jax-ml/jax · error · ValueError

precision matrix must match data dims

Error message

precision matrix must match data dims

What it means

In _gaussian_kernel_eval the precision matrix must be a full (d, d) matrix matching the data dimensionality, because it is Cholesky-factorized for whitening. Passing a diagonal vector of precisions, a scalar, or a wrongly sized matrix raises this error.

Source

Thrown at jax/_src/scipy/stats/kde.py:264

def _gaussian_kernel_convolve(chol, norm, target, weights, mean):
  diff = target - mean[:, None]
  alpha = linalg.cho_solve(chol, diff)
  arg = 0.5 * jnp.sum(diff * alpha, axis=0)
  return norm * jnp.sum(jnp.exp(-arg) * weights)


@api.jit(static_argnums=0)
def _gaussian_kernel_eval(in_log, points, values, xi, precision):
  points, values, xi, precision = promote_dtypes_inexact(
      points, values, xi, precision)
  d = points.shape[1]

  if xi.shape[1] != d:
    raise ValueError("points and xi must have same trailing dim")
  if precision.shape != (d, d):
    raise ValueError("precision matrix must match data dims")

  whitening = linalg.cholesky(precision, lower=True)
  points = jnp.dot(points, whitening)
  xi = jnp.dot(xi, whitening)
  log_norm = jnp.sum(jnp.log(
      jnp.diag(whitening))) - 0.5 * d * jnp.log(2 * np.pi)

  def kernel(x_test, x_train, y_train):
    arg = log_norm - 0.5 * jnp.sum(jnp.square(x_train - x_test))
    if in_log:
      return jnp.log(y_train) + arg
    else:
      return y_train * jnp.exp(arg)

  reduce = special.logsumexp if in_log else jnp.sum
  reduced_kernel = lambda x: reduce(api.vmap(kernel, in_axes=(None, 0, 0))
                                    (x, points, values),
                                    axis=0)

View on GitHub (pinned to 1e1c6a8fc0)

Solutions

1. Wrap per-dimension precisions into a diagonal matrix: jnp.diag(precision_vec)
2. Ensure precision.shape == (d, d) where d == points.shape[1]
3. If precision is a Cholesky factor L, pass L @ L.T

Example fix

// before
vals = _gaussian_kernel_eval(False, pts, vals, xi, prec_vec)  # shape (d,)
// after
vals = _gaussian_kernel_eval(False, pts, vals, xi, jnp.diag(prec_vec))

d = points.shape[1]
if precision.ndim == 1:
    precision = jnp.diag(precision)
assert precision.shape == (d, d)

def precision_matches(precision, points) -> bool:
    d = jnp.asarray(points).shape[1]
    return jnp.asarray(precision).shape == (d, d)

When it happens

Trigger: Direct calls to _gaussian_kernel_eval with precision of shape (d,), (1,), or (k, k) with k != d; upstream, this usually surfaces as a wrongly-built inv_cov passed through evaluate.

Common situations: Constructing a per-feature precision vector instead of a matrix; using a Cholesky factor or correlation matrix of the wrong size when wiring custom precision into the KDE.

Related errors

• `weights` input should be one-dimensional.
• mean does not have dimension {self.d}
• covariance does not have dimension {self.d}
• KDEs are not the same dimensionality
• points have dimension {}, dataset has dimension {}

AI-assisted analysis of jax-ml/jax@1e1c6a8fc0 (2026-08-27). Data as JSON: /api/errors/7c87aa96bde59a99. Report an issue: GitHub.