Fixation models¶

A fixation model maps a pair of fitnesses \((f_i, f_j)\) to a probability that a new allele \(j\) arising on background \(i\) fixes in the population. gpvolve-v2 ships five built-ins.

Built-ins¶

Name	Required params	Bounded in `[0, 1]`?	Reference
`sswm` (alias `strong_selection_weak_mutation`)	none	yes	Gillespie 1984
`moran`	`population_size`	yes	Sella and Hirsch 2005
`mccandlish` (alias `mcclandish` for v1 compat)	`population_size`	yes	McCandlish 2011
`bloom_dms`	`pi_table`	yes	empirical (Bloom DMS)
`weak_mutation`	none	no	Lynch and Conery 2003

Only the bounded kernels can build a row-stochastic transition matrix. Passing weak_mutation to build_transition_matrix raises NonStochasticError.

Registering a custom kernel¶

from gpvolve import register_fixation_model
import numpy as np

@register_fixation_model(
    name="my_kernel",
    bounded_unit_interval=True,
    required_params=frozenset({"sigma"}),
)
def my_kernel(fi, fj, sigma, **_):
    s = (fj - fi) / fi
    return 0.5 * (1.0 + np.tanh(s / sigma))

The decorator stamps the Protocol attributes (name, bounded_unit_interval, required_params) onto the function and registers it in the module-level registry. Look up the registration with gpvolve.get_fixation_model("my_kernel").

v1 spelling¶

The original gpvolve had a misspelled kernel name (mcclandish instead of mccandlish). v2 ships both spellings as aliases for the same function so v1 scripts keep working.