I have two vectors $\mathbf v_1$ and $\mathbf v_2$:

$$\mathbf v_1 = \begin{pmatrix}x_1\\y_1\\z_1\end{pmatrix}, \mathbf v_2 = \begin{pmatrix}x_2\\y_2\\z_2\end{pmatrix}$$

The components of these vectors each differ by at most a constant $\epsilon$:

\begin{align} |x_2 - x_1| &\le \epsilon\\ |y_2 - y_1| &\le \epsilon\\ |z_2 - z_1| &\le \epsilon\\ \end{align}

The angle $\theta$ between these vectors is given by:

\begin{align} \theta = \arccos \frac{\mathbf v_1 \cdot \mathbf v_2}{\left\| \mathbf v_1 \right\| \, \left\| \mathbf v_2 \right\|} \end{align}

How can I derive a bound on the maximum on the angle between vectors $\mathbf v_1$ and $\mathbf v_2$ only based on the constant $\epsilon$?

  • 2
    $\begingroup$ There is no helpful bound as you can achieve any angle using suitably small values of $x_1,x_2,y_1,y_2,z_1,z_2$. It might be different if $\mathbf v_1$ was fixed with $|x_1| + |y_1|+|z_1| \gg \epsilon$ and only $\mathbf v_2$ was variable $\endgroup$
    – Henry
    Sep 1 '15 at 10:02
  • $\begingroup$ @Henry how would the bound look like if $\mathbf v_1$ was fixed as you suggested? $\endgroup$
    – pistermink
    Sep 1 '15 at 10:17
  • $\begingroup$ I suspect there are bounds of the form $\dfrac{k\epsilon}{|\mathbf v_1|}$ perhaps with $k$ slightly above $\sqrt 3$ for $|\mathbf v_1| \gg \epsilon$ $\endgroup$
    – Henry
    Sep 1 '15 at 11:40
  • $\begingroup$ @Henry could you please post this as an answer (with an explanation how you derived this bound) so I can accept this? $\endgroup$
    – pistermink
    Sep 2 '15 at 12:49

As requested in the comments:

I suspect there are bounds of the form $\dfrac{k\epsilon}{|\mathbf{v}_1|}$ perhaps with $k$ slightly above $\sqrt3$ for $|\mathbf{v}_1|\gg \epsilon$.

This comes from empirically trying to maximise the angle by taking a large $|\mathbf{v}_1|$ and putting a $\pm \epsilon$ cube round it and looking at its eight corners.


Assume that $\mathbf{v}_1$ is fixed and $\mathbf{v}_2$ can be varied within the constraints. Furthermore, assume none of $x_1,y_1,z_1$ are zero, and $\epsilon \ll \min(|x_1|,|y_1|,|z_1|)$.

WLOG consider only $\mathbf{v}_1$ in the positive octant, i.e. $x_1,y_1,z_1>0$ and with $x_1\ge y_1\ge z_1$. The angle between $\mathbf{v}_1$ and $\mathbf{v}_2$ is maximised when the endpoint of $\mathbf{v}_2$ is at one of the vertices of a cube of edge length $2\epsilon$ centered at $(x_1,y_1,z_1)$ and with faces parallel to the coordinate planes. It can be shown that the particular vertex at which the angle is maximum is $(x_1-\epsilon,y_1+\epsilon,z_1+\epsilon)$ if $x\ge(y+z)$; else $(x_1+\epsilon,y_1-\epsilon,z_1-\epsilon)$ if $x\le(y+z)$. In either case, the approximation for the upper bound on angle will be the same.

Then we have for the choice of $(x_1-\epsilon,y_1+\epsilon,z_1+\epsilon)$:

$$\begin{align} \cos\theta&=\frac{\mathbf{v}_1 \cdot \mathbf{v}_2}{||\mathbf{v}_1||\,||\mathbf{v}_2||} \\[1em] &=\frac{(x_1-\epsilon,y_1+\epsilon,z_1+\epsilon)\cdot(x_1,y_1,z_1)}{||(x_1-\epsilon,y_1+\epsilon,z_1+\epsilon)||\,||(x_1,y_1,z_1)||} \\[1em] &=\frac{x_1(x_1-\epsilon)+y_1(y_1+\epsilon)+z_1(z_1+\epsilon)}{\sqrt{x_1^2+y_1^2+z_1^2}\cdot\sqrt{(x_1-\epsilon)^2+(y_1+\epsilon)^2+(z_1+\epsilon)^2}} \tag{1} \end{align}$$

Now let $$s_1=x_1-y_1-z_1 \tag{2}$$ and $$s_2=x_1^2+y_1^2+z_1^2 \tag{3}$$ which are fixed by our choice of $\mathbf{v}_1$. Then (1) reduces to

$$\cos\theta=\frac{s_2-s_1\epsilon}{\sqrt{s_2}\cdot\sqrt{s_2-2s_1\epsilon+3\epsilon^2}}$$ so $$\begin{align} \sec\theta &= \frac{\sqrt{1-2\dfrac{s_1}{s_2}+3\dfrac{\epsilon^2}{s_2}}}{1-\dfrac{s_1}{s_2}\epsilon} = \frac{\sqrt{1-2\dfrac{s_1}{s_2}+\dfrac{s_1^2}{s_2^2}\epsilon^2+\dfrac{(3s_2-s_1^2)}{s_2^2}\epsilon^2}}{\sqrt{1-2\dfrac{s_1}{s_2}\epsilon+\dfrac{s_1^2}{s_2^2}\epsilon^2}} \\[1em] &= \sqrt{1+\frac{\dfrac{(3s_2-s_1^2)}{s_2^2}\epsilon^2}{\left(1-\dfrac{s_1}{s_2}\epsilon\right)^2}} \\[1em] &\approx 1+\frac{(3s_2-s_1^2)}{2s_2^2}\epsilon^2 \tag{4} \end{align}$$

because $\sqrt{1+x}=1+\frac{1}{2}x+\frac{(\frac{1}{2})(-\frac{1}{2})}{2}x^2+\ldots \approx 1+\frac{1}{2}x$ by the binomial series, and in the denominator $1-\frac{s_1}{s_2}\epsilon \to 1$ as $\epsilon \to 0$.

But the Taylor Series for $\sec\theta$ is:

$$\sec\theta=1+\frac{1}{2}\theta^2+\ldots \approx 1+\frac{1}{2}\theta^2 \tag{5}$$

So equating the approximations in (4) and (5):

$$\theta_{max} \approx \frac{\sqrt{3s_2-{s_1}^2}}{s_2}\,\epsilon \tag{6}$$


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.