The commonly used $3\times 3$ Laplacian convolution kernel $\frac{1}{6}\begin{bmatrix}1 & 4 & 1\\4 & -20 & 4\\1 & 4 & 1\end{bmatrix}$ is only an approximation of rotational symmetry, as mentioned in the Wikipedia article on the Discrete Laplace operator.

Also, I found the first order Sobel operator to have an exactly rotational symmetric form without any proof.

So my questions: How can we compute the exact value of a rotational symmetric $n\times n$ convolution kernel for

  1. the $\nabla$ operator?
  2. the Laplacian operator?

1 Answer 1


I don't think you can. The problem is that you represent a function $f(\mathbf x)$ on a Cartesian mesh and you apply the operator to that. If you rotate the function, i.e., you try $f(A\mathbf x)$ (where $A$ is a $2\times 2$ rotation matrix), then you will get different function values at the vertices of the mesh and different values if you apply the discrete Laplace to it. In other words, even though the Laplacian operator is rotationally invariant, the Laplace operator on a mesh is not and can not be.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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