Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Let $F \colon \mathbb R^2 \to \mathbb R$ be the function $$ F(x,y):=xye^x + ye^y - e^x+1 $$ and denote with $C$ the set of zeroes of $F$, i.e. $C:=\{(x,y) \in \mathbb R^2 : F(x,y)=0\}$. Let also $f \colon \mathbb R^2 \to \mathbb R$ be a function which is $C^2$ in a neighbourhood of $(0,0)$ and such that $$ \nabla f(0,0) = (-2,2), \qquad \qquad H_f(0,0) = \begin{pmatrix} 2 & 1 \\ 1 & 0 \end{pmatrix}. $$ Question: is $(0,0)$ a minimum for $f$ on $C$?

Well, I show you what I have done so far. First of all, some routine calculations yield $$ \nabla F(0,0) = (-1,1) \qquad \qquad H_F(0,0) = \begin{pmatrix} -1 & 1 \\ 1 & 2 \end{pmatrix} $$

Indeed, the fact that $$ \nabla f(0,0) = 2\nabla F(0,0) $$ does agree with the theory of Lagrange multipliers: the gradients are parellel, so I think that $(0,0)$ is an extremum for $f$ on $C$. The problem is how to classify it without any information on $f$: we have only its hessian matrix, which is - I suppose - the key to solve this. Both $H_f(0,0)$ and $H_F(0,0)$ are indefinite.

How can we establish the nature of the critical point $(0,0)$? Thanks in advance for your help.

share|cite|improve this question
You have to figure out how the Hessian looks on the space orthogonal to $\nabla f$. – Fabian Aug 16 '12 at 21:12
@Fabian Thanks for your comment. Why do I have to look a the space orthogonal to $\nabla f$? It is of course the subspace of $\mathbb R^{2}$ generated by $(1,1)$ but I do not know how to use this information. Would you mind explaining me a little bit, please? Thanks. – Romeo Aug 16 '12 at 21:21
up vote 2 down vote accepted

Suppose you have a parametrization $x=X(t), y=Y(t)$ of the curve $C$ near $(0,0)$, with $X(0)=0$ and $Y(0)=0$. Thus if $V$ and $A$ are the corresponding velocity and acceleration vectors, $$\dfrac{d}{dt} f(X(t),Y(t)) = X' f_x(X,Y) + Y' f_y(X,Y) = (\nabla f)\cdot V$$

$$\dfrac{d^2}{dt^2} f(X(t),Y(t)) = X'' f_x(X,Y) + Y'' f_y(X,Y) + (X')^2 f_{xx}(X,Y) + 2 X' Y' f_{xy}(X,Y) + (Y')^2 f_{yy}(X,Y) = (\nabla f)\cdot A + V^T H(f) V $$

Similarly for $f$ replaced by $F$, but in this case $\dfrac{d}{dt} F(X(t),Y(t)) = \dfrac{d^2}{dt^2} F(X(t),Y(t)) = 0$.

Now since $\nabla f(0,0) = 2 \nabla F(0,0)$, $(\nabla F(0,0)) \cdot V(0) = 0$ implies $\dfrac{d}{dt} f(X(t),Y(t))|_{t=0} = (\nabla f(0,0)) \cdot V(0) = 0$. And $$\dfrac{d^2}{dt^2} f(X(t),Y(t))|_{t=0} = \dfrac{d^2}{dt^2} f(X(t),Y(t))|_{t=0} - 2 \dfrac{d^2}{dt^2} F(X(t),Y(t))|_{t=0} = V(0)^T (H(f) - 2 H(F)) V(0)$$

Now $V(0)$ is orthogonal to $\nabla F(0,0) = (-1,1)$, so it is a multiple of $(1,1)$, and $$ (1,1) \left( \pmatrix{2 & 1\cr 1 & 0\cr} - 2 \pmatrix{-1 & 1\cr 1 & 2\cr}\right) \pmatrix{1\cr 1\cr} = -2$$ so the second derivative is negative, and you have a local maximum.

share|cite|improve this answer

In order to find the Hessian, we need to know the change in $f$ to second order in the deviation from $(0,0)$.

Expanding $F(x,y)$ around this point, we have up to second order $$ F(x,y) = \nabla F(0,0)\begin{pmatrix} x \\y\end{pmatrix} + \tfrac12 \begin{pmatrix} x & y\end{pmatrix} H_F \begin{pmatrix} x \\y\end{pmatrix}.$$

Similarly, expanding $f(x,y)$ we obtain $$ f(x,y) = \nabla f(0,0)\begin{pmatrix} x \\y\end{pmatrix} + \tfrac12 \begin{pmatrix} x & y\end{pmatrix} H_f \begin{pmatrix} x \\y\end{pmatrix}.$$

In the last step, we need to implement the constraint $F(x,y)=0$. To first order, we have $\nabla F(0,0) \begin{pmatrix} x \\y\end{pmatrix}=0$ and because $\nabla f \parallel \nabla F$, we have that $f(0,0)$ is an extremum. In particular, we have $y=x$ to first order.

To second order, we have $$\tfrac12 \begin{pmatrix} x & y\end{pmatrix} H_F \begin{pmatrix} x \\y\end{pmatrix} = \tfrac12 \begin{pmatrix} x & x\end{pmatrix} H_F \begin{pmatrix} x \\x\end{pmatrix} = \frac{3}{2} x^2. $$ In order that $F(x,y)=0$ to second order, we need to compensate this via the first term, so we have $y= x - \frac{3}{2} x^2$.

Plugging this into the expansion of $f(x,y)$, we obtain a term from the gradiant and a term from the Hessian (Tunococ just considered the term from the Hessian). In total, we have $$f(x,y) = -3 x^2 + 2 x^2 =-x^2.$$ So it is a local maximum.

share|cite|improve this answer
I don't understand how you came up with the correction factor $-\frac 32 x^2$. With your $y$, you get $y'(0) = -2$. That means you're going in the direction violating the constraint. – Tunococ Aug 16 '12 at 23:07
@Tunococ: $y'(x) =1-3 x$, setting $x=0$, i have $y'(0)=1$. – Fabian Aug 17 '12 at 5:27

Edit: I did not take into account the non-linearity of the constraint so it was wrong. This is the wrong version:

The complementary space of $\nabla F(0, 0)$ is spanned by $(1, 1)$. The (unscaled) reduced Hessian of $f$ is a $1$-by-$1$ matrix: $\begin{pmatrix}1 & 1\end{pmatrix}\begin{pmatrix}2 & 1 \\ 1 & 0\end{pmatrix}\begin{pmatrix}1 \\ 1\end{pmatrix} = 4 > 0$, so it is a minimum.

For correct answer, see Robert's answer above.

share|cite|improve this answer
Thanks for your answer, but I haven't understood. Why are you looking at the quadratic form given by the hessian on the complementary space of $\nabla f(0,0)$? Are my calculations of $\nabla F$ and $H_F$ useless? Thanks a lot. – Romeo Aug 16 '12 at 21:24
@Tunococ: I don't think that this answer is completely correct. To second order you should take into account that the space with $F(x,y)=0$ is curved. – Fabian Aug 16 '12 at 21:41
Sorry I was wrong. I switched $f$ and $F$. I've fixed that. – Tunococ Aug 16 '12 at 22:35

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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