Equivalence of $Ax = b$ and the quadratic form?

499 Views Asked by At

I want to prove the following:

Given $A \in \mathbf{R}^{n \times n}$ is symmetric positive definite. Prove that $\hat{x}$ solves $Ax = b$ if and only if $\hat{x}$ minimizes the quadratic function $f: \mathbf{R}^n \to \mathbf{R}$ given by:

$$f(x) = \frac{1}{2}x^TAx - x^Tb$$

Attempt:

Since $A$ is positive definite, it is invertible since its eigenvalues are all strictly positive. Let $x = A^{-1}b$ and determine $f(y) - f(x)$ for any $y \in \mathbf{R}^n$. Since $Ax = b$:

$$\begin{align} f(y) - f(x) &= \frac{1}{2}y^TAy - y^Tb - \frac{1}{2}x^TAx + x^Tb \\ &= \frac{1}{2}y^TAy - y^TAx + \frac{1}{2}x^TAx \\ &= \frac{1}{2}(y - x)^TA(y-x)\end{align}$$ Since $A$ is positive definite, the last expression is nonnegative and therefore $f(y) \geq f(x)$ for all $y \in \mathbf{R}^n$, which gives x = $A^{-1}b$ as the global minimum of $f(x)$ and $$f(A^{-1}b) = -\frac{1}{2}b^TA^{-1}b$$

Concerns:

I'm concerned that this proof is determining what the global minimum of the equivalent system is not necessarily that $\hat{x}$ solves $Ax = b$ if and only if $\hat{x}$ minimizes the quadratic function. Any hints in the right direction would be greatly appreciated!

1

There are 1 best solutions below

0
On BEST ANSWER

Because $A$ is invertible, "$\hat{x}$ solves $Ax=b$" is just "$\hat{x} = A^{-1} b$."

So to show the "if and only if" claim, you need to show that $A^{-1}b$ is the unique minimizer of $f$.

So far you have shown that $A^{-1}b$ is a minimizer (since $f(y) \ge f(A^{-1} b)$ for all $y$), but you still need to show it is the only one. To do this, it suffices to adjust your argument to show the stronger inequality $f(y) > f(A^{-1} b)$ for all $y \ne A^{-1} b$. Can you see how to do this? (Hint: $A$ is positive definite.)