diff --git a/gpytorch/variational/variational_strategy.py b/gpytorch/variational/variational_strategy.py
index 453447da2..3435599df 100644
--- a/gpytorch/variational/variational_strategy.py
+++ b/gpytorch/variational/variational_strategy.py
@@ -1,5 +1,6 @@
 from __future__ import absolute_import, division, print_function, unicode_literals
 
+import math
 import torch
 import gpytorch
 from ..lazy import RootLazyTensor, PsdSumLazyTensor, DiagLazyTensor
@@ -132,7 +133,7 @@ def forward(self, x):
                 if gpytorch.beta_features.diagonal_correction.on():
                     fake_diagonal = (inv_product * induc_data_covar).sum(0)
                     real_diagonal = data_data_covar.diag()
-                    diag_correction = DiagLazyTensor(real_diagonal - fake_diagonal)
+                    diag_correction = DiagLazyTensor((real_diagonal - fake_diagonal).clamp(0, math.inf))
                     predictive_covar = PsdSumLazyTensor(predictive_covar, diag_correction)
 
             return MultivariateNormal(predictive_mean, predictive_covar)