3

Let $$P(x,y) = Ax^2 + 2Bxy + Cy^2 \\A \neq 0 \quad x,y \in \mathbb R.$$ Without using the second derivative test, prove that

  1. If $AC - B^2 > 0$, then
    (i) $P$ has no zeroes outside the origin and
    (ii) $\operatorname{sgn}(P) = \operatorname{sgn}(A)$ for all $x,y$
  2. If $AC - B^2 = 0$, then
    (i) $P$ has zeroes outside the origin and
    (ii) outside these zeroes, $\operatorname{sgn}(P) = \operatorname{sgn}(A)$
  3. If $AC - B^2 < 0$, then
    (i) $P$ has zeroes outside the origin
    (ii) there exist $x,y$ such that $\operatorname{sgn}(P) = \operatorname{sgn}(A)$ and
    (iii) there also exist $x,y$ such that $\operatorname{sgn}(P) = -\operatorname{sgn}(A)$

Please note:

  1. The second derivative test of calculus is not available, as this question is intended as a lemma for proving the second derivative test
  2. This post is related but distinct

My proof is below: I request verification, critique, or alternate proofs.


Let $$f(x,y) = (x -\frac B A y)^2 \\g(y) = \left(\frac y A \right)^2.$$ Then $$P(x,y) = A \cdot [f(x,y) + (AC-B^2)g(y)]\\ \operatorname{sgn}(P(x,y)) = \operatorname{sgn}(A) \cdot \operatorname{sgn}[f(x,y) + (AC-B^2)g(y)].$$

Observe that the range of both $f$ and $g$ is $[0, \infty)$.

Case 1: If $AC - B^2 > 0$, then $\operatorname{sgn}(P) = \operatorname{sgn}(A)$ unless both $f$ and $g$ are zero. But $g(y) = 0$ implies $y = 0$, and $f(x,0) = 0$ implies $x = 0$, so this can only happen at the origin.

Case 2: If $AC - B^2 = 0$, then $\operatorname{sgn}(P) = \operatorname{sgn}(A)$ unless $f$ is zero, which happens for infinite values of $x, y$.

Case 3: If $AC - B^2 < 0$, then for $|x| \gg |y|$, the first term dominates, and $\operatorname{sgn}(P) = \operatorname{sgn}(A)$, and for $|y| \gg |x|$, the second term dominates, and $\operatorname{sgn}(P) = -\operatorname{sgn}(A)$. Furthermore, for any fixed $y$, then varying $x$ causes $f$ to take on its entire range, so there exist infinite values of $x,y$ which make $P$ zero.

SRobertJames
  • 4,278
  • 1
  • 11
  • 27

1 Answers1

3

As an alternative, we could proceed as follows, for $y\neq 0$ by $t=\frac x y$

$$P(x,y)=Ax^2 + 2Bxy + Cy^2=y^2\left(At^2+2Bt+C\right)=y^2p(t)$$

then the sign of $P(x,y)$ is equal to the sign of $p(t)$ which can be studied by its discriminant $4B^2-4AC$.

The case $y=0$ can be considered a part.

user
  • 154,566
  • Outstanding! I think it should be discriminant (not determinant). How did you think of setting $t = \frac x y$? Or, more directly: How could I realize to do something like that? – SRobertJames Jan 18 '24 at 22:39
  • @SRobertJames Yes of course! Thanks I fix that. This is exactly the way the second derivative test is explained in calcus courses using the quadratic equation, so it seems very natural to me. See for example this reference https://ocw.mit.edu/courses/18-02-multivariable-calculus-fall-2007/resources/lec_week4/ – user Jan 18 '24 at 22:41
  • 1
    When you study a homogeneous function, it is natural to dehomogenize by evaluating $f(x,y)$ at either $(1,y/x)$ or $(x/y,1)$. – Ted Shifrin Jan 18 '24 at 22:53
  • 1
    Follow up question at https://math.stackexchange.com/questions/4847344/when-and-how-do-we-dehomogenize-a-homogeneous-function – SRobertJames Jan 18 '24 at 23:34
  • Maybe @TedShifrin can give some extended exaplanation on that! – user Jan 18 '24 at 23:37