Given a vector $v$ and an angle $\theta$, find a vector $w$ such that the angle between $v$ and $w$ is exactly $\theta$

Question

Given:

• an $n$-dimensional vector $v = (v_1,\ldots,v_n) \in \mathbb{R}^n$
• an angle $\theta \in (0, \frac{\pi}{2})$

I am looking for a technique to construct a vector $w = (w_1,\ldots, w_n) \in \mathbb{R}^n$ such that:

$$\frac{v\cdot w}{\|v\|\|w\|} = \cos(\theta)$$

I am specifically interested in a way to do this for $n > 3$ dimensions (for $n \leq 3$, I think you could do it with rotation matrices).

I realize that when $n \geq 3$, there are an infinite number of feasible vectors $w$. I would be OK with any $w$ that satisfies the following criteria:

1. $v$ and $w$ are distinct: $v_i \neq w_i$ for all $i$.

2. all components of $w$ are non-zero: $w_i \neq 0$ for all $i$.

Answer 1

A "just wing it" algorithm which I think would be numerically stable is

• Find a nonzero vector $u$ that is perpendicular to $v$, for example $$ u=(0,\ldots,0,v_j,0,\ldots,0,-v_i,0,\ldots,0)$$ where $v_i$ and $v_j$ are the numerically largest coordinates of $v$.

• Normalize $u$ so its norm equals $\|v\|$.

• Take $w=\cos(\theta)v+\sin(\theta)u$.

Note that $w$ does not depend continuously on $v$, but that is impossible when $n$ is odd, due to the hairy ball theorem. For even $n$ you can get continuity by taking $$ u = (v_2,-v_1,v_4,-v_3,\ldots) $$

---

If it's important to you to make all elements of $w$ nonzero, instead choose all elements of $u$ except one randomly (with some continuous distribution). Let the one you don't choose randomly correspond to the numerically largest coordinate of $v$; then you can easily solve for the value of that coordinate that will make $u\cdot v=0$.

Then each element of $ w $ will almost surely be nonzero.

Answer 2

I'm adding another solution (credit goes to one of the PhD students in my lab), since I found it elegant / wanted to post some code for others.

---

Solution:

Given $v$, we want to find a $w$ such that:

$$\frac{v.w}{\|v\|\|w\|} = \cos(\theta)$$

For clarity of exposition, let's assume that $\|v\| =1$ and $\|w\| = 1$ (if $v$ is not normalized, then normalize it). We can now express our required condition as:

$$v.w = \cos(\theta)$$

We will construct a suitable vector as $$w = U.r$$ where:

• $U = [\, v \,|\, u_2 \,|\, \ldots \,|\,u_n] \in \mathbb{R}^{n\times n}$ is an orthonormal matrix with $v$ in the first column. We can construct this matrix given $v$ by using the Gram-Schmidt process to determine the remaining basis vectors $u_2,\ldots, u_d$.

• $r = (r_1, r_2, \ldots, r_d) \in \mathbb{R}^n$ is a unit vector where $r_1 = \cos(\theta)$ and $\|r\| = 1$

To see that these choices of $U$ and $r$ produce a suitable $w$, note:

$$\begin{align} v.w &= v.(Ur)\\ &= v.(u_1.r_1 + u_2.r_2, \ldots, u_d. r_d)\\ &= (v.u_1) r_1 + (v.u_2)r_2 + \ldots + (v.u_d)r_d\\ \end{align}$$ By construction of $U$, we have that $u_1 = v$ and $v.u_i = 0$ for $i = 2,\ldots, n$, so: $$\begin{align} (v.u_1) r_1 + (v.u_2)r_2 + \ldots + (v.u_d)r_d = \|v\|^2 r_1 = r_1 \end{align}$$ Since we have set $r_1 = \cos(\theta)$, we therefore have that: $$\frac{v.w}{\|v\|\|w\|} = r_1 = \cos(\theta)$$ as desired.

Note that we are not quite done, since we still have choose $r_2,\ldots, r_d$ so that $\|r\| = 1$. This requires: $$ \begin{align} \|r\| &= 1\\ {r_1}^2 + {r_2}^2 + \ldots + {r_n}^2 &= 1\\ \cos^2(\theta) + {r_2}^2 + \ldots + {r_n}^2 &= 1\\ {r_2}^2 + \ldots + {r_n}^2 &= 1 - \cos^2(\theta) \\ {r_2}^2 + \ldots + {r_n}^2 &= \sin^2(\theta) \end{align} $$ In other words, $r_2,\ldots, r_n$ must lie on the surface of a $n$-dimensional sphere with radius $\sin(\theta)$. While there are many ways to choose these values, one easy way is to sample this point uniformly from the surface of the sphere. Based on this post, this can be achieved by sampling a $n -1$ random variables $r_i \sim N(0,1)$ and scaling them so that their norm is exactly $\sin(\theta)$

---

Code:

import numpy as np
from numpy.linalg import norm

def rotate_vector(v, angle):

    """
    returns vector w such that |w|=|v| and v.w/|v||w| = cos(angle) and 
    :param v: numpy array with dimension d >= 2
    :param angle: angle of rotation in radians 
    :return: w, numpy array of dimension v.dim
    """

    assert isinstance(v, np.ndarray)
    assert np.all(np.isfinite(v)) and v.ndim == 1 and len(v) >= 2
    norm_value = norm(v)
    d = len(v)

    # normalize vector
    v = v / norm(v)

    # construct orthonormal matrix U where U[:, 0] = v
    while True:
        u_diag = np.random.randn(d)
        if not np.any(np.isclose(u_diag, 0.0, atol = 0.01)):
            break

    U = np.diag(u_diag)
    U[:, 0] = v
    U, _ = np.linalg.qr(U)
    assert np.all(np.isclose(np.abs(U[:, 0]), np.abs(v)))

    # overcases cases where U flips the sign
    if np.all(np.isclose(U[:, 0], -v)):
        U = -U

    # generate a vector r = [r_1, r_2, .., .. r_d ]
    # r = cos(angle)
    # r_2,..., r_d are any non-zero values such that (r_2)^2 + ... + (r_d)^2 = sin^2(alpha)

    # r_2,..,r_d is just a point on a d-1 dimensional sphere with radius sin(alpha)
    # this can be sampled uniformly using a multivariate gaussian
    # see https://math.stackexchange.com/a/1585996/7145
    while True:
        z = np.random.randn(d-1)
        if not np.any(np.isclose(z, 0.0, atol = 0.01)):
            break

    r = np.sin(angle) * (z / norm(z))
    r = np.insert(r, 0, np.cos(angle))

    # compute rotated unit vector
    w = np.dot(U, r)
    assert np.isclose(norm(w), 1.0)
    assert np.isclose(angle, np.arccos(np.dot(v, w)))

    # rescale vector with the same norm as original input
    w = w * norm_value
    return w