fix: correct dimension handling and stabilize sparse GP posterior

- Fixed incorrect dimension usage for `y_cov_factor` by ensuring it matches the number of observations (`n_obs`) instead of landmarks (`n_landmarks`).
- Added a new helper function `add_projected_variance` to properly project the observation noise covariance and stabilize the covariance matrix.
- Replaced direct variance addition in `_LandmarksConditional` with the new `add_projected_variance` function to maintain numerical stability during Cholesky decomposition.
- Ensured the covariance matrix remains positive definite by adjusting diagonal elements based on a `jitter` threshold.

This commit resolves the dimension mismatch issue and preserves the stabilizing effect, ensuring correct posterior GP computations with uncertainty.

CHANGELOG.md

@@ -4,6 +4,7 @@
  - bugfix DimensionalityEstimator dimensionality initialization
  - implement 'fixed' gaussian proces type to allow more inducing points than datapoints
  - implement `copy()` method for `Predictor` class
+ - fix: uncertainty computation of sparse GP
 
 # v1.4.3
 

mellon/conditional.py

@@ -5,7 +5,7 @@
 from jax.numpy import diag as diagonal
 from jax.numpy.linalg import cholesky
 from jax.scipy.linalg import solve_triangular
-from .util import ensure_2d, stabilize, DEFAULT_JITTER, add_variance
+from .util import ensure_2d, stabilize, DEFAULT_JITTER, add_variance, add_projected_variance
 from .base_predictor import Predictor, ExpPredictor, PredictorTime
 from .decomposition import DEFAULT_SIGMA
 
@@ -278,9 +278,9 @@ def __init__(
             LLB = stabilize(LLB, jitter)
         else:
             logger.debug("Assuming y is not the mean of the GP.")
-            y_cov_factor = _sigma_to_y_cov_factor(sigma, y_cov_factor, xu.shape[0])
+            y_cov_factor = _sigma_to_y_cov_factor(sigma, y_cov_factor, x.shape[0])
             sigma = None
-            LLB = add_variance(LLB, y_cov_factor, jitter=jitter)
+            LLB = add_projected_variance(LLB, A, y_cov_factor, jitter=jitter)
 
         L_B = cholesky(LLB)
         r = y - mu

mellon/util.py

@@ -315,6 +315,33 @@ def add_variance(K, M=None, jitter=DEFAULT_JITTER):
     return K
 
 
+def add_projected_variance(K, A, y_cov_factor, jitter=DEFAULT_JITTER):
+    """
+    Adds the projected observation noise covariance to K and stabilizes it.
+
+    Parameters
+    ----------
+    K : array_like, shape (n_landmarks, n_landmarks)
+        The initial covariance matrix.
+    A : array_like, shape (n_landmarks, n_obs)
+        The projection matrix from observations to inducing points.
+    y_cov_factor : array_like, shape (n_obs, n_obs)
+        The observation noise covariance matrix.
+    jitter : float, optional
+        A small number to stabilize the covariance matrix. Defaults to 1e-6.
+
+    Returns
+    -------
+    stabilized_K : array_like, shape (n_landmarks, n_landmarks)
+        The stabilized covariance matrix with added projected variance.
+    """
+    noise = A @ y_cov_factor @ A.T
+    noise_diag = np.diag(noise)
+    diff = where(noise_diag < jitter, jitter - noise_diag, 0)
+    K = K + noise + np.diag(diff)
+    return K
+
+
 def mle(nn_distances, d):
     R"""
     Nearest Neighbor distribution maximum likelihood estimate for log density