simplify(x, 0) reproduces every sample exactly. The walk is conservative there: it drops a knot only when it can prove the drop costs nothing, so at zero it may well keep everything.

Simplification is bounded elsewhere. No sample moves by more than tol. simplify_step takes tol in ticks and applies it directly. simplify_points measures its budget against the unrounded curve, so pass en / ed = tol + 1/2 to get the same bound on the reconstruction.

Installation

pip install xinterp

Usage

Points

import numpy as np
from xinterp import forward_points, inverse_points

xp = np.array([0, 10, 20])
fp = np.array([0, 1000, 2000], dtype="datetime64[s]")

x = np.array([0, 5, 10, 15, 20])
result = forward_points(x, xp, fp)
expected = np.array([0, 500, 1000, 1500, 2000], dtype="datetime64[s]")
assert np.array_equal(result, expected)

x = np.array([1, 499, 1001, 1503, 1997], dtype="datetime64[s]")
result = inverse_points(x, xp, fp, method="nearest")
expected = np.array([0, 5, 10, 15, 20])
assert np.array_equal(result, expected)

Step

The rate is a ratio, so a 2.5 s sampling interval is num, den = 5, 2 and stays exact. Values landing on a half tick round to even:

import numpy as np
from xinterp import forward_step, inverse_step

tie_indices = np.array([0, 4])
tie_values = np.array([0, 10], dtype="datetime64[s]")
num, den = 5, 2

x = np.array([0, 1, 2, 3, 4])
result = forward_step(x, tie_indices, tie_values, num, den)
expected = np.array([0, 2, 5, 8, 10], dtype="datetime64[s]")
assert np.array_equal(result, expected)

f = np.array([0, 2, 5, 8, 10], dtype="datetime64[s]")
result = inverse_step(f, tie_indices, tie_values, num, den)
assert np.array_equal(result, x)

infer_step recovers that ratio from an axis, and reports the worst per-segment deviation from it, so you can tell whether a single rate is honest before adopting one. deviation_step gives the residual segment by segment:

import numpy as np
from xinterp import deviation_step, infer_step

x = np.array([0, 10, 20, 30])
f = np.array([0, 25, 50, 75], dtype="datetime64[s]")

assert infer_step(x, f) == (5, 2, 0)  # 2.5 s per index, fitting every segment
assert np.array_equal(deviation_step(x, f, 5, 2), [0, 0, 0])

Simplification

simplify_points returns a mask of the knots worth keeping. Tolerance is the exact ratio en / ed, measured against the unrounded curve -- to bound the reconstructed value by tol ticks, pass tol + 1/2:

import numpy as np
from xinterp import forward_points, simplify_points

xp = np.array([0, 10, 20, 30, 40])
fp = np.array([0, 1001, 1999, 3002, 4000], dtype="datetime64[s]")

keep = simplify_points(xp, fp, 5, 2)  # 5/2 == tol 2 ticks + 1/2
assert np.array_equal(keep, [True, False, False, False, True])

recon = forward_points(xp, xp[keep], fp[keep])
assert np.abs((recon - fp).astype("i8")).max() <= 2  # the promised bound

simplify_step is the twin for rate-described segments: it fuses consecutive runs whose declared step agrees, and re-anchors each survivor. It returns the keep mask alongside the re-anchored tie values:

import numpy as np
from xinterp import simplify_step

tie_values = np.array([0, 1000, 2000])
tie_lengths = np.array([10, 10, 10])

keep, fused = simplify_step(tie_values, tie_lengths, [100], [1], 0)
assert np.array_equal(keep, [True, False, False])  # one run, not three
assert np.array_equal(fused, [0])

Migrating from 0.1

forward and inverse are now forward_points and inverse_points. The old names still work as aliases but emit a DeprecationWarning, and are scheduled for removal in 0.5:

from xinterp import forward_points as forward  # was: from xinterp import forward

Nothing else in the 0.1 surface changed. See CHANGELOG.md for the full list.

Development

The project is developed with uv and maturin:

uv sync        # build the extension and install the dev dependencies
uv run pytest  # Python suite
cargo test     # Rust suite

See CONTRIBUTING.md for the full toolchain.