[BUG]: plc.eigenvector_centrality fails to converge on simple graph

[eriknw]

Version

23.12 dev

Which installation method(s) does this occur on?

Source

Describe the bug.

plc.eigenvector_centrality fails to converge on a simple graph. NetworkX converges in a few iterations.

The adjacency matrix of this graph is:

[[0 1 0]
 [1 0 1]
 [0 1 0]]

The expected result is [0.5, 0.5**0.5, 0.5].

Minimum reproducible example

import cupy as cp
import numpy as np
import pylibcugraph as plc

src_indices = cp.array([0, 1, 1, 2], dtype=np.int32)
dst_indices = cp.array([1, 0, 2, 1], dtype=np.int32)

G = plc.SGGraph(
    resource_handle=plc.ResourceHandle(),
    graph_properties=plc.GraphProperties(
        is_multigraph=False,
        is_symmetric=True,
    ),
    src_or_offset_array=src_indices,
    dst_or_index_array=dst_indices,
    weight_array=None,  # Also fails with array of all 1s
    store_transposed=True,
    renumber=False,
    do_expensive_check=False,
)
# cpp/src/centrality/eigenvector_centrality_impl.cuh line=151: Eigenvector Centrality failed to converge.
node_ids, values = plc.eigenvector_centrality(
    resource_handle=plc.ResourceHandle(),
    graph=G,
    epsilon=0.1,  # This is really high!
    max_iterations=1000,  # This is super high! We should converge in a few steps
    do_expensive_check=False,
)

### Relevant log output

_No response_

### Environment details

_No response_

### Other/Misc.

_No response_

### Code of Conduct

- [X] I agree to follow cuGraph's Code of Conduct
- [X] I have searched the [open bugs](https://github.com/rapidsai/cugraph/issues?q=is%3Aopen+is%3Aissue+label%3Abug) and have found no duplicates for this bug report

[ChuckHastings]

I was able to reproduce this in the C API. I will investigate further.

[ChuckHastings]

I have a bug fix for this. Will push a PR later today, perhaps with a few other bug fixes.