1
$\begingroup$

In this paper I'm stumped on the proof of Lemma 5 (page 4).

Not the whole lemma, just the part that says if $\mathbf{x}\in\{0,1\}^{n}$, such that $\mathbf{p\cdot x} = p$ and $\mathbf{q\cdot x} = q$, for $p,q \in [0,1]$ then using Lagrangian multipliers, together with the fact that $D$ is convex, it can be shown that $D(\mathbf{p}\|\mathbf{q})$ is minimised when:

(1) $p_{i}$ is the same for all $i$ with $x_{i}=1$

(2) $q_{i}$ is the same for all $i$ with $x_{i}=1$

(3) $p_{i}$ is the same for all $i$ with $x_{i}=0$

(4) $q_{i}$ is the same for all $i$ with $x_{i}=0$

When I use Lagrange multipliers to minimise $D(\mathbf{p}\|\mathbf{q})$ I end up with $\mathbf{p}=\mathbf{q}$, as expected, but $\mathbf{p}$ can't equal $\mathbf{q}$, otherwise it's trivial and I'm not sure how to get those 4 conditions above.

EDIT: Here's my working so far (and I think I made some progress):

$$L = \sum_{i}p_{i}\log{\frac{pi}{qi}} + \lambda (p - \sum_{i} p_{i}x_{i} ) + \nu (q - \sum_{i} q_{i}x_{i})$$

$$\nabla_{p_{i}}L = 1 + \log{\frac{p_{i}}{q_{i}}} - \lambda x_{i}= 0$$

$$\nabla_{\lambda}L = \sum_{i} p_{i}x_{i} - p = 0$$

$$\nabla_{\nu}L = \sum_{i} q_{i}x_{i} - q = 0$$

Which gives me $$p_{i} = q_{i}e^{\lambda - 1}$$

for $x_{i} = 1$. And

$$p_{i} = q_{i}e^{- 1}$$

for $x_{i} = 0$.

Let $x_{i} = 1$, for $i = \{1,...,d\}$:

Then

$$\sum_{i = 1}^{d} p_{i} = p$$

and we can let $p_{i} = \frac{p}{d}$. Similarly we can have $q_{i} = \frac{q}{d}$.

Let $x_{i} = 0$, for $i = \{d+1,...,n\}$:

Then

$$\sum_{i = d+1}^{n} p_{i} = 1 -p$$

and we have $p_{i} = \frac{1-p}{n-d}$. Similarly we have $q_{i} = \frac{1-q}{n-d}$.

But I'm still not clear where it comes out that the $p_{i}$'s must be the same for $i$ with $x_{i} = 1$ etc.

  • 0
    can you show your work?2017-01-12
  • 0
    Ya, it'd be nice to see what you've done already.2017-01-13
  • 0
    Sorry yes, I've edited my post.2017-01-13

0 Answers 0