19
$\begingroup$

$\def\RR{\mathbb{R}}$Let $\omega$ be a $2$-form on $\RR^{2n}$, where $\RR^{2n}$ has the usual Euclidean norm. The comass of $\omega$ is defined to be $\max_{|u|, |v| \leq 1} \omega(u,v)$. Here is the main question:

If $\omega_1$, $\omega_2,\ldots,\omega_n$ are $2$-forms of comass $\leq 1$ on $\RR^{2n}$, what is the maximum possible value of $$|\omega_1 \wedge \omega_2 \wedge \cdots \wedge \omega_n|?$$ In particular, if $\omega_1 = \omega_2 = \cdots = \omega_n = e_1 \wedge e_2 + e_3 \wedge e_4 + \cdots + e_{2n-1} \wedge e_{2n}$, then $\omega_1 \wedge \omega_2 \wedge \cdots \wedge \omega_n = n! (e_1 \wedge \cdots \wedge e_{2n})$, is this $n!$ optimal?

This is a followup to previous posts on Mathoverflow and math.SE; here are observations from those posts (plus a few more new ones):

  1. Define the skew-symmetric matrix $A^k$ by $A^k_{ij} = \omega_k(e_i, e_j)$. Then the comass of $\omega_k$ is the spectral radius (largest norm of an eigenvalue) of $A^k$. Recall that the eigenvalues of a skew symmetric matrix are of the form $\pm \lambda_1 \sqrt{-1}$, $\pm \lambda_2 \sqrt{-1}, \ldots, \pm \lambda_n \sqrt{-1}$, so the condition that the comass is $\leq 1$ is equivalent to $-1 \leq \lambda_1, \lambda_2, \ldots, \lambda_n \leq 1$.

  2. In fact, we may assume that, for each $k$, the eigenvalues of $A^k$ are $\pm \sqrt{-1}$, see David Speyer's comment on the math.SE post. In this case, the condition on $A^k$ is equivalent to saying that $A^k$ is an orthogonal matrix with $(A^k)^2 = - \text{Id}$.

  3. $\omega_1 \wedge \omega_2 \cdots \wedge \omega_n$ is multilinear in the coefficients $A^k_{ij}$ of the skew-symmetric matrices. Specifically, it is the unique symmetric multilinear polynomial $P(A^1, A^2, \ldots, A^n)$ in $n$ skew-symmetrix matrices such that $P(A, A, \ldots, A)$ is $n! \operatorname{Pfaffian}(A)$. If $A$ has eigenvalues $\pm \lambda_1 \sqrt{-1}, \ldots, \pm \lambda_n \sqrt{-1}$, then $\operatorname{Pfaffian}(A) = \pm \lambda_1 \lambda_2 \cdots \lambda_n$, so, in the case $A^1 = A^2 = \cdots = A^n$, equality holds exactly when all the eigenvalues are $\pm \sqrt{-1}$.

  4. Previous posts have proved the $n!$ bound in the cases $n=2$ and $n=3$. There is also an easy $n^n$ bound; see David Speyer's comment on the previous MO post.

  5. A speculation from David Speyer: If $A$ is a skew symmetric matrix with eigenvalues $\pm \lambda_1 \sqrt{-1}, \ldots, \pm \lambda_n \sqrt{-1}$, then $\sum_{i<j} A_{ij}^2 = \sum \lambda_i^2$. I wonder whether the $n!$ bound might continue to be correct if all that we assume about $A^1, A^2, \ldots, A^n$ is that $\sum_{i<j} (A^k_{ij})^2 \leq n$. (Note that $\sum_{i=1}^n \lambda_i^2 \leq n$ implies $\prod \lambda_i \leq 1$ by AM-GM, so the case $A^1 = A^2 = \cdots = A^n$ still works with this weaker hypothesis.) The case of $n=2$ and two distinct two-forms also works.

A related article appeared here.

$\endgroup$
14
  • 4
    $\begingroup$ Would you mind if I edited this to include the background from the previous questions? While at it, I'd also really like to tweak the notation in two ways: (1) Ask about $n$ two-forms, so as not to suggest an infinite sequence of questions :) and (2) use the notation $\alpha_1$, $\alpha_2$, ..., $\alpha_n$ rather than $\alpha$, $\beta$, $\gamma$, ... since the alphabetic notation is a pain to engage with. (EG: Imagine writing "let the eigenvalues of $\alpha$ be $(\rho_1, \rho_2, \dots, \rho_{2n})$, the eigenvalues of $\beta$ be $(\sigma_1, \sigma_2, \dots, \sigma_{2n})$ etcetera ...") $\endgroup$ Commented Jun 13, 2023 at 13:59
  • $\begingroup$ @DavidESpeyer I would much appreciate it, go ahead. $\endgroup$ Commented Jun 13, 2023 at 14:00
  • 1
    $\begingroup$ Okay, edits made. As you can tell, I like this question! $\endgroup$ Commented Jun 13, 2023 at 14:18
  • 2
    $\begingroup$ I'll make one more comment. Here is a variant of the problem which doesn't mention skew-symmetry and which might already be known. Let $D(A^1, A^2, \ldots, A^m)$ be the unique symmetric multilinear polynomial in $m$ many $m \times m$ matrices such that $D(A, A, \dots, A) = \det(A)$. Suppose that each $A^k$ has operator norm $\leq 1$. Can we conclude that $|D(A^1, A^2, \ldots, A^m)| \leq 1$? $\endgroup$ Commented Jun 13, 2023 at 14:59
  • 1
    $\begingroup$ I will be utterly astonished if there is a larger maximum than the n! as in the question. $\endgroup$ Commented Jun 15, 2023 at 21:59

3 Answers 3

14
+150
$\begingroup$

Yes! Here is a proof.

Define the norm $|\omega|$, for a $p$-form on $\mathbb R^m$, to be the least upper bound of $|\omega(v_1,\dots ,v_p)|$ for vectors $v_i$ of length one. For any $p$ and $q$ we can ask for the smallest number $C$ such that for every $p$-form $\alpha$ and every $q$-form $\beta$ on $\mathbb R^m$ we have $|\alpha\wedge \beta|\le C|\alpha||\beta|$. Call it $C_{p,q}$. It does not depend on $m$, because if $C$ is valid for forms in $\mathbb R^{p+q}$ and if $v_1,\dots,v_{p+q}$ span $V\cong \mathbb R^{p+q}$ in $\mathbb R^m$ then $$ |(\alpha\wedge \beta)(v_1,\dots ,v_{p+q})|\le C|\alpha_V||\beta_V||v_1|\dots |v_{p+q}|\le C|\alpha||\beta||v_1|\dots |v_{p+q}|,$$ where $\alpha_V$ and $\beta_V$ are the restrictions to $V$.

I claim that $C_{2,2n-2}\le n$. This implies that for $2$-forms $\omega_1,\dots ,\omega_n$ we have $|\omega_1\wedge \dots \wedge \omega_n|\le n|\omega_1\wedge \dots \wedge \omega_{n-1}||\omega_n|$, and therefore by induction on $n$ we have $|\omega_1\wedge \dots \wedge \omega_n|\le n!|\omega_1|\dots |\omega_n|$.

To prove the claim, first observe that $C_{p,q}$ has this other interpretation: it is the smallest $C$ such that the inner product of two $p$-forms on $\mathbb R^{p+q}$ always satisfies $\alpha\cdot \beta\le C|\alpha||\beta|$. This is so because the norm satisfies $|\beta|=|\ast\beta|$ (Hodge star) and $|\alpha\wedge \ast\beta|=|\alpha\cdot\beta|$.

So a restatement of the claim to be proved is that for $2$-forms on $\mathbb R^{2n}$ we have $\alpha\cdot\beta\le n|\alpha||\beta|$. This can be seen by making a change of orthogonal basis so that $$\alpha = a_1e_1\wedge e_2+\dots a_ne_{2n-1}\wedge e_{2n}$$ for some $a_1,\dots ,a_n$. Write $$\beta=b_1e_1\wedge e_2+\dots b_ne_{2n-1}\wedge e_{2n}+\dots ,$$ where the other terms will not matter. Then $$|\alpha|=max|a_i|$$ $$|\beta|\ge max |b_i|$$ $$\alpha\cdot\beta=\sum_i a_ib_i\le n\ max|a_ib_i|\le n|\alpha||\beta|$$.

I did not say why this norm is invariant under Hodge star. Suppose that $\beta$ is a $p$-form in $\mathbb R^{p+q}$. The claim is that $|(\ast \beta)(v_1,\dots,v_q)|\le |\beta||v_1|\dots |v_q|$. Without loss of generality the vectors $v_i$ are orthogonal and of length one. Complete them to an orthonormal basis by vectors $w_1,\dots ,w_p$. Then $$ |(\ast \beta)(v_1,\dots,v_q)|= |\beta(w_1,\dots ,w_p)|=|\beta|.$$

$\endgroup$
1
  • 3
    $\begingroup$ Nice answer, this one might go in THE BOOK! $\endgroup$
    – Ian Agol
    Commented Jun 18, 2023 at 15:18
5
$\begingroup$

This is an answer to the point/question 5, suggested by David Speyer (as well as a strenghtening of the main inequality, see the Edit below). Consider the norm $$\|\omega\|=\left(\sum_1^n\lambda_j^2\right)^{\frac12}=\sqrt{-{\rm Tr} A^2}.$$ This is a Euclidian norm over the space of skew-symmetric matrices. Then one applies the van der Korput-Schaake inequality (actually discovered by Szegö, and probably known to Banach, see the MO answer). It says that if $P$ is a symmetric $N$-linear form over a Euclidian space $X$, and if $$\forall x\in X,\qquad |P(x,\ldots,x)|\le c\|x\|^N,$$ then $$\forall x_1,\ldots,x_N\in X,\qquad |P(x_1,\ldots,x_N)|\le c\prod_1^N\|x_k\|.$$ In the present case, we have $$|\omega\wedge\cdots\wedge\omega|=n!|\lambda_1\cdots\lambda_n|\le c_n\|\omega\|^n,\qquad c_n=n!n^{-n/2}.$$ We deduce therefore $$|\omega_1\wedge\cdots\wedge\omega_n|\le n!n^{-n/2}\prod_1^n\|\omega_k\|.$$

Edit. As remarked by Mikhail, this can be combined with the obvious inequality $$\|\omega\|\le n^{1/2}\max_j|\lambda_j|,$$ to recover the main inequality that when the comass of each $\omega_k$ is $\le1$, then $$|\omega_1\wedge\cdots\wedge\omega_n|\le n!.$$

$\endgroup$
2
  • 1
    $\begingroup$ Can this be considered a strengthening of the inequality for 2-forms of comass 1, seeing that $\max \lambda_j\geq \frac{\|\omega\|}{\sqrt{n}}$ ? $\endgroup$ Commented Jun 19, 2023 at 8:26
  • 1
    $\begingroup$ @MikhailKatz It seems so. $\endgroup$ Commented Jun 19, 2023 at 9:17
4
$\begingroup$

I was informed by a colleague that there is also a solution to the Pfaffian formulation of the problem, in Lemma 2.1 of the following article:

Roos, Bero, On Bobkov’s approximate de Finetti representation via approximation of permanents of complex rectangular matrices, Proc. Am. Math. Soc. 143, No. 4, 1785-1796 (2015). ZBL1345.28002.

$\endgroup$
3
  • 1
    $\begingroup$ You can use "insert citation" button to find well-formatted citation. $\endgroup$
    – C.F.G
    Commented Jun 18, 2023 at 16:17
  • $\begingroup$ And according to Roos, the proof is due to Hörmander in his paper "On a theorem of Grace", where it appears to be Theorem 4. $\endgroup$
    – Deane Yang
    Commented Jun 18, 2023 at 16:27
  • 3
    $\begingroup$ @DeaneYang In the real case, Roos points out that the proof is due already to Banach. $\endgroup$ Commented Jun 18, 2023 at 16:29

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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