pespace.inference.likelihood_ti

Taichi backend of the Whittle likelihood kernel.

This module contains the taichi kernel for the Whittle likelihood computation, used by LikelihoodBilbyInterfaceTaichi.

Important: this module must not use from __future__ import annotations (PEP 563). Taichi parses kernel annotations at decoration time and requires the actual annotation objects (e.g., ti.template()); with PEP 563 all annotations become strings and the kernel definition fails with a TaichiSyntaxError.

Functions

compute_whittle_likelihood_ti(channels, ...)

Compute the Whittle likelihood.
pespace.inference.likelihood_ti.compute_whittle_likelihood_ti(channels, observed_data, response_data, psd, df)[source]

Compute the Whittle likelihood.

Parameters:

  • channels (ti.template()) – Specifying the TDI channels to iterate over. Used with ti.static() for compile-time loop unrolling.

  • observed_data (ti.template()) – taichi.field containing the observed frequency-domain data for each channel. Expected to be a field of ti.types.vector(2, float) indexed by frequency bin and channel.

  • response_data (ti.template()) – taichi.field containing detector responses for each channel, having the same structure as observed_data.

  • psd (ti.template()) – taichi.field containing the noise power spectral density for each channel. Expected to be a real-valued field indexed by frequency bin and channel.

  • df (float) – Spacing between frequency bins in Hz.

Returns:

The log-likelihood value computed using the Whittle approximation.

Return type:

float

Notes

  • Uses Array-of-Structures (AoS) layout for StructField, with the channel loop placed inside the frequency loop for optimal memory accessing

  • Atomic operations are used to accumulate the log-likelihood to ensure thread safety in parallel execution