11
$\begingroup$

Is there a generalization of the Cauchy-Schwarz inequality along the following lines? Let $V$ be an inner product space (for simplicity of notation, let us work over the real numbers). Let $v_1, \ldots, v_n$ be in $V$. Let $G$ denote the Gram matrix of the $v_i$, namely, $G$ consists of all possible $(v_i, v_j)$, as $i,j = 1, \ldots, n$, where $(-,-)$ is the inner product in $V$. The usual Cauchy-Schwarz inequality, with $n=2$, can be written as follows, to get rid of square roots:

$$ \det(G) = (v_1,v_1)(v_2,v_2) - (v_1,v_2)^2 \geq 0, $$

with equality iff $v_1$ and $v_2$ both belong to some $1$-dimensional subspace of $V$. So in this case, for $n=2$, the LHS is a homogeneous polynomial in $G$ of degree $2$, and equality is achieved iff $v_1$ and $v_2$ both belong to some $1$-dimensional subspace.

For the general $n$ case, is there a higher degree homogeneous polynomial in $G$ which is non-negative for any $v_1, \ldots, v_n$ in $V$, and which vanishes iff the $v_i$, for $i = 1,\ldots, n$ all lie in some $1$-dimensional subspace of $V$?

(I suspect there may be such a polynomial of degree $2 \lfloor \frac{n(n+1)}{4} \rfloor$. So for instance, if $n=2$, the expected degree is $2$. If $n=3$, the expected degree is $6$, and so on.)

$\endgroup$
  • 1
    $\begingroup$ Indeed the Gramian is positive semi-definite, so its determinant is always nonnegative, and is positive just when the vectors are linearly independent. See en.wikipedia.org/wiki/Gramian_matrix#Gram_determinant $\endgroup$ – Kevin Casto Jul 31 at 19:35
  • $\begingroup$ @KevinCasto, yes but, the determinant of the Gramian vanishes iff the vectors are linearly dependent. What I would like is though, a polynomial which vanishes iff the vectors lie in the same $1$-dimensional subspace. $\endgroup$ – Malkoun Jul 31 at 19:49
  • 6
    $\begingroup$ Actually, there's an octic polynomial: $$Q(v_1,v_2,\ldots,v_n) = \sum_{1\le i<j\le n} ((v_i,v_i)(v_j,v_j)-(v_i,v_j)^2)^2.$$ $\endgroup$ – Robert Bryant Jul 31 at 19:52
  • $\begingroup$ @RobertBryant, ah yes true! The famous sum of squares trick, when working over $\mathbb{R}$. Thank you. How can one obtain all such polynomials? Can one use some form of the positivstellensatz perhaps? $\endgroup$ – Malkoun Jul 31 at 19:55
  • $\begingroup$ @RobertBryant, could you please write it as an answer? The answer turned out to be simple (and I should have thought about it), but it is guiding me in the right direction (for the problem I am interested in, which inspired this post). Is the post too trivial? Should I delete it? $\endgroup$ – Malkoun Jul 31 at 20:02
16
$\begingroup$

Yes, because the OP stated that the ground field is $\mathbb{R}$, one can simply take the octic polynomial $$ Q(v_1,v_2,\ldots,v_n) = \sum_{1\le i < j\le n} \bigl((v_i,v_i)(v_j,v_j)-(v_i,v_j)^2\bigr)^2, $$ which will do the trick.

| cite | improve this answer | |
$\endgroup$
  • 7
    $\begingroup$ Come to think about it, are the squares necessary? We already know that the expressions inside parentheses are nonnegative. $\endgroup$ – Malkoun Jul 31 at 20:28
  • 1
    $\begingroup$ @Malkoun: Yes, you are absolutely correct. I didn't think of that, but, yes, simply the sum of the terms would be enough. Of course, this implies that this construction would work for any ordered field. $\endgroup$ – Robert Bryant Jul 31 at 21:28
  • 7
    $\begingroup$ If one removes the square, then this (now quartic) polynomial also has a geometric interpretation: it is the trace of the exterior square $\bigwedge^2(T^* T)$ of the square $T^* T$ of the linear transformation $T: {\bf R}^n \to V$ that maps the standard basis to $v_1,\dots,v_n$. (It is also the square of the Frobenius norm of $\bigwedge^2 T$.) $\endgroup$ – Terry Tao Aug 1 at 3:24
  • 1
    $\begingroup$ I guess another way to express what Terry is saying is that the identity $$ (v,v)(w,w) = (v,w)^2 + |v\wedge w|^2 = (v,w)^2+(v{\wedge}w,v{\wedge}w)$$ (with the natural inner product on $\Lambda^2(V)$) already shows that $(v,v)(w,w) - (v,w)^2$ is a sum of squares anyway, so the natural quartic polynomial would be $$P(v_1,\ldots,v_n) = \sum_{1\le i < j\le n} |v_i\wedge v_j|^2.$$ In particular, it is expressed as a sum of squares. $\endgroup$ – Robert Bryant Aug 1 at 23:20

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.