5

Let $P$ and $V$ be vector spaces of dimension $k$ and $n$, respectively. I want to know if it's possible to endow $\text{Hom}(P,V)$ with a natural inner product, but without using the (non-canonical) isomorphism between $V$ and $V^*$ and then considering something like $\langle{\phi},{\psi}\rangle = \text{tr}(\phi^{\top} \circ \psi)$.

What comes to my mind and what is more suitable to the setup I'm working on is something like this: assume that $P$ and $V$ have inner products $\langle \cdot, \cdot \rangle_P$ and $\langle \cdot, \cdot \rangle_V$, respectively. Then, we can identify $P \otimes V$ with $\text{Hom}(P,V)$ using the map $p \otimes v \mapsto \langle \cdot, p \rangle_P v$ and we can consider the canonical inner product on $P \otimes V$ given by $\langle p \otimes v, q \otimes w \rangle_{P \otimes V} := \langle p, q \rangle_P \langle v, w \rangle_P$. But how this translates to the Hom space?

I've tried this: choosing basis $\{p_1,\ldots,p_k\}$ and $\{v_1,\ldots,v_n\}$ (and I can't assume those basis are orthogonal) and defining $G$ as the Gram matrix of $V$, ie, $G = (\langle v_i, v_j \rangle_V)_{i,j}$, I've identified $\text{Hom}(P,V)$ as the space of matrices $n \times k$ written in those basis and then defined the inner product as $\text{tr}(A^{\top}GB)$. But this formula is wrong because on $P \otimes V$ we have $\langle{p_i \otimes v_r},{p_j \otimes v_s}\rangle_{P \otimes V} = \langle{p_i},{p_j}\rangle_P \langle{v_r},{v_s}\rangle_V$, but the inner product $\text{tr}(A^{\top}GB)$ on $\text{Hom}(P,V)$ computed on the corresponding matrices of the linear transformations $A := \langle \cdot,p_i \rangle_P v_r$ and $B := \langle \cdot,p_j \rangle_P v_s$ gives me $$\sum_{\lambda=1}^k \langle p_{\lambda},p_i \rangle_P \langle v_r,v_s \rangle_V \langle p_{\lambda},p_j \rangle_V$$ and this is not what I expected. Any idea how should I proceed?

  • 1
    Does the Cartain-Eilenberg adjunction ${\mathrm Hom}(X\otimes Y,Z)\approx {\mathrm Hom}(X,\mathrm{Hom}(Y,Z))$ give anything that appeals to you? – paul garrett Jun 02 '21 at 21:31
  • @paulgarrett, I don't even know how to endow $\text{Hom}(X,Y)$ with inner product, so, I don't see how adding another layer of Hom could help me. ): – Bias of Priene Jun 02 '21 at 21:35
  • The answer to the question in the first paragraph is certainly "no". Regardless, the rest is a more concrete question. Why do you not want to pick orthonormal bases? They will simplify your calculation. – Thorgott Jun 02 '21 at 22:01
  • @Bias Regarding the second paragraph: the inner product that this induces on the Hom space is $\langle \phi, \psi \rangle = \operatorname{tr}(\phi^\top \circ \psi)$. – Ben Grossmann Jun 02 '21 at 22:42
  • @Thorgott, I can't assume orthogonality because this problem appeared to me in a computational problem and orthogonalizing is kinda expensive, so, I'm trying to come up with analytic formulas without assuming it. (There's also a deeper reason, but in the end it's just the computational aspect that matters) – Bias of Priene Jun 02 '21 at 23:11
  • @BenGrossmann, well, I've computed it in basis and it didn't work, but maybe I've made a mistake. But now I'm convinced I should consider the adjoint map (in the inner product sense) and use the formula given in this answer to achieve what I want (I'll check this out tomorrow): https://math.stackexchange.com/questions/1490851/why-do-we-need-an-orthonormal-basis-to-represent-the-adjoint-of-the-operator – Bias of Priene Jun 02 '21 at 23:15
  • Well, if you just stubbornly spell it out, you should arrive $\langle f,g\rangle=\sum_{i=1}^n\sum_{j=1}^k\sum_{i^{\prime}=1}^n\sum_{j^{\prime}=1}^k\langle f(p_j),v_i\rangle_V\langle g(p_{j^{\prime}}),v_{i^{\prime}}\rangle_V\langle v_i,v_{i^{\prime}}\rangle_V\langle p_j,p_{j^{\prime}}\rangle_P$ for $f,g\in\operatorname{Hom}(P,V)$. Not sure if that is helpful though. – Thorgott Jun 02 '21 at 23:27

1 Answers1

1

$\def\hom{\operatorname{Hom}} \def\tr{\operatorname{tr}}$I will sum up what has already been discussed in the comments. Let $V,W$ be finite-dimensional inner product spaces over $k\in\{\mathbb{R},\mathbb{C}\}$. If $A:V\to W$ is $k$-linear, denote $A^*: W\to V$ to the adjoint operator (it is the unique $k$-linear map such that $\langle Av,w\rangle_W=\langle v,A^*w\rangle_V$ for all $v\in V,w\in W$). The canonical inner product in $\hom(V,W)$ is called the Frobenius inner product and is defined in the following way: for $A,B\in\hom(V,W)$, we set $$ \langle A,B\rangle=\tr(A^*B)=\tr(BA^*), $$ where $\tr$ denotes the trace (although computation of the trace requires picking a basis, the final result can be checked to be independent of the choice of basis).

Suppose $\{e_i\}$ and $\{c_i\}$ are respectively orthonormal bases of $V$ and of $W$, and denote $\rho_{ij}\in\hom(V,W)$ to the unique $k$-linear map such that $\rho_{ij}(e_k)=\delta_{ik}c_j$ (we will use Einstein summation convention). Then $\{\rho_{ij}\}$ is an orthonormal basis of $\hom(V,W)$. If $A=A^{ij}\rho_{ij}$, then $A^{ij}=c^j(Ae_i)$, where $\{c^i\}\subset W^*$ is the dual basis of $\{c_i\}$. Let $A,B\in\hom(V,W)$. Then

\begin{align*} \langle A,B\rangle &=\sum_{ij}(A^{*})^{ij}B^{ij}\\ &=\sum_{ij}\overline{c^j(Ae_i)}c^j(Be_i)\\ &=\sum_{ij}\langle c^j(Ae_i)c_j,c^j(Be_i)c_j\rangle_W\\ &=\sum_{ijk} \langle c^j(Ae_i)c_j,c^k(Be_i)c_k\rangle_W\\ &=\sum_i \langle Ae_i,Be_i\rangle_W, \end{align*} which gives a way of computing the inner product of $\hom(V,W)$ in terms of an orthonormal basis of $V$ and the inner product in $W$.

The generalization to infinite-dimensional Hilbert spaces is the Hilbert-Schmidt inner product.