Nov 17

Cauchy-Schwarz inequality and optimization 2

Cauchy-Schwarz inequality and optimization 2

Here we consider a problem from mt3042 Optimization Guide by M. Baltovic:

Example 8.1. Given the initial state s_{1}=W for some W>0 and a discount factor \delta\in(0,1), we need to maximize the discounted stream of future rewards

(1) \sum_{i=1}^{\infty }\delta^t\sqrt{a_t}.

The maximization is over actions (a_1,a_2,...) (you can think of them as consumptions). The initial state s_1=W is the initial wealth; at each step, the consumption is chosen between zero and the wealth left from the previous period, 0\leq a_t\leq s_t. This reduces the amount available for consumption in the next period:

(2) s_{t+1}=s_t-a_t, t\geq 1.

Step 1. Here we reformulate the problem to remove the states from the constraints. From (2) we have a_t=s_t-s_{t+1}. Fixing an integer T>0 then we obtain

(3) a_1+a_2+...+a_{T-1}+a_T=s_1-s_2+s_2-s_3+...+s_{T-1}-s_T+s_T-s_{T+1}=s_1-s_{T+1}.

Since in each state s_t consumption a_t is nonnegative, the states can only decrease over time. On the other hand, they cannot be negative. Hence, the sequence \{s_T\} is nonincreasing and bounded from below by zero. Therefore the limit S=\lim_{T\rightarrow\infty}s_T exists and is nonnegative. Letting T\rightarrow\infty in (3) we get \sum_{t=1}^{\infty}a_t=W-S\leq W.

The problem becomes: maximize (1) subject to

(4) \sum_{t=1}^{\infty}a_t\leq W.

Step 2. Getting rid of square roots. Denoting x_t=\sqrt{a_t}, we have to maximize \sum_{i=1}^{\infty }\delta^tx_t subject to \sum_{t=1}^{\infty }x_t^2\leq W.

Step 3. Applying the result on the Cauchy-Schwarz inequality. The result from the previous post does not depend on the number of variables. For our present needs it is formulated as follows: Let a_1,a_2,...\geq 0 be fixed numbers and let x_1,x_2,...\geq 0 be variables satisfying the restriction x_1^2+x_2^2+...\leq W. Then the infinite weighted sum a_1x_1+a_2x_2+... is maximized at the point

(x_1,x_2,...)=\frac{W^{1/2}}{\left( a_1^2+a_2^2+...\right)^{1/2}}(a_1,a_2,...)

and the maximized value is a_1x_1+a_2x_2+...=W^{1/2}\left(a_1^2+a_2^2+...\right)^{1/2}.

Letting a_{t}=\delta^t we get a_1^2+a_2^2+...=1/(1-\delta^2) and

(x_1,x_2,...)=[(1-\delta^2)W]^{1/2}(1,\delta,\delta^2,...) with the maximized value \sqrt{W/(1-\delta^2)}.

And no clumsy terminology and notation from the Infinite Horizon Dynamic Programming method!

Nov 17

Cauchy-Schwarz inequality and optimization 1

Cauchy-Schwarz inequality and optimization 1

Between Cauchy and Schwarz there was also Bunyakovsky who contributed to this inequality.

Inequalities, if they are precise, can be used for optimization, even in the infinite-dimensional case. The method explained here does not rely on the Lagrange method or the Kuhn-Tucker theorem or convexity. It illustrates the point that it's not a good idea to try fit all problems into one framework. The solution method should depend on the problem.

Problem. Let a_1,a_2\geq 0 be fixed numbers and let x_1,x_2\geq 0 be variables satisfying the restriction

(1) x_1^2+x_2^2\leq W.

The problem is to maximize the weighted sum a_1x_1+a_2x_2 subject to the above restriction. This can be done using the convexity method. However, our future application will contain an infinite number of variables in which case the convexity method becomes conceptually complex. Therefore here we provide a different method based on the Cauchy-Schwarz inequality. In our case it looks like this

(2) \left\vert a_1x_1+a_2x_2\right\vert \leq \left(a_1^2+a_2^2\right)^{1/2}\left(x_1^2+x_2^2\right)^{1/2}.

We solve the problem in two steps. First we bound a_1x_1+a_2x_2 from above and then we find the point (x_1,x_2) at which the upper bound is attained.

Step 1. (1) and (2) obviously imply

(3) a_1x_1+a_2x_2\leq\left(a_1^2+a_2^2\right)^{1/2}W^{1/2}.

Step 2. We need to find (x_1,x_2) such that (1) is satisfied and (3) turns into an equality. This involves a little guessing. The guiding idea is that on the right in (3) we have squares of a_1,a_2. For them to appear at the left, we choose x_1=ca_1,x_2=ca_2 with some constant c. Then we want


to be true. This gives the value of the constant c=\frac{W^{1/2}}{\left(a_1^2+a_2^2\right)^{1/2}}. The point we are looking for becomes

(4) (x_1,x_2)=\frac{W^{1/2}}{\left(a_1^2+a_2^2\right)^{1/2}}(a_1,a_2). It satisfies (1):


Further, the maximized value of a_1x_1+a_2x_2 is \left(a_1^2+a_2^2\right)^{1/2}W^{1/2}.