I understand from wikipedia that a variogram model must be positive definite to be used for kriging:

Note that the experimental variogram is an empirical estimate of the covariance of a Gaussian process. As such, it may not be positive definite and hence not directly usable in kriging, without constraints or further processing. This explains why only a limited number of variogram models are used: most commonly, the linear, the spherical, the Gaussian and the exponential models.

...But, I don't understand why the variogram model must be positive definite?

I know that covariance matrices are always positive semi-definite, so I might be able to understand why a positivesemi-definite matrix is necessary. But, why positive definite?

I'm having trouble tying these concepts together: kriging, positive definite matrix, covariance, variogram.


The terminology about positive definite matrices is very inconsistent. Some authors use "positive definite" to mean all-positive eigenvalues and "positive semidefinite" to mean all-nonnegative eigenvalues; some use "positive definite" to mean all-nonnegative eigenvalues and "strictly positive definite" to mean all-positive. Here, I think the latter use is meant: to be able to perform the process of kriging, the matrix must be positive semidefinite.

That said, kriging with singular covariance matrices is mildly unusual, and some code / default ways of thinking about the process might be broken by it.

| cite | improve this answer | |
  • $\begingroup$ okay, thanks. But what is implied by an eigenvalue of 0 in a covariance matrix (since that's what positive semi-definite means)? $\endgroup$ – makansij Sep 18 '16 at 1:50
  • $\begingroup$ @Hunle Let $X \sim \mathcal{N}(\mu, \Sigma)$; an eigenvalue of zero means that there is some $v$ with $\Sigma v = 0$. But that means that if we observe along the direction $v$, i.e. we take $v X$, then that has $\mathrm{Var}[v X] = v^T \mathrm{Var}[X] v = v^T \Sigma v = 0$. So an eigenvalue of 0 means that $v X$ is known exactly. $\endgroup$ – djs Sep 18 '16 at 1:52
  • 1
    $\begingroup$ I do not fully understand these things, but I have seen codes to generate gridded Gaussian random fields via FFT that use "circulant embedding", where in practice they truncate negative (eigen-)values to 0. $\endgroup$ – GeoMatt22 Sep 18 '16 at 1:56
  • $\begingroup$ @Dougal, I'm sorry, but I don't fully understand your comment. what is $v$ in your example? $\endgroup$ – makansij Sep 20 '16 at 6:38
  • $\begingroup$ Once we've seen a bunch of data, we have a probability distribution (a Gaussian process) of what everything should look like. Choose points to observe, and stack up their values in a random vector $X$. $X$ is distributed as a normal distribution with mean $\mu$ and variance $\Sigma$. If $\Sigma$ is singular, with a vector $v$ in its null space, then $v X$ will be known exactly. If $\Sigma$ were the zero matrix, then this is true for any vector $v$, so $X$ would be known exactly. If it had the ones vector in its null space, then we might not know each component of $X$ but we'd know its sum. $\endgroup$ – djs Sep 20 '16 at 6:46

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.