Let $F:\mathbb{R}^n\rightarrow \mathbb{R}$ a scalar field which has a quadratic from,

$$F(\mathbf{x}) = \frac{1}{2}\mathbf{d}(\mathbf{x})^\top\Lambda\mathbf{d}(\mathbf{x})$$ with $\mathbf{d}:\mathbb{R}^n\rightarrow\mathbb{R}^m$ being a twice differentiable vector field, and $\Lambda\in \mathbb{R}^{m\times m}$ being a symmetric, positive semi-definite matrix. Thus, $F$ is the objective function in generalized least square problems such as Gauss-Newton. Let $\mathtt{J}_\mathbf{d}$ be the Jacobian of $\mathbf{d}$ and $\mathtt{H}_\mathbf{d}$ its second derivative, the $n\times m \times m$ Hessian tensor.

Due to the product rule, the first derivative of $F$ becomes: \begin{equation} \nabla F = \frac{1}{2}( \mathbf{d}(\mathbf{x})^{\top} \Lambda \mathtt{J}_ \mathbf{d}(\mathbf{x}) )^\top + \frac{1}{2}( \mathtt{J}_ \mathbf{d} (\mathbf{x})^{\top} \Lambda \mathbf{d}(\mathbf{x}) = \mathtt{J}_\mathbf{d}(\mathbf{x})^{\top} \Lambda \mathbf{d}(\mathbf{x}), \end{equation} using the fact that $\Lambda$ is symmetric. Again by means of the product rule, the second derivation or Hessian of $F$ is \begin{equation} \mathtt{H}_F (\mathbf{x}) = \mathtt{J} _\mathbf{d}(\mathbf{x})^\top \Lambda \mathtt{J} _ \mathbf{b}(\mathbf{x}) + \mathtt{H}_\mathbf{d}(\mathbf{x}) \Lambda \mathbf{d}(\mathbf{x}), \end{equation} right?

Question: Is $\mathtt{H}_F$ is positive semi-definite?


It isn't (you have proven it yourself). Take $n=m=1$, $d(x) = 1-x^2$ and $\Lambda =1$. Then we have $$F(x) = d(x)^2/2 = (1-x^2)^2/2$$ and the Hessian is given by $$H_F(x) = \frac{d^2}{dx^2} F(x) = -2x ( 1-x^2)$$ which is smaller than 0 for $0<x<1$ and $x<-1$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.