`tf.pow` for complex64: `pow(0+0j, 0+0j)` returns `NaN+NaNj` in eager mode

Issue type

Bug

Have you reproduced the bug with TensorFlow Nightly?

Yes

Source

source

TensorFlow version

TensorFlow 2.22.0-dev20260503

Custom code

Yes

OS platform and distribution

No response

Mobile device

No response

Python version

No response

Bazel version

No response

GCC/compiler version

No response

CUDA/cuDNN version

No response

GPU model and memory

No response

Current behavior?

Summary

tf.pow(0+0j, 0+0j) returns NaN+NaNj in eager mode, while NumPy returns (1+0j). The eager implementation computes complex power via exp(b * log(a)), which hits log(0) = -inf and produces NaN.

Under XLA (jit_compile=True), tf.pow(0+0j, 0+0j) correctly returns 1+0j, matching NumPy.

Scope

All complex zero variants are affected:

Base	Exp	Eager	XLA	NumPy
0+0j	0+0j	NaN+NaNj	1+0j	1+0j
-0+0j	0+0j	NaN+NaNj	1+0j	1+0j
0-0j	0+0j	NaN+NaNj	1+0j	1+0j

Root cause

TensorFlow's eager complex pow kernel computes pow(a, b) = exp(b * log(a)). When a = 0+0j, log(0+0j) = -inf + ..., and subsequent multiplication produces NaN. The special case pow(0, 0) = 1 (from IEEE 754 for real numbers and C99 for complex) is not handled.

Expected behavior

Eager should return (1+0j) for pow(0+0j, 0+0j), matching NumPy and XLA.

Environment

• TensorFlow 2.22.0-dev20260503
• CPU only

Standalone code to reproduce the issue

## Reproduction


import tensorflow as tf
import numpy as np

z = tf.constant([0+0j], dtype=tf.complex64)

print(tf.pow(z, z).numpy())
# [(nan+nanj)]  — eager: wrong

print(tf.function(lambda a, b: tf.pow(a, b), jit_compile=True)(z, z).numpy())
# [(1+0j)]  — XLA: correct

print(np.power(0+0j, 0+0j))
# (1+0j)  — NumPy reference


Reproduces on TF 2.22.0-dev20260503 (nightly), CPU.

Relevant log output