10
$\begingroup$

Given a symmetric bilinear form $f:V\times V \to K$ , where $V$ is a vector space and $K$ is an appropriate field, define the quadratic form $q:V \to K$ as $q(v):= f(v,v)$.

The Polarisation Formula states that $f(x,y) = 1/2\big( q(x+y) - q(x) - q(y)\big)$, which is easily proven.

This means that any symmetric bilinear form $f:V\times V \to K$ is fully determined by the values $f(v,v)$ for all $v \in V$.

I now want to prove the following theorem: Prove that any symmetric $k$-linear form $M:V\times\cdots \times V \to K$ is determined by the values $M[v]^k := M[v,...,v]$ for all $v\in V$.

How does that work?

$\endgroup$
4
  • $\begingroup$ I'll rephrase my question: Why is any symmetric k-linear form M:Vx...xV -> determined by the values of M[v]^k := M[v,...,v] for all v in V? I am looking for a PROOF. $\endgroup$ Jun 3, 2012 at 15:41
  • 4
    $\begingroup$ This looks like homework. Voting to close until some context is given $\endgroup$
    – Igor Rivin
    Jun 3, 2012 at 16:37
  • $\begingroup$ This is not my homework. I have just completed Linear Algebra 1, which is more basic than the question I posted. I stumbled upon this in a German Analysis book in my free time! I need to understand this because the book uses multilinear forms to define higher order differentials of functions f:R^n -> R. I have never heard of multilinear forms before today and have a hard time dealing with them. For M(x1,---xk) I tried to substitute linear combinations of a basis v1,..,vn but I couldn't simplify it further. Please help me get this right, I am really trying to understand it. Thank you, Feli $\endgroup$ Jun 3, 2012 at 19:54
  • $\begingroup$ Answers to (and the general context of) mathoverflow.net/questions/85213/… might also be somewhat enlightening. $\endgroup$ May 29, 2013 at 9:01

5 Answers 5

8
$\begingroup$

You can be completely explicit in this matter. For $T_j$ in a commutative algebra $$ T_1T_2\dots T_k=\frac{1}{2^k k!}\sum_{\epsilon_j=\pm 1} \epsilon_1\dots\epsilon_k(\epsilon_1T_1 +\dots+\varepsilon_{k}T_{k})^k. $$ The following lemma in available in the Euclidean case.

Lemma. Let $V$ be an Euclidean finite-dimensional vector space, and $A$ a symmetric $k$-multilinear form. We have $ \sup_{\Vert T\Vert=1} \vert{A T^k}\vert =\sup_{\Vert{T_j}\Vert=1} \vert{AT_1\dots T_k}\vert. $

This lemma is a consequence of the 1928 paper by O.D. Kellogg [MR1544896]. This is not true in the non-Euclidean case where the inequality $$ \sup_{\Vert T\Vert=1} \vert{A T^k}\vert \le \sup_{\Vert{T_j}\Vert=1} \vert{AT_1\dots T_k}\vert\le \kappa_k \sup_{\Vert T\Vert=1} \vert{A T^k}\vert, $$ holds true in general with the best constant $ \kappa_{k}= k^k/k!. $

$\endgroup$
3
$\begingroup$

The reason is Schur–Weyl duality. $\newcommand{\GL}{\operatorname{GL}}$

The subspace $W = \langle \forall v \in V \, \mid \, v \otimes v \otimes \cdots \otimes v\rangle$ forms a $\GL(V)$ subrepresentation of $\bigotimes^kV$ if we allow $\GL(V)$ to act diagonally on tensors.

If we consider the dual action, which is the symmetric group $S_k$ permuting tensor factors, we see that all of the generators of $W$ have the symmetry type of the trivial representation of $S_k$ since they are invariant under these permutations. It follows that the $\GL(V)$ subrepresentation $W$ is contained within $\mbox{Sym}^kV \subset \bigotimes^kV$.

However, by Schur-Weyl duality, the symmetric tensors form an irreducible representation of $\GL(V)$ — the subrepresentation $W$ is either $0$ or all of $\mbox{Sym}^kV$.

It isn't $0$, so every symmetric tensor $s$ can be written

$$s = \alpha_0 \cdot v_0 \otimes v_0 \otimes \cdots \otimes v_0 + \alpha_1 \cdot v_1 \otimes v_1 \otimes \cdots \otimes v_1 + \cdots \alpha_l \cdot v_l \otimes v_l \otimes \cdots \otimes v_l$$

for some suitable choice of $v_i$ and $\alpha_i$.

In other words, the $k^{\rm th}$ powers of the elements of $V$ span $\mbox{Sym}^k V$. It follows that knowing a symmetric multilinear form on the $k^{\rm th}$ powers is enough to determine the form.

$\endgroup$
7
  • 4
    $\begingroup$ This could be considered overkill. You can write down a simple formula for $M(x_1,\dots ,x_k)$ as a linear combination of terms of the form $M(x,\dots ,x)$. And it does look like homework. $\endgroup$ Jun 3, 2012 at 17:08
  • $\begingroup$ This is the most general version of polarization that I know. I leave my answer in hopes that it helps a future google searcher. $\endgroup$ Jun 3, 2012 at 17:21
  • 2
    $\begingroup$ Most general? The proof by explicit formulas works over any ring in which k factorial is invertible, and shows that symmetric multilinear functors of $k$ variables correspond bijectively with homogeneous degree $k$ polynomials. $\endgroup$ Jun 3, 2012 at 20:29
  • 1
    $\begingroup$ @Tom Goodwillie By "most general" I meant that any symmetry type is accommodated. Of course, symmetric and exterior powers are familiar, and so the polarization coefficients can be obtained through a little experimentation. But there are other symmetry types corresponding to other irreps of S_k and the corresponding polarizations would be much harder to guess. Continued... $\endgroup$ Jun 4, 2012 at 0:09
  • 1
    $\begingroup$ For example, if I have a linear map from the degree 4 portion of the free Lie algebra on V and I have already specified its values on brackets of the form [v,[w,[v,w]]], is the map determined? If so, how can I write a general [x,[y,[z,t]]] as a linear combination of brackets of this unusual form? $\endgroup$ Jun 4, 2012 at 0:09
2
$\begingroup$

This Wikipedia article contains the Polarization of an algebraic form in general.

$\endgroup$
2
$\begingroup$

The following link contains an explicit formula for bilinear and trilinear forms:

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.53.3481&rep=rep1&type=ps

It is a ps file by Erik G.F. Thomas "A polarization identity for multilinear maps".

A general formula can be constructed in Mathematica by means of the shift operator mentioned in that paper.

I hope it will be useful. It is not so easy to find an explicit formula in the net!

$\endgroup$
1
$\begingroup$

A later version of the paper referenced above now appears on ArXiv https://arxiv.org/abs/1309.1275 edited by TH Koornwinder after Thomas's death. The appendix by Koornwinter attributes the result originally to

S. Mazur and W. Orlicz, Grundlegende Eigenschaften der polynomischen Operationen. Studia Math. 5 (1934) 50–68.

with

J. Bochnak and J. Siciak, Polynomials and multilinear mappings in topological vector spaces. Studia Math. 39 (1971), 59–76.

giving a slightly more general formula. They agree with Bazin's formula above but with 0 and 1 instead of +1 and -1 (and hence without the power of 2).

$$M(v_1,...,v_k) = \frac{1}{k!} \sum\limits_{\epsilon_1,...,\epsilon_k=0}^1 (-1) ^{n- \epsilon_1-\cdots-\epsilon_k} M(\epsilon_1 v_1+ \cdots \epsilon_k v_k)^k $$

$\endgroup$

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.