5
$\begingroup$

Introduction: Let's assume we have a 2-form $\alpha=(1/2)\sum_{j,k=1}^n a_{jk}\ e_j\wedge e_k$, where $n=2m$, and $a_{jk}\in\mathbb C$. We know that $\alpha^{\wedge m}=\alpha\wedge\alpha\dots\wedge\alpha\ (m\ {\rm times}) = n!\ ({\rm Pf}\ A)\cdot e_1\wedge\dots\wedge e_n$, where ${\rm Pf}\ A$ is the Pfaffian of the matrix $A$ whose elements are $a_{jk}$. In other words, if we need to calculate $(\alpha^{\wedge m},I)$, where $(\cdot,\cdot)$ is the scalar product and $I=e_1\wedge\dots\wedge e_n$, then we can easily do that: it is $n!\ ({\rm Pf}\ A)$, and the Pfaffian of a matrix $A$ is easily computed numerically, in polynomial time, for large $n$.

Question: Now, I have two 2-forms: one is $\alpha$ defined above, and the other is $\beta=(1/2)\sum_{j,k=1}^n b_{jk}\ e_j\wedge e_k$. I want to calculate $\alpha^{\wedge p}\wedge\beta^{\wedge m-p}$ for some $p$. What will be the coefficient in front of $I=e_1\wedge\dots\wedge e_n$? Can it be computed efficiently, say polynomially in $n$? In other words, how to numerically calculate $(\alpha^{\wedge p}\wedge\beta^{\wedge m-p},I)$ for some general skew-symmetric complex-valued matrices $A$ and $B$? What if I have not two, but $m$ 2-forms $\alpha_1,\dots,\alpha_m$, determined by the complex-valued matrices $A_1\dots A_m$, and we want to calculate ($\alpha_1\wedge\alpha_2\dots\wedge\alpha_m,I)$? Is it something easily computable, like Pfaffian of a matrix in the introduction above? In general, the answer is no, but what if the matrices $A_k$ are of low rank, say rank 4? (the case of rank 2 is trivial, since each form becomes a wedge product of two vectors)

Did someone study this, or saw some discussions about similar or related problems? Where can I try to find the information about such problems? I would appreciate any suggestion. I am not even sure what is the commonly used term for this coefficient.

What I found so far: in "Multilinear algebra" of W. Greub, the Pfaffian of skew-linear transformations is introduced (Ch. 8.4, Springer edition 1978) for several transformations. I.e. Greub starts as follows: $E$ is an inner product space of dimension $m=2n$, and $\phi_1, \phi_2,\dots \phi_m$ are the skew-linear transformations. Each $\phi_k$ determines an element $\Psi_k\in \bigwedge^2 E$, so that $\Psi_1\wedge\Psi_2\wedge\dots\Psi_m\in \bigwedge^n E$. Taking a basis vector $a\in E$, he defines ${\rm Pf}_a(\phi_1\dots\phi_m) = (\Psi_1\wedge\dots\wedge\Psi_m, a)$. This is exactly what I need. Unfortunately, already in the next paragraph he switches to the standard case $\phi_1=\phi_2=\dots=\phi_m=\phi$, which gives the standard answer for the Pfaffian of a matrix for $\Psi$, and does not discuss the case of different $\phi$'s anymore.

I would assume that people have studied this problem (or this is something simple, and I am just missing something obvious). I would be very grateful for help!

$\endgroup$
1
$\begingroup$

For two 2-forms I can offer you a blueprint of an algorithm. Details would have to be filled in.

Consider two skewsymmetric $n\times n$-matrices $A$ and $B$. The task is to compute $\operatorname{Pf}(sA+tB)$ since this gives all $\alpha^p\wedge\beta^{m-p}$ simultaneously.

Let's assume first that $A$ is invertible. Then there is a skew-analogue of Gram-Schmidt (first detail to be filled in) which brings $A$ into some sort of diagonal form $J$. Let $Sp(n)$ be the symplectic group with respect to $J$.

Now transform $B$ using a symplectic base change to bring it into an analogue of a companion matrix (second detail to be filled in). This is possible since the $Sp(n)$ action on skewsymmetric matrices is a $\theta$-representation in the sense of Vinberg.

Now $\text{Pf}(sA+tB)$ is readily computable.

If $A$ is not invertible, change it to $A+cB$ for a suitable constant $c$. If no such $c$ exists, the Pfaffian is identically zero.

$\endgroup$
  • $\begingroup$ First of all, thank you! The approach you suggested looks very interesting and thought-provoking; maybe it can even be extended to the case of many matrices $A_1\dots A_m$. Now, about the details to be filled in: the first one I clear, it is not difficult to do. But I am a little puzzled by the second one, and I would greatly appreciate if you could elucidate it a little. What do you mean by an "analogue of the companion matrix"? Companion matrix to which polynomial? Did you mean just the convenient form that the companion matrix has, or do I miss something in your answer? $\endgroup$ – BarTov May 16 '16 at 3:22
  • $\begingroup$ What I meant was "rational canonical form" aka. "Frobenius normal form". The polynomial $\text{Pf}(sJ+B)$ is a skew-symmetric analogue of the characteristic polynomial. Maybe we don't have to reinvent the wheel.I googled for "ratonal canonical form" and "skew-symmetric matrix" and stumbled across the paper "Skew-Symmetric Matrix Polynomials and their Smith Forms" by Mackey, Mackey, Mehly, and Mehrmann. The proof of Thm. 5.4 seems to give the algorithm you are looking for. $\endgroup$ – Friedrich Knop May 16 '16 at 6:45
  • $\begingroup$ Finally, no, I don't think that this generalizes to more than two $2$-forms. The invariant theory of $GL(n)$ acting on three or more copies of $\wedge^2\mathbb{C}^n$ is too bad. $\endgroup$ – Friedrich Knop May 16 '16 at 6:49
0
$\begingroup$

Since the two forms in the definition of "multilinear Pfaffian" commute one can use polarization and write down explicit formula. See e.g. (7) of https://arxiv.org/abs/1309.1275 This formula, however, has $2^m$ terms.

$\endgroup$

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.