Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

How one can prove that $$ \int_{\mathbb{R}^N} \mathcal{N}(\mathbf{y}| \boldsymbol{\mu}_1, K_1) \log \mathcal{N}(\mathbf{y}| \boldsymbol{\mu}_0, K_0) d \mathbf{y} = -\frac12 \left[ N \log 2 \pi + \log |K_0| + \mathrm{tr} (K_0^{-1} K_1) + (\boldsymbol{\mu}_0 - \boldsymbol{\mu}_1)^T K_0^{-1} (\boldsymbol{\mu}_0 - \boldsymbol{\mu}_1) \right]. $$ There $\mathbf{y} \in \mathbb{R}^N$, $|M|$ is the determinant of matrix $M$, $\mathrm{tr}(M)$ is the trace of matrix $K$, $$ \mathcal{N}(\mathbf{y}| \boldsymbol{\mu}, K) = \frac{1}{(2 \pi)^{\frac{N}{2}} |K|^{\frac12}} \exp \left(-\frac12 (\mathbf{y} - \boldsymbol{\mu})^T K^{-1} (\mathbf{y} - \boldsymbol{\mu})\right). $$

For a one-dimensional case the problem is simple, but for higher dimensions I can't directly prove this formula, but it seems to be correct (for example, it can be used to calculate Kullback-Leibler divergence between two gaussian distributions).

Also, it would be sufficient in case one provides a source with the equation mentioned above.

share|cite|improve this question
up vote 0 down vote accepted

The only problem with this integral is the term $$ \int_{\mathbb{R}^N} (\mathbf{y} - \boldsymbol{\mu_1})^T K_1^{-1} (\mathbf{y} - \boldsymbol{\mu_1}) \frac{1}{(2 \pi)^{\frac{N}{2}} |K_0|^{\frac12}} \exp\left(-\frac12 (\mathbf{y} - \boldsymbol{\mu}_0)^T K_0^{-1} (\mathbf{y} - \boldsymbol{\mu}_0)) \right) d\mathbf{y}. $$ Let's make substitution $\mathbf{x} = \mathbf{y} - \boldsymbol{\mu_0}$. We get $$ \int_{\mathbb{R}^N} (\mathbf{x} - \boldsymbol{\mu_1} + \boldsymbol{\mu_0})^T K_1^{-1} (\mathbf{x} - \boldsymbol{\mu_1} + \boldsymbol{\mu_0}) \frac{1}{(2 \pi)^{\frac{N}{2}} |K_0|^{\frac12}} \exp\left(-\frac12 \mathbf{x}^T K_0^{-1} \mathbf{x} \right) d\mathbf{x}. $$ Denote $\boldsymbol{\mu} = \boldsymbol{\mu_1} - \boldsymbol{\mu_0}$. Note that $$ (\mathbf{x} - \boldsymbol{\mu})^T K_1^{-1} (\mathbf{x} - \boldsymbol{\mu}) = \mathbf{x} K_1^{-1} \mathbf{x} - 2 \boldsymbol{\mu}^T K_1^{-1} \mathbf{x} + \boldsymbol{\mu}^T K_1^{-1} \boldsymbol{\mu}. $$

The integral $$ \int_{\mathbb{R}^N} - 2\boldsymbol{\mu}^T K_1^{-1} \mathbf{x} \frac{1}{(2 \pi)^{\frac{N}{2}} |K_0|^{\frac12}} \exp\left(-\frac12 \mathbf{x}^T K_0^{-1} \mathbf{x} \right) d\mathbf{x} = 0. $$

Now we need to calculate $$ \int_{\mathbb{R}^N} \mathbf{x}^T K_1^{-1} \mathbf{x} \frac{1}{(2 \pi)^{\frac{N}{2}} |K_0|^{\frac12}} \exp\left(-\frac12 \mathbf{x}^T K_0^{-1} \mathbf{x} \right) d\mathbf{x}. $$

Matrix $K_1 = U^T U$, $K_1^{-1} = U^{-1} U^{-T}$ ($K_1 > 0$ and symmetric for distribution to be correct). So, we make another substitution $$ \mathbf{z} = U^{-T} \mathbf{x}, \\ \mathbf{x} = U^T \mathbf{z}, \\ d \mathbf{x} = |U|^T d \mathbf{z}. $$ The integral for this substitution has the form: $$ \int_{\mathbb{R}^N} \mathbf{z}^T \mathbf{z} \frac{1}{(2 \pi)^{\frac{N}{2}} |K_0|^{\frac12}} \exp\left(-\frac12 \mathbf{z}^T U K_0^{-1} U^T \mathbf{z} \right) |U|^T d\mathbf{z} = \\ \int_{\mathbb{R}^N} \mathbf{z}^T \mathbf{z} \frac{1}{(2 \pi)^{\frac{N}{2}} |U^{-T} K_0 U^{-1}|^{\frac12}} \exp\left(-\frac12 \mathbf{z}^T (U^{-T} K_0 U^{-1})^{-1} \mathbf{z} \right) d\mathbf{z} = \\ \int_{\mathbb{R}^N} \sum_{i = 1}^N z_i^2 \frac{1}{(2 \pi)^{\frac{N}{2}} |U^{-T} K_0 U^{-1}|^{\frac12}} \exp\left(-\frac12 \mathbf{z}^T (U^{-T} K_0 U^{-1})^{-1} \mathbf{z} \right) d\mathbf{z} = \\ =\mathrm{tr} (U^{-T} K_0 U^{-1}) = \mathrm{tr} (U^{-1} U^{-T} K_0) = \mathrm{tr} (K_1^{-1} K_0). $$

Now we can prove: $$ \int_{\mathbb{R}^N} (\mathbf{y} - \boldsymbol{\mu_1})^T K_1^{-1} (\mathbf{y} - \boldsymbol{\mu_1}) \frac{1}{(2 \pi)^{\frac{N}{2}} |K_0|^{\frac12}} \exp\left(-\frac12 (\mathbf{y} - \boldsymbol{\mu}_0)^T K_0^{-1} (\mathbf{y} - \boldsymbol{\mu}_0)) \right) d\mathbf{y} = \mathrm{tr} (K_1^{-1} K_0) + (\boldsymbol{\mu}_1 - \boldsymbol{\mu}_0)^T K_1^{- 1} (\boldsymbol{\mu}_1 - \boldsymbol{\mu}_0). $$ And finally it is easy to prove the result in question.

share|cite|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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