MHB Q(x) is a strictly convex function, show that G is positive definite

numbersense
Messages
4
Reaction score
0
Consider the quadratic function $\displaystyle q(\textbf{x}) = \frac{1}{2} \textbf{x}^T G \textbf{x} + \textbf{d}^T \textbf{x} + c$ on $\mathbb{R}^n$, where $\textbf{d}$ is a constant $n \times 1$ vector, $G$ is a constant $n \times n$ symmetric matrix and $c$ is a scalar.

The gradient is $\nabla q(\textbf{x}) = G \textbf{x} + \textbf{d}$ and the Hessian is $\nabla^2 q(\textbf{x}) = G$.

If $q(\textbf{x})$ is a strictly convex function then show that $G$ is positive definite.I am not sure whether I should start with the convex function definition or start by considering the gradient or the Hessian.

I tried expanding the inequality in the convex function definition but didn't get anywhere.

There is a proposition that says $f$ is strictly convext on $\mathbb{R}^n$ $\implies$ any stationary point is the unique global minimizer. (I can't even prove that a stationary point exists) Another theorem says that positive definiteness is a sufficient condition for being a unique global minimizer and positive semi definiteness is a necessary condition for being a local minimizer. I can't see how to use these statements to prove what the question is asking.
 
Mathematics news on Phys.org
numbersense said:
Consider the quadratic function $\displaystyle q(\textbf{x}) = \frac{1}{2} \textbf{x}^T G \textbf{x} + \textbf{d}^T \textbf{x} + c$ on $\mathbb{R}^n$, where $\textbf{d}$ is a constant $n \times 1$ vector, $G$ is a constant $n \times n$ symmetric matrix and $c$ is a scalar.

The gradient is $\nabla q(\textbf{x}) = G \textbf{x} + \textbf{d}$ and the Hessian is $\nabla^2 q(\textbf{x}) = G$.

If $q(\textbf{x})$ is a strictly convex function then show that $G$ is positive definite.I am not sure whether I should start with the convex function definition or start by considering the gradient or the Hessian.

I tried expanding the inequality in the convex function definition but didn't get anywhere.

There is a proposition that says $f$ is strictly convext on $\mathbb{R}^n$ $\implies$ any stationary point is the unique global minimizer. (I can't even prove that a stationary point exists) Another theorem says that positive definiteness is a sufficient condition for being a unique global minimizer and positive semi definiteness is a necessary condition for being a local minimizer. I can't see how to use these statements to prove what the question is asking.
I don't think you need to use the gradient or the Hessian to show that $G$ is positive definite. The fact that $q$ is strictly convex tells you that $q(\textbf{0}) <\frac12\bigl(q(\textbf{x}) + q(-\textbf{x})\bigr)$, for any nonzero vector $\textbf{x}.$ It follows very easily that $0 < \textbf{x}^T G \textbf{x}$.
 
Opalg said:
I don't think you need to use the gradient or the Hessian to show that $G$ is positive definite. The fact that $q$ is strictly convex tells you that $q(\textbf{0}) <\frac12\bigl(q(\textbf{x}) + q(-\textbf{x})\bigr)$, for any nonzero vector $\textbf{x}.$ It follows very easily that $0 < \textbf{x}^T G \textbf{x}$.

Thanks. It follows like this.
\begin{align*}
q\left(\frac{1}{2} \textbf{x} + \left(1 - \frac{1}{2}\right) (-\textbf{x})\right) = q(0) = c <& \frac{1}{2} q(\textbf{x}) + \left(1 - \frac{1}{2}\right) q(-\textbf{x}) = \textbf{x}^T G \textbf{x} + c\\
\textbf{x}^T G\textbf{x} >& 0
\end{align*}

This is quite clever. I didn't think of obtaining an inequality by using specific values in the convex function definition.
 
Insights auto threads is broken atm, so I'm manually creating these for new Insight articles. In Dirac’s Principles of Quantum Mechanics published in 1930 he introduced a “convenient notation” he referred to as a “delta function” which he treated as a continuum analog to the discrete Kronecker delta. The Kronecker delta is simply the indexed components of the identity operator in matrix algebra Source: https://www.physicsforums.com/insights/what-exactly-is-diracs-delta-function/ by...
Fermat's Last Theorem has long been one of the most famous mathematical problems, and is now one of the most famous theorems. It simply states that the equation $$ a^n+b^n=c^n $$ has no solutions with positive integers if ##n>2.## It was named after Pierre de Fermat (1607-1665). The problem itself stems from the book Arithmetica by Diophantus of Alexandria. It gained popularity because Fermat noted in his copy "Cubum autem in duos cubos, aut quadratoquadratum in duos quadratoquadratos, et...
I'm interested to know whether the equation $$1 = 2 - \frac{1}{2 - \frac{1}{2 - \cdots}}$$ is true or not. It can be shown easily that if the continued fraction converges, it cannot converge to anything else than 1. It seems that if the continued fraction converges, the convergence is very slow. The apparent slowness of the convergence makes it difficult to estimate the presence of true convergence numerically. At the moment I don't know whether this converges or not.
Back
Top