Variational Gaussian approximation with only $O(N)$ free parameters

Nov 27, 2023·
Kalvik
· 9 min read
Table of Contents

Variational Gaussian Approximation

It is often the case that there is no closed-form solution to the posterior $p(\mathbf{x} \mid \mathbf{y})$:

$$ p(\mathbf{x}|\mathbf{y}) = \frac{p(\mathbf{y}|\mathbf{x}) p(\mathbf{x})}{p(\mathbf{y})} $$

For example, when the likelihood $p(\mathbf{y} \mid \mathbf{x})$ is non-Gaussian and the prior $p(\mathbf{x})$ is Gaussian, one has to resort to approximations of the posterior. The variational approximation is a popular method for this task. Given the mathematically elegant properties of Gaussian distributions, practitioners often choose to approximate the posterior as a Gaussian, even when the true posterior is not Gaussian.

But a Gaussian distribution $\mathcal{N}(\mu, \mathbf{\Sigma})$ over $N$ random variables with mean $\mu$ and covariance $\mathbf{\Sigma}$, has $\mathcal{O}(N^2)$ parameters ($N$ mean parameters and $N^2$ covariance parameters) that need to be optimized, referred to as variational parameters. Consequently, the variational Gaussian approximation does not scale well. However, Opper and Archambeau, 2009 showed that if the prior distribution $p(\mathbf{x})$ is a Gaussian and the likelihood $p(\mathbf{y} \mid \mathbf{x})$ is factorized, the number of free variational parameters is only $\mathcal{O}(N)$, i.e., only $\mathcal{O}(N)$ of the $\mathcal{O}(N^2)$ variational parameters need to be optimized.

I found the original derivation of Opper and Archambeau, 2009 to be terse and difficult to follow initially. Therefore, this article will provide a detailed derivation of their key results.

Variational Free Energy (VFE)

alt text

The variational approach aims to approximate the posterior $p(\mathbf{x} \mid \mathbf{y})$ with the variational distribution $q(\mathbf{x})$, where the distribution’s parameters determine its density. The optimal $q^*(\mathbf{x})$ is obtained by minimizing the KL divergence:

$$ q^*(\mathbf{x}) = \text{arg} \min_{q(\mathbf{x})} \text{KL}(q(\mathbf{x}) || p(\mathbf{x} | \mathbf{y})) $$

The KL can be reformulated as follows:

$$ \begin{aligned} \text{KL}(q(\mathbf{x}) || p(\mathbf{x} | \mathbf{y})) &= \mathbb{E}_{q(\mathbf{x})} [\ln q(\mathbf{x})] - \mathbb{E}_{q(\mathbf{x})} [\ln p(\mathbf{x}|\mathbf{y})] \\ &= \mathbb{E}_{q(\mathbf{x})} [\ln q(\mathbf{x})] - \mathbb{E}_{q(\mathbf{x})} \left[ \ln \frac{p(\mathbf{x}, \mathbf{y})}{p(\mathbf{y})} \right] \\ &= \mathbb{E}_{q(\mathbf{x})} [\ln q(\mathbf{x})] - \mathbb{E}_{q(\mathbf{x})} [\ln p(\mathbf{x}, \mathbf{y})] + \mathbb{E}_{q(\mathbf{x})}[\ln p(\mathbf{y})] \\ &= \mathbb{E}_{q(\mathbf{x})} [\ln q(\mathbf{x})] - \mathbb{E}_{q(\mathbf{x})} [\ln p(\mathbf{x}, \mathbf{y})] + \ln p(\mathbf{y}) \\ \end{aligned} $$

In the last equation above, $\ln p(\mathbf{y})$ is usually computationally intractable. Since we cannot compute the above KL, we instead optimize the variational free energy $\mathcal{F}$, which is equivalent to the above KL up to an added constant:

$$ \begin{aligned} \mathcal{F} &= \underbrace{\mathbb{E}_{q(\mathbf{x})} [\ln q(\mathbf{x})]}_{-\text{entropy of } q(\mathbf{x})} - \mathbb{E}_{q(\mathbf{x})} [\ln p(\mathbf{x}, \mathbf{y})] \\ &= -\frac{1}{2} \ln |\mathbf{\Sigma}| -\frac{N}{2} \ln(2\pi e) - \langle \ln p(\mathbf{x}, \mathbf{y}) \rangle_{q(\mathbf{x})} \end{aligned} $$

When the variational distribution $q(\mathbf{x})$ is a Gaussian $\mathcal{N}(\mu, \mathbf{\Sigma})$, the solution $q^*(\mathbf{x})$ can be obtained by taking the derivative of the variational free energy $\mathcal{F}$ with respect to the parameters $\mu$ and $\mathbf{\Sigma}$ of the variational distribution $q(\mathbf{x})$ and setting them equal to zero:

$$ \nabla_{\mu} \mathcal{F} = - \nabla_{\mu} \langle \ln p(\mathbf{x}, \mathbf{y}) \rangle_{q(\mathbf{x})} $$ $$ \begin{aligned} \nabla_{\mathbf{\Sigma}} \mathcal{F} &= - \frac{1}{2} \mathbf{\Sigma}^{-\top} - \nabla_{\mathbf{\Sigma}} \langle \ln p(\mathbf{x}, \mathbf{y}) \rangle_{q(\mathbf{x})} \\ \implies \mathbf{\Sigma}^{-\top} &= - 2\nabla_{\mathbf{\Sigma}} \langle \ln p(\mathbf{x}, \mathbf{y}) \rangle_{q(\mathbf{x})} \end{aligned} $$

VFE for Gaussian Prior and Factorized Likelihood

So far, the only assumption we made was that the variational distribution $q(\mathbf{x})$ is Gaussian. As such, the above variational approach does not scale well (i.e., $\mathcal{O}(N^2)$ variational parameters). However, with two additional assumptions:

$$ \begin{aligned} p(\mathbf{x}) &\sim \mathcal{N}(0, \mathbf{K}) \\ p(\mathbf{y}|\mathbf{x}) &= \prod_n p(y_n|x_n) \end{aligned} $$

i.e., the prior distribution $p(\mathbf{x})$ is also a Gaussian (which can also have a non-zero mean), and if the likelihood of the data $p(\mathbf{y}\mid\mathbf{x})$ factorizes, it can be shown that the number of free variational parameters is only $\mathcal{O}(N)$. Opper and Archambeau demonstrated this by plugging the joint distribution $p(\mathbf{x}, \mathbf{y})$ with the above assumptions into the variational free energy $\mathcal{F}$. To see this, we first compute $\ln p(\mathbf{x}, \mathbf{y})$:

$$ \begin{aligned} p(\mathbf{x}, \mathbf{y}) &= p(\mathbf{x}) p(\mathbf{y}|\mathbf{x}) \\ &= \frac{1}{(2\pi)^\frac{N}{2}|\mathbf{K}|^\frac{1}{2}} \exp \left\{ -\frac{1}{2}({\mathbf{x}^\top \mathbf{K}^{-1} \mathbf{x}}) \right\} \exp \left\{\ln \prod_n p(y_n|x_n) \right\} \\ &= \frac{1}{Z_0} \exp \left\{ -\frac{1}{2}({\mathbf{x}^\top K^{-1} \mathbf{x}}) \right\} \exp \left\{- \sum_n - \ln p(y_n|x_n) \right\} \\ \implies & \ln p(\mathbf{x}, \mathbf{y}) = -\ln Z_0 -\frac{1}{2}({\mathbf{x}^\top \mathbf{K}^{-1} \mathbf{x}}) - \sum_n -\ln \left[p(y_n|x_n) \right] \\ \end{aligned} $$

In the above, note that the likelihoods $p(y_n \mid x_n)$ can have any distribution and are not limited to Gaussians. Although the likelihoods appear in an exponential form, the natural log cancels the exponent, yielding the original product of likelihoods $\prod_n p(y_n \mid x_n)$. We also collect the normalization constants, including $\mid \mathbf{K}\mid $ in $Z_0$. Now, we can compute the expectation of $\ln p(\mathbf{x}, \mathbf{y})$ with respect to the variational distribution $q(\mathbf{x})$, which will be used to compute the variational free energy $\mathcal{F}$:

$$ \begin{aligned} \langle & \ln p(\mathbf{x}, \mathbf{y}) \rangle_{q(\mathbf{x})} = \langle -\frac{1}{2}({\mathbf{x}^\top \mathbf{K}^{-1} \mathbf{x}}) - \sum_n -\ln \left[p(y_n|x_n) \right] - \ln Z_0 \rangle_{q(\mathbf{x})} \\ &= -\langle \frac{1}{2}({\mathbf{x}^\top \mathbf{K}^{-1} \mathbf{x}}) \rangle_{q(\mathbf{x})} - \sum_n \langle -\ln \left[p(y_n|x_n) \right] \rangle_{q(x_n)} - \langle \ln Z_0 \rangle_{q(\mathbf{x})} \\ &= -\langle \frac{1}{2}\text{Tr}(\mathbf{K}^{-1} \mathbf{x}\mathbf{x}^\top) \rangle_{q(\mathbf{x})} - \sum_n \langle -\ln \left[p(y_n|x_n) \right] \rangle_{q(x_n)} - \ln Z_0 \\ &= - \frac{1}{2}\text{Tr}(\mathbf{K}^{-1} (\mathbf{\Sigma} + \mu\mu^\top)) - \sum_n -\langle \ln \left[p(y_n|x_n) \right] \rangle_{q(x_n)} - \ln Z_0 \\ &= - \frac{1}{2}\text{Tr}(\mathbf{K}^{-1}\mathbf{\Sigma}) - \frac{1}{2}\text{Tr}(K^{-1}\mu\mu^\top) - \sum_n \langle -\ln \left[p(y_n|x_n) \right] \rangle_{q(x_n)} - \ln Z_0 \\ &= - \frac{1}{2}\text{Tr}(\mathbf{K}^{-1}\mathbf{\Sigma}) - \frac{1}{2}\mu^\top K^{-1}\mu - \sum_n \langle -\ln \left[p(y_n|x_n) \right] \rangle_{q(x_n)} - \ln Z_0 \\ \end{aligned} $$

Now we can plug the above expectation into the variational free energy $\mathcal{F}$ to get the following:

$$ \begin{aligned} \mathcal{F} &= \mathbb{E}_{q(\mathbf{x})} [\ln q(\mathbf{x})] - \mathbb{E}_{q(\mathbf{x})} [\ln p(\mathbf{x, y})] \\ &= -\frac{1}{2} \ln |\mathbf{\Sigma}| -\frac{N}{2} \ln(2\pi e) + \frac{1}{2}\text{Tr}(\mathbf{K}^{-1}\mathbf{\Sigma}) + \\ & \quad \quad \frac{1}{2}\mu^\top \mathbf{K}^{-1}\mu + \sum_n \langle -\ln \left[p(y_n|x_n) \right] \rangle_{q(x_n)} + \ln Z_0 \\ \end{aligned} $$

We can compute the solution covariance matrix $\mathbf{\Sigma}$ of the variational distribution $q(\mathbf{x})$ by taking the derivative w.r.t $\mathbf{\Sigma}$ and setting it equal to zero:

$$ \begin{aligned} \nabla_{\mathbf{\Sigma}} \mathcal{F} &= \nabla_{\mathbf{\Sigma}} \left[ -\frac{1}{2} \ln |\mathbf{\Sigma}| -\frac{N}{2} \ln(2\pi e) + \frac{1}{2}\text{Tr}(\mathbf{K}^{-1}\mathbf{\Sigma}) \right. \\ & \quad \quad \quad \quad \left. + \frac{1}{2}\mu^\top \mathbf{K}^{-1}\mu + \sum_n \langle -\ln \left[p(y_n|x_n) \right] \rangle_{q(x_n)} + \ln Z_0 \right] \\ &= -\nabla_{\mathbf{\Sigma}} \frac{1}{2} \ln |\mathbf{\Sigma}| + \nabla_{\mathbf{\Sigma}} \frac{1}{2}\text{Tr}(\mathbf{K}^{-1}\mathbf{\Sigma}) + \nabla_{\mathbf{\Sigma}} \sum_n \langle -\ln \left[p(y_n|x_n) \right] \rangle_{q(x_n)} \\ &= - \frac{1}{2} \mathbf{\Sigma}^{-\top} + \frac{1}{2}\mathbf{K}^{-1} + \sum_n \nabla_{\mathbf{\Sigma}} \langle -\ln \left[p(y_n|x_n) \right] \rangle_{q(x_n)} \\ \implies& \frac{1}{2} \mathbf{\Sigma}^{-\top} = \frac{1}{2}\mathbf{K}^{-1} + \sum_n \nabla_{\mathbf{\Sigma}} \langle -\ln \left[p(y_n|x_n) \right] \rangle_{q(x_n)} \\ \implies& \mathbf{\Sigma}^{-\top} = \mathbf{K}^{-1} + 2\sum_n \nabla_{\mathbf{\Sigma}} \langle -\ln \underbrace{\left[p(y_n|x_n) \right]}_\text{likelihood} \rangle_{q(x_n)} \\ \implies& \mathbf{\Sigma}^{-\top} = \left( \mathbf{K}^{-1} + \mathbf{\Lambda} \right) \\ \end{aligned} $$

Due to our factorization assumption for the data likelihood, i.e., the likelihood of each $y_n$ depends only on $x_n$, the expectation of each likelihood $p(y_n\mid x_n)$ with respect to the variational distribution depends solely on $q(x_n)$—a univariate variational Gaussian distribution over $x_n$. Similarly, the derivative of each likelihood with respect to the variational distribution depends only on the $n^\text{th}$ mean term $\mu_n$ and covariance term $\Sigma_{nn}$ (an element of the diagonal).

Therefore, we can conclude that the solution Gaussian variational distribution $q^*(\mathbf{x})$ will have a precision matrix $\mathbf{\Sigma}^{-1}$ whose off-diagonal terms are the same as the prior precision matrix $\mathbf{K}^{-1}$, and the diagonal terms will have the additional updates from the diagonal matrix $\mathbf{\Lambda}$, obtained from the derivatives of the likelihood.

This implies that our Gaussian variational distribution $q^*(\mathbf{x})$ has $N$ mean parameters and only $N$ effective covariance parameters, resulting in $\mathcal{O}(N)$ parameters. This represents a significant reduction from the conventional $\mathcal{O}(N^2)$ parameters required when optimizing the full covariance matrix instead of a diagonal matrix, making the approach scalable for variational approximations involving a large number of variables.

New Efficient Parameterization

Based on the above results, Opper and Archambeau suggested defining the mean parameters $\mu$ of the variational distribution as follows, where $\mathcal{v}$ is a vector consisting of the new mean parameters:

$$ \mu = \mathbf{K}\mathcal{v} $$

This simplifies the variational free energy $\mathcal{F}$ into the following which does not require us to invert the prior covariance $\mathbf{K}$:

$$ \begin{aligned} \mathcal{F} =& -\frac{1}{2} \ln |\mathbf{\Sigma}| -\frac{N}{2} \ln(2\pi e) + \frac{1}{2}\text{Tr}(\mathbf{K}^{-1}\mathbf{\Sigma}) + \\ & \quad \frac{1}{2} \mu^\top \mathbf{K}^{-1}\mu + \sum_n \langle -\ln \left[p(y_n|x_n) \right] \rangle_{q(x_n)} + \ln Z_0 \\ =& -\frac{1}{2} \ln |\mathbf{\Sigma}| -\frac{N}{2} \ln(2\pi e) + \frac{1}{2}\text{Tr}(\mathbf{K}^{-1}\mathbf{\Sigma}) + \\ & \quad \frac{1}{2} \mathcal{v}^\top\mathbf{K}^\top\mathbf{K}^{-1}\mathbf{K}\mathcal{v} + \sum_n \langle -\ln \left[p(y_n|x_n) \right] \rangle_{q(x_n)} + \ln Z_0 \\ =& -\frac{1}{2} \ln |\mathbf{\Sigma}| -\frac{N}{2} \ln(2\pi e) + \frac{1}{2}\text{Tr}(\mathbf{K}^{-1}\mathbf{\Sigma}) + \\ & \quad \frac{1}{2} \mathcal{v}^\top\mathbf{K}^\top\mathcal{v} + \sum_n \langle -\ln \left[p(y_n|x_n) \right] \rangle_{q(x_n)} + \ln Z_0 \\ \end{aligned} $$

The above gives us the following gradients with respect to the new parameterization $\mathcal{v}$:

$$ \begin{aligned} \frac{\partial\mathcal{F}}{\partial\mathcal{v}} &= \frac{\partial\mu}{\partial\mathcal{v}} \frac{\partial\mathcal{F}}{\partial\mathcal{\mu}} \\ &= \frac{\partial}{\partial\mathcal{v}}\mathbf{K}\mathcal{v} \left[ \frac{\partial}{\partial\mathcal{\mu}} \frac{1}{2}\mu^\top \mathbf{K}^{-1}\mu - \underbrace{-\frac{\partial}{\partial\mathcal{\mu}}\sum_n \langle -\ln \left[p(y_n|x_n) \right] \rangle_{q(x_n)}}_{\bar{\mathcal{v}}} \right] \\ &= \mathbf{K} \left[ \mathbf{K}^{-1}\mu - \bar{\mathcal{v}} \right] \\ &= \mathbf{K} \left[ \mathbf{K}^{-1}\mathbf{K}\mathcal{v} - \bar{\mathcal{v}} \right] \\ &= \mathbf{K} \left[\mathcal{v} - \bar{\mathcal{v}} \right] \\ \end{aligned} $$

Finally, Opper and Archambeau suggested defining the covariance parameters with the vector $\mathbf{\lambda}$, which is the diagonal of the diagonal matrix $\mathbf{\Lambda}$, giving us the following gradients:

$$ \begin{aligned} \frac{\partial\mathcal{F}}{\partial\mathbf{\Lambda}} &= \frac{\partial\mathbf{\Sigma}}{\partial\mathbf{\Lambda}} \frac{\partial\mathcal{F}}{\partial\mathcal{\mathbf{\Sigma}}} \\ &= \left[ \frac{\partial}{\partial\mathbf{\Lambda}} \left( \mathbf{K}^{-1} + \mathbf{\Lambda} \right)^{-\top} \right] \frac{1}{2} \Biggl[ \underbrace{- \mathbf{\Sigma}^{-\top} + \mathbf{K}^{-1}}_{-\mathbf{\Lambda}} \\ & \quad \quad \quad \quad \quad \quad \quad \quad \quad + \underbrace{2 \sum_n \frac{\partial}{\partial\mathbf{\Sigma}} \langle -\ln \left[p(y_n|x_n) \right] \rangle_{q(x_n)}}_{\bar{\mathbf{\Lambda}}} \Biggr] \\ &= \frac{1}{2}\frac{\partial}{\partial\mathbf{\Lambda}} \left( \mathbf{K}^{-1} + \mathbf{\Lambda} \right)^{-\top}[-\mathbf{\Lambda}+\bar{\mathbf{\Lambda}}]\\ &= \frac{1}{2}-\left( \mathbf{K}^{-1} + \mathbf{\Lambda} \right)^{-\top} \frac{\partial \left( \mathbf{K}^{-1} + \mathbf{\Lambda} \right)}{\partial\mathbf{\Lambda}} \left( \mathbf{K}^{-1} + \mathbf{\Lambda} \right)^{-\top} [-\mathbf{\Lambda}+\bar{\mathbf{\Lambda}}] \\ &= \frac{1}{2}[\mathbf{\Sigma} \mathbf{I} \mathbf{\Sigma}] [\mathbf{\Lambda}-\bar{\mathbf{\Lambda}}]\\ &= \frac{1}{2}[\mathbf{\Sigma} \circ \mathbf{\Sigma}] [\mathbf{\Lambda}-\bar{\mathbf{\Lambda}}]\\ &\implies \frac{\partial\mathcal{F}}{\partial\mathbf{\lambda}} = \frac{1}{2}[\mathbf{\Sigma} \circ \mathbf{\Sigma}] [\mathbf{\lambda}-\bar{\mathbf{\lambda}}] \end{aligned} $$

Conclusion

In short, we resort to approximate inference methods when computing the original distributions becomes challenging. In Variational Gaussian approximation, a variational Gaussian distribution is employed to approximate the distribution of interest. However, this method entails optimizing $N$ mean parameters and $N^2$ covariance parameters for the solution variational Gaussian distribution over $N$ random variables.

Opper and Archambeau, 2009 demonstrated that assuming a Gaussian prior and a factorized likelihood allows for the optimization of $N$ mean parameters and only $N$ covariance parameters. This significantly enhances the scalability of the method for larger problems. Furthermore, they proposed a new parameterization for the variational Gaussian distribution to achieve these improved results. Notably, their proposed parameterization is akin to the natural parametrization of Gaussian distributions, as recently shown by Khan and Lin 2017 to also result in only $\mathcal{O}(N)$ free variables in this problem context.

Finally, it’s worth noting that Opper and Archambeau considered a full approximation, i.e., they used an $N$-dimensional variational Gaussian distribution to approximate the true posterior distribution of $N$ random variables. However, if we are interested in a sparse approximation, considering only an $M$-dimensional variational Gaussian distribution where $M\ll N$, then we will have to optimize the full covariance matrix of the variational distribution. Although in such cases, the difference between $\mathcal{O}(M^2)$ and $\mathcal{O}(N)$ will be significant enough that the full covariance matrix of the sparse approximation would still have fewer variational parameters than a full approximation with a diagonal covariance matrix.

Relevant Identities

  • $\mathbb{E}[\mathbf{x}\mathbf{x}^\top] = \mathbb{E}[\mathbf{x}]\mathbb{E}[\mathbf{x}]^\top + \text{cov}[\mathbf{x}]$

  • $\nabla_{\mathbf{X}} \text{Tr}(\mathbf{A}\mathbf{X}) = \mathbf{A}$

  • $\nabla_{\mathbf{X}} \ln \mid\mathbf{X}\mid = \mathbf{X}^{-\top}$

  • $\nabla_{\mathbf{X}} \mathbf{X}^{-1} = -\mathbf{X}^{-1} [\nabla_{\mathbf{X}}\mathbf{X}] \mathbf{X}^{-1}$

Bayesian Learning