[DOC] Add arXiv link to KFAC for modern NN architectures paper (#57)

singd/optim/optimizer.py

@@ -156,8 +156,8 @@ def __init__(
                 or `Linear` modules that process matrix- or higher-dimensional
                 features.
                 Possible values are `'expand'` and `'reduce'`.
-                See [Eschenhagen et al., 2023](TODO Insert arXiv link) for an
-                explanation of the two approximations.
+                See [Eschenhagen et al., 2023](https://arxiv.org/abs/2311.00636)
+                for an explanation of the two approximations.
             warn_unsupported: Only relevant if `params` is unspecified. Whether to
                 warn if `model` contains parameters of layers that are not supported.
                 These parameters will not be trained by the optimizer. Default: `True`

test/optim/test_kfac.py

@@ -100,8 +100,8 @@ def test_kfac(
 ):
     """Test KFAC using (deep) linear models with weight sharing and the MSE loss.
 
-    See [Eschenhagen et al., 2023](TODO Insert arXiv link) for more details on
-    the conditions for KFAC being exact.
+    See [Eschenhagen et al., 2023](https://arxiv.org/abs/2311.00636) for more
+    details on the conditions for KFAC-expand and KFAC-reduce being exact.
 
     Args:
         model: Tuple of model name and function takes `bias` as input and