Use expm1 for log-normal transforms

Summary: Replacing `exp(x) - 1` with the numerically more stable `expm1`.

Differential Revision: D62789111

botorch/models/transforms/utils.py

@@ -60,7 +60,7 @@ def norm_to_lognorm(mu: Tensor, Cov: Tensor) -> tuple[Tensor, Tensor]:
     diag = torch.diagonal(Cov, dim1=-1, dim2=-2)
     b = mu + 0.5 * diag
     mu_ln = torch.exp(b)
-    Cov_ln = (torch.exp(Cov) - 1) * torch.exp(b.unsqueeze(-1) + b.unsqueeze(-2))
+    Cov_ln = torch.special.expm1(Cov) * torch.exp(b.unsqueeze(-1) + b.unsqueeze(-2))
     return mu_ln, Cov_ln
 
 
@@ -88,7 +88,7 @@ def norm_to_lognorm_variance(mu: Tensor, var: Tensor) -> Tensor:
         The `batch_shape x n` variance vector of the log-Normal distribution.
     """
     b = mu + 0.5 * var
-    return (torch.exp(var) - 1) * torch.exp(2 * b)
+    return torch.special.expm1(var) * torch.exp(2 * b)
 
 
 def expand_and_copy_tensor(X: Tensor, batch_shape: torch.Size) -> Tensor: