I want to take the inverse of a dot product, where both vectors have complex components. In other words, if $\textbf{A} \cdot \textbf{B} = d$, and I know $\textbf{A}$ and $d$, I want to find a $\textbf{B}$. I know that I cannot do so uniquely, which is fine; I have a procedure for creating a set of vectors that will satisfy $\textbf{A} \cdot \textbf{B} = d$. But, it relies on finding vectors that are orthogonal to $\textbf{A}$ and each other. Normally this would not be a problem; just take $\textbf{A}$, zero all but two of its components, switch the last two, and negate one of them. Make sure it's a different pair every time. For more orthogonal vectors, you can take cross products.

The difficulties arise when I consider vectors with complex components. I want to normalize each of the orthogonal vectors, which means divide by their magnitudes. I have read that you need to divide by $\sqrt{|v_x|^2 + |v_y|^2 + |v_z|^2 ...}$. What is the justification for this? Since I am just normalizing vectors, do I absolutely have to take the magnitudes of the components? Also, My technique of finding the inverse dot product relies on the identity $\textbf{A} \cdot \textbf{B} = d = |\textbf{A}||\textbf{B}|$. What modifications might I need to make? I can post more details if people want. Also, if anyone has links to stuff to read about, especially with regards to the reasons why, I would be most appreciative.


I wouldn't call this the "inverse of a dot product".

Given complex vectors $v = (v_{1}, \ldots, v_{n})$ and $w = (w_{1}, \ldots, w_{n})$, their scalar (=dot) product is given by $v \cdot w = \sum_{j=1}^{n} v_{j} \overline{w}_{j} = v_{1} \overline{w}_{1} + \cdots + v_{n} \overline{w}_{n}$. Why? Well, you want to generalize the usual dot product on $\mathbb{R}^{n}$, but you also want that $v \cdot v \geq 0$ for all $v$, and the vector $(i,0,\ldots,0)$ shows that you can't do without the conjugation. You might ask why, I don't care if $v \cdot v \geq 0$ for all $v$. Most people do: the expression $\|v\| = \sqrt{v \cdot v}$ should define a norm and $\|v - w\|$ a metric on $\mathbb{C}^{n}$, and taking square-roots of non-positve numbers (or even complex numbers) simply isn't well-defined. Note that $\|v\| = \sqrt{v \cdot v} = \sqrt{\sum_{j=1}^{n} |v_{j}|^{2}}$.

Having settled this, let $v$ be a non-zero vector. Its orthogonal complement $U = v^{\perp} = \{u \in \mathbb{C}^{n}\,:u \cdot v = 0\}$ is the set of all vectors orthogonal to $v$. Since $U$ is determined by the single linear equation $u_{1}\overline{v}_{1} + \cdots + u_{n}\overline{v}_{n} = 0$, it is an $(n-1)$-dimensional subspace of $\mathbb{C}^{n}$. Finding solutions is easily achieved using Gauss elimination, this will give you vectors $u_{1}, \cdots, u_{n-1}$, which you can make into an orthonormal basis of $U$ using Gram-Schmidt (note that the notation $\langle u, v \rangle = u \cdot v$ is to be understood. The fact that you're working with $\mathbb{C}$ and not with $\mathbb{R}$ is immaterial, just be careful to note that $v \cdot (\lambda w) = \overline{\lambda} (v \cdot w)$, i.e., the dot product is conjugate-linear in the second variable.

Finally, in order to solve the equation $w \cdot v = d$ simply take any $u \in U = v^{\perp}$ and put $w = u + \frac{d}{v \cdot v} v$, and note that $w \cdot v = (u + \frac{d}{v \cdot v}v) \cdot v = (u \cdot v) + \frac{d}{v \cdot v} v \cdot v = 0 + d = d$.


Yes, to take the length of a complex vector you need the squared magnitudes of the components. So the vector $(1+i,1-i)$ has length $\sqrt{|1+i|^2+|1-i|^2}=\sqrt{2+2}=2$

It is still true that the dot product of orthogonal vectors is zero in a complex space. So once you find a $\textbf{B}$ parallel to $\textbf{A}$ with $\textbf{A} \cdot \textbf{B} = d$ you can add any vector orthogonal to $\textbf{A}$ to it and the dot product will not change. But when you take the dot product, remember to take the complex conjugate of the components of $\textbf{B}$


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.