[DiadkoShmek]

logdet() can raise AttributeError: 'Tensor' object has no attribute 'linear_ops' when an added diagonal is a Kronecker product of constant-diagonal factors and the matrix reaches the existing fast-path threshold. A 1024-dimensional covariance triggers this with default settings, also blocking Gaussian log-probability evaluation.

Recognize this constant diagonal locally in _logdet and reuse the existing spectral shift, logdet(K + dI) = sum(log(eigenvalues(K) + d)). This removes the broken factor reconstruction and its incorrect scaling, while preserving the existing small-matrix fallback and solve/root routes.

import torch
from linear_operator.operators import (
    ConstantDiagLinearOperator,
    DiagLinearOperator,
    KroneckerProductAddedDiagLinearOperator,
    KroneckerProductDiagLinearOperator,
    KroneckerProductLinearOperator,
)

a = torch.linspace(1, 2, 32, dtype=torch.double)
b = torch.linspace(3, 5, 32, dtype=torch.double)
k = KroneckerProductLinearOperator(DiagLinearOperator(a), DiagLinearOperator(b))
d = KroneckerProductDiagLinearOperator(
    ConstantDiagLinearOperator(torch.tensor([2.0], dtype=torch.double), diag_shape=32),
    ConstantDiagLinearOperator(torch.tensor([3.0], dtype=torch.double), diag_shape=32),
)
matrix = KroneckerProductAddedDiagLinearOperator(k, d)
result = matrix.logdet()  # Default settings, 1024 x 1024 matrix.
torch.testing.assert_close(result, torch.log(torch.kron(a, b) + 6).sum())
print(result)

BASE raises at matrix.logdet(); the change returns approximately 2536.626652357273, matching the diagonal reference without overriding any settings.

CPU validation:

• The complete affected native module passes 247 tests. The new regression raises in six BASE subtests and passes on the change, covering float32/float64, positive/zero/two-negative constant factors, and selected noise gradients/Hessians.
• A separate NumPy/analytic/finite-difference reference passes 29 cases/165 checks; independent review passes 53 scenarios/240 checks. These include batch metadata, different diagonal factorizations, nonconstant-diagonal controls and the default-size Gaussian value/first-gradient path.
• Exact workflow Flake8/pre-commit, all-files hooks, and wheel/sdist builds pass. Both HTML documentation builds succeed with zero Sphinx logger warnings, while retaining the same 25 existing Python RuntimeWarnings from the unchanged type formatter in docs/source/conf.py. No warning suppression or warning-free claim.

This is a constant-noise log-determinant repair, not a general Kronecker higher-derivative, solve/root, GPU or performance claim.

I directed this work and take responsibility for the contribution and its follow-up. AI assistance was used in producing the change.