5
$\begingroup$

From the slides of variational inference, it shows the evidence lower bound ($L$) and the derivative over a variational distribution $q(z_k)$, quoted as follows

$$ L_k = \int q(z_k) E_{-k} \bigg[ \log p(z_k|z_{-k},x) \bigg] dz_k - \int q(z_k) \log q(z_k) dz_k \tag{22} $$

equation (22) focuses on the $k$th latent variable $z_k$, and assumes other latent variables $z_{-k}$ as constants, so that from (22), the derivative of $z_k$ can be derived

$$ \frac{dL}{dq(z_k)} = E_{-k} \bigg[ \log p(z_k|z_{-k},x) \bigg] - \log q(z_k) - 1 = 0 \tag{23} $$

and then by a simple reformulation, we can now have the coordinate ascent update rule for $q(z_k)$

$$ q^*(z_k) \propto \exp \bigg\{ E_{-k} \bigg[ \log p(z_k|z_{-k},x) \bigg] \bigg\} \tag{24} $$

Question, what I don't understand is how (23) can be like that?

Because according to my understanding of (22), the integral is over $z_k$ which is $\int_{z_k}$, not over $q(z_k)$ which is $\int_{q(z_k)}$. If so, instead of deriving $\frac{dL}{dz_k}$, how to derive $\frac{dL}{dq(z_k)}$?

$\endgroup$
1

1 Answer 1

3
$\begingroup$

I'll start with a quick derivation of the Euler-Lagrange formula, then I'll show how you can use it, then I'll show how it applies to your problem.

Background

Equation 22 is a functional -- a function that takes a function as input and outputs a real number.

We want to find the function $q_k(z_k)$ such that the functional in equation 22 is maximized. In other words, we'd like to pick a function $q_k(z_k)$ such that the left hand side, $L_k$, of equation 22 is as large as possible. You know how to do this if I give you a function $f(x)$ of some variable x and ask you to maximize it; you start by looking for stationary points -- you differentiate $f(x)$ w/r/t x, set equal to zero and solve for x to find the stationary points. The difference here is that instead of trying to find the values of the input variable which are stationary points for a function, you are trying to find the input functions which are stationary points of the functional.

Let's assume we're dealing with a normed linear space of functions defined on some domain [a,b] in x, i.e. a set of functions $V$ that satisfy some basic criteria ($f + g = g + f \hspace{0.4cm} \forall \hspace{0.2cm}f,g \in V$, there is a zero element 0 such that $0 + f = f$, for each $f$ in $V$ there is some element $g$ such that $f + g = 0$, etc.) and has an associated norm $\Vert f\Vert$ that yields the "length" of element $f$ from the set of functions. This enables us to quantify the "distance" between two functions f and g in this function space as $\Vert f-g\Vert$. One example of a possible norm on a space of functions of a single variable on a domain [a,b] might be:

$$\int_a^b|f(x)|dx$$

Let's further assume all the functions in this space are well-behaved on the domain of interest -- both continuous and twice-differentiable. We can define an operation on functionals that accept as input functions from this space called the Gateaux derivative which is analogous to the directional derivative in 3-space. In other words, let's say we have a functional $F[f(x)]$ where $f(x)$ is from the set $V$ and we add to it some function $h(x)$ also from $V$ multiplied by a small constant $\epsilon$. Let's stipulate that $h(x)$ is a function that is zero at both ends of the domain on which the functions in $V$ are defined, i.e. $h(a) = h(b) = 0$. The Gateaux derivative is then:

$$\delta F[f(x), h(x)] =\lim_{\epsilon\to0}\frac{F[f(x) + \epsilon h(x)] - F[f(x)]}{\epsilon} = \frac{d}{d\epsilon}F[f(x) + \epsilon h(x)] |_{\epsilon=0}$$

It can be shown that for a function $g(x)$ to be a stationary point of the functional $F[f(x)]$ (i.e. a possible maximum or minimum), a necessary condition is that $\delta F[g(x), h(x)]$ must be zero for all $h(x)$ that meet our criteria.

Now say you have a functional $F[f(x)]$ that takes function $f(x)$ as input and outputs a real number, where $f(x)$ is an element from this set of functions we're working with. Let's start from a function $f_0$ and add to it $\epsilon h$, where $h$ is another function in the set and $\epsilon$ is a constant small enough that $f_0 + \epsilon h$ is still in the set $V$. Let's say that $F$ is of the form:

$$F[f_0(x)] = \int_a^bL(x, f_0(x) + \epsilon h(x), f_0'(x) + \epsilon h'(x))dx$$

where $f_0'(x)$ and $h'(x)$ are the derivatives of $f_0(x)$ and $h(x)$. In other words, we're working with a functional that is an integral on [a,b] of an expression $L$ that contains $x$, $f_0(x)$ and $f_0'(x)$, and we're asking what happens if we add another function in the set multiplied by a small constant $\epsilon$ to the starting input $f_0$. You can probably see a few analogies to differentiating a function at a point $x_0$. For simplicity, I'll now start writing $f_0$ and $h$ in place of $f_0(x)$ and $h(x)$.

Now let's differentiate both sides of our functional w/r/t the constant epsilon. (We can take the derivative inside the integral because we already stipulated all our functions are well-behaved.)

$$\frac{d}{d\epsilon}F[f_0 + \epsilon h(x)] = \int_a^b \frac{\partial}{\partial \epsilon} L(x, f_0 + \epsilon h, f_0' + \epsilon h') dx = $$

$$\frac{d}{d\epsilon}F[f_0 + \epsilon h(x)] = \int_a^b \frac{\partial}{\partial f_0} L(x, f_0 + \epsilon h, f_0' + \epsilon h') h + \frac{\partial}{\partial f_0'} L(x, f_0 + \epsilon h, f_0' + \epsilon h') h' dx$$

We said that $\delta F[f(x), h(x)] = \frac{d}{d\epsilon}F[f(x) + \epsilon h(x)] |_{\epsilon=0}$; we just found $\frac{d}{d\epsilon}F[f(x) + \epsilon h(x)]$, so:

$$\delta F[f(x), h(x)] = \frac{d}{d\epsilon}F[f(x) + \epsilon h(x)] |_{\epsilon=0} = \int_a^b \frac{\partial}{\partial f_0} L(x, f_0, f_0') h + \frac{\partial}{\partial f_0'} L(x, f_0, f_0') h' dx$$

We want to find where this is equal to zero. So do integration by parts on the right hand side; because we stipulated that $h(a) = h(b) = 0$, this is easily seen to be:

$$\int_a^b \frac{\partial}{\partial f_0} L(x, f_0, f_0') h + \frac{\partial}{\partial f_0'} L(x, f_0, f_0') h dx$$

The fundamental lemma of the calculus of variations (I won't prove this here) says that if $\int_a^b f(x) h(x) dx = 0$ for all twice differentiable, continuous $h(x)$ such that $h(a) = h(b) = 0$, it must be true that $f(x) = 0$ for all x in the domain [a,b] of interest. This gives us the Euler-Lagrange formula:

$$\frac{\partial}{\partial f(x)}L(x, f(x), f'(x)) - \frac{d}{dx}\frac{\partial}{\partial f'(x)}L(x, f(x), f'(x)) = 0$$

To recap, we now know that if we want to find stationary points of a functional of the form:

$$F[f(x)] = \int_a^b L(x, f(x), f'(x))dx$$

and the functions we're considering meet some basic criteria, these stationary points (maxima and minima) must satisfy Euler-Lagrange. You want to maximize equation 22, the criteria we named all apply, and equation 22 is of this form, therefore you'll use Euler-Lagrange to find functions that are stationary points (and therefore possible maxima of your lower bound). Now let's look at a couple examples of how to use this.

Example

To use Euler-Lagrange, we're going to take the expression inside the integral in your functional, take the derivative of it with respect to $f(x)$ and $f'(x)$, plug these into Euler-Lagrange and solve the resulting differential equation. You'll take the derivative w/r/t $f(x)$ in the same way that you would if f(x) were a variable a instead of a function. So for example, if your functional is:

$$\int_a^b f(x)^2 + xf'(x)dx$$

then $L(x, f(x), f'(x)) = f(x)^2 + xf'(x)$, and $\frac{\partial}{\partial f(x)} = 2f(x)$ while $\frac{\partial}{\partial f'(x)} = x$.

A common example is the task of finding the shortest path between two points in 2-space. We can use the arc length integral:

$$\int_a^b \sqrt{1 + f'(x)^2}dx$$

In which case $L(x, f(x), f'(x)) = \sqrt{1 + f'(x)^2}$. The derivative w/r/t f(x) is 0. You can easily see that the derivative w/r/t f'(x) is $\frac{f'(x)}{\sqrt{1 + f'(x)^2}}$ (just mentally pretend that $f'(x)$ is some variable a and take the derivative w/r/t a). Plug this into Euler-Lagrange and you get:

$$\frac{d}{dx}\frac{f'(x)}{\sqrt{1 + f'(x)^2}} = 0$$

This differential equation clearly yields that:

$$\frac{f'(x)}{\sqrt{1 + f'(x)^2}} = C$$

Solve for $f'(x)$ and you'll see that $f'(x) = C_2$ for some constant $C_2$. Integrate w/r/t x and you'll see that $y=C_2x + C_3$ -- you can apply the boundary conditions to find the constants. Stationary points can be maxima or minima; in this case there is only one stationary point and we can show it is a minimum, thus establishing that the shortest path between two straight points is a line. (You knew this anyway, but this is another way to get to the same outcome.)

Equation 22

For your equation 22, you have a functional of $q_k(z_k)$ and would like to maximize it. The formula inside the integral, the $L$ expression or Lagrangian, is:

$$q_k(z_k)E_{-k}[log(p(z_k | z_{-k},x))] - q_k(z_k) \log q_k(z_k)$$

In this case the function that is input to the functional is $q_k(z_k)$. So we'll take the derivative of $L$ w/r/t $q_k(z_k)$ and $q_k'(z_k)$ and plug these into Euler-Lagrange. The derivative w/r/t $q_k'(z_k)$ is zero. The derivative w/r/t $q_k(z_k)$ (again, just mentally treat $q_k(z_k)$ as a variable a and take the derivative w/r/t it) is clearly:

$$E_{-k}[\log(p(z_k | z_{-k},x))] - \log q_k(z_k) - 1$$

Plug this into Euler-Lagrange and we have that the stationary point of your functional (the only point that could be a maximum of your functional, the KL divergence from the true posterior) is found when:

$$E_{-k}[\log(p(z_k | z_{-k},x))] - \log q_k(z_k) - 1 = 0$$

and there you have it!

You may find chapter 4 of Logan's Applied Mathematics helpful as well, this will provide a similar but more detailed overview.

$\endgroup$
7
  • $\begingroup$ > "... doesn't include $q_k'(z_k)$, so the second term in Euler-Lagrange is zero"; What if $f(x) = x^2/2$, so that $f'(x) = x$. I think you are correct in this application, but the notation from Wikipedia, doesn't make things clear or might outright be wrong. $\endgroup$ Commented May 26, 2023 at 14:19
  • $\begingroup$ There is no contradiction there. Euler-Lagrange differentiates a functional w/r/t the function f(x) in the functional. If your function is $f(x) = x^2 / 2$ and your functional is $F[f(x)] = \int_a^b f(x) dx$, to apply Euler-Lagrange, you differentiate the expression inside the integral w/r/t f(x) using the formula above. Same if your functional is $F[f(x)] = \int_a^b f(x) f'(x) dx$. Maybe I need to clarify in the answer to make this easier to follow. $\endgroup$
    – HappyDog
    Commented May 26, 2023 at 16:25
  • $\begingroup$ Here are some nice examples that may help you: bjlkeng.io/posts/the-calculus-of-variations . I'll add a more detailed walk-through of the last steps in deriving the update equations for mean-field to the answer later to avoid any confusion. $\endgroup$
    – HappyDog
    Commented May 26, 2023 at 16:44
  • $\begingroup$ $x^2$ wasn't a great choice, if you consider $f(x) = \exp(x)$ you can see that $f(x) = f'(x)$, but somehow $\frac{dL}{df(x)} \neq \frac{dL}{df'(x)}$. $L$ still being a function not a functional. However reading your link, is just bad, but perhaps common notation. If instead we define $L(a,b,c)$ and consider $\frac{dL}{da} - \frac{d}{da}\frac{dL}{dc}$ evaluated at $(x, f(x), f'(x))$ I would be happy, although unsure if it really is $\frac{d}{da}$ or $\frac{d}{dx}$ . $\endgroup$ Commented May 26, 2023 at 17:23
  • 1
    $\begingroup$ @HappyDog, I have formatted your equations. Please check them. If you want, you can rollback them. $\endgroup$ Commented Jun 2, 2023 at 15:47

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.