Spherical functions

Python/numba package for evaluating and transforming Wigner $\mathfrak{D}$ matrices and spin-weighted spherical harmonics directly in terms of quaternions, and in more standard forms

Project maintained by moble

Wigner $\mathfrak{D}$ matrices

The Wigner $\mathfrak{D}$ matrices allow us to represent the rotation group by means of finite-dimensional matrices. They reduce to spherical functions in special cases, which further allows us to derive the rotation of those spherical functions—as described in more detail on this page. They are derived in terms of a particular split of the quaternion group into two parts.

Explicitly, a quaternion $\quat{Q}$ can be expressed in terms of two complex numbers $\quat{Q}_a = \quat{Q}_1 + i\, \quat{Q}_z$ and $\quat{Q}_b = \quat{Q}_y + i\, \quat{Q}_x$.1 This is only important because it allows us to verify the multiplication law \begin{align} \label{eq:QuaternionComponentProducts} (\quat{P}\,\quat{Q})_a &= \quat{P}_a\, \quat{Q}_a - \co{\quat{P}}_b\, \quat{Q}_b, \\\
(\quat{P}\,\quat{Q})_b &= \quat{P}_b\, \quat{Q}_a + \co{\quat{P}}_a\, \quat{Q}_b. \end{align} Given a rotor $\rotor{R}$, these two complex numbers are the quantities actually used in computing the $\mathfrak{D}$ matrix elements. Note that there is a series of basic choices leading to both the decomposition of a quaternion into two complex numbers, and the product law. None of these choices is set in stone; we just choose something with an eye to the desired result.

The following is essentially the same as Wigner’s original derivation, but with more elegance, and more sensitivity to numerical issues and special cases. This version of the derivation comes from a paper I wrote a couple years ago, and is the source of the code used in this module.

The basic idea of the derivation is to construct a $(2\ell+1)$-dimensional vector space of homogeneous polynomials in these complex numbers $\quat{Q}_a$ and $\quat{Q}_b$. To make that a little more concrete, the basis of this vector space is \begin{equation*} \label{WignerBasisComponent} \mathbf{e}_{(m)}(\quat{Q}) \defined \frac{\quat{Q}_{a}^{\ell+m}\, \quat{Q}_{b}^{\ell-m}} {\sqrt{ (\ell+m)!\, (\ell-m)! }}. \end{equation*} Here, $m$ ranges from $-\ell$ to $\ell$ in steps of $1$, but $\ell$ (and hence $m$) can be a half-integer. Now, the key idea is that a rotation of $\quat{Q}$ by some new rotor $\quat{R}$ gives us a new vector basis, which we can represent in terms of the basis shown above. In fact, we get a matrix transforming one set of basis vectors to another. We’ll write this as \begin{equation*} \mathbf{e}_{(m’)}(\quat{R}\, \quat{Q}) = \sum_{m} \mathfrak{D}^{(\ell)}_{m’,m}(\quat{R})\, \mathbf{e}_{(m)}(\quat{Q}). \end{equation*}

So now, we’ve defined the $\mathfrak{D}$ matrices. But we can also plug $\quat{R}\, \quat{Q}$ into the original expression for $\mathbf{e}$, and figure out what $\mathfrak{D}$ should actually be. We’ll have to use the expressions for $(\quat{R}\, \quat{Q})_a$ and $(\quat{R}\, \quat{Q})_b$ given above, and we’ll find that we have polynomials with terms given as sums of two different things.

This brings us to the first fork in the road. If either $\quat{R}_a$ or $\quat{R}_b$ is tiny (i.e., the absolute value is numerically small), we can (and in fact must) treat these as special cases. First, if $\lvert \quat{R}_a \rvert \lesssim 10^{-15}$, for example, we can just ignore it; since $\lvert \quat{R} \rvert=1$ (within numerical precision), we are assured that $\lvert \quat{R}_b \rvert \approx 1$. Thus, we get \begin{align*} \mathbf{e}_{(m’)}(\quat{R}\, \quat{Q}) &\approx \frac{ (- \co{\quat{R}}_{b}\, \quat{Q}_{b})^{\ell+m’}\, (\quat{R}_{b}\, \quat{Q}_{a})^{\ell-m’} } { \sqrt{ (\ell+m’)!\, (\ell-m’)! } }, \\\
&\approx (- \co{\quat{R}}_{b})^{\ell+m’}\, (\quat{R}_{b})^{\ell-m’}\, \mathbf{e}_{(-m’)}(\quat{Q}). \end{align*} In this case, it’s not hard to see that the expression for the $\mathfrak{D}$ matrix is \begin{equation} \label{eq:D_RaApprox0} \mathfrak{D}^{(\ell)}_{m’,m}(\quat{R}) = (-1)^{\ell+m’}\, \quat{R}_b^{-2m’} \delta_{-m’,m} = (-1)^{\ell-m}\, \quat{R}_b^{2m} \delta_{-m’,m}. \end{equation} In the same way, we can calculate this for $\lvert \quat{R}_b \rvert \lesssim 10^{-15}$: \begin{equation} \label{eq:D_RbApprox0} \mathfrak{D}^{(\ell)}_{m’,m}(\quat{R}) = \quat{R}_a^{2m’} \delta_{m’,m} = \quat{R}_a^{2m} \delta_{m’,m}. \end{equation}

Now, the other fork in that road is the general case, when both components have larger magnitudes. We have powers of the sum of those terms in Eq. \eqref{eq:QuaternionComponentProducts}. This leads us to use (two applications of) the binomial expansion. After a little simplification, we can express the result as \begin{multline} \label{eq:DAnalytically} \mathfrak{D}^{(\ell)}_{m’,m}(\quat{R}) = \sum_{\rho} \binom{\ell+m’} {\rho}\, \binom{\ell-m’} {\ell-\rho-m}\, (-1)^{\rho}\, \\\\ \times \quat{R}_{a}^{\ell+m’-\rho}\, \co{\quat{R}}_{a}^{\ell-\rho-m}\, \quat{R}_{b}^{\rho-m’+m}\, \co{\quat{R}}_{b}^{\rho}\, \sqrt{ \frac{ (\ell+m)!\, (\ell-m)! } { (\ell+m’)!\, (\ell-m’)! } }. \end{multline} It turns out that this expression is not the best way to implement the calculation in the code. The reason is that we would need to take a bunch of exponents of complex numbers, and there’s the possibility that the sum would cancel out to give a very small number, which would be polluted with roundoff, etc. So we manipulate it to put it in a better form. For example, we can simplify the above as \begin{align} \nonumber \mathfrak{D}^{(\ell)}_{m’,m}(\quat{R}) &= \sqrt{ \frac{ (\ell+m)!\, (\ell-m)! } { (\ell+m’)!\, (\ell-m’)! } }\, \quat{R}_{a}^{\ell+m’}\, \co{\quat{R}}_{a}^{\ell-m}\, \quat{R}_{b}^{-m’+m} \\\
\nonumber &\qquad \times \sum_{\rho} \binom{\ell+m’} {\rho}\, \binom{\ell-m’} {\ell-\rho-m}\, (-1)^{\rho}\, \quat{R}_{a}^{-\rho}\, \co{\quat{R}}_{a}^{-\rho}\, \quat{R}_{b}^{\rho}\, \co{\quat{R}}_{b}^{\rho}, \\\
\nonumber &= \sqrt{ \frac{ (\ell+m)!\, (\ell-m)! } { (\ell+m’)!\, (\ell-m’)! } }\, \lvert \quat{R}_{a} \rvert^{2\ell-2m}\, \quat{R}_{a}^{m’+m}\, \quat{R}_{b}^{-m’+m} \\\
\label{eq:D_RaGeqRb} &\qquad \times \sum_{\rho} \binom{\ell+m’} {\rho}\, \binom{\ell-m’} {\ell-\rho-m}\, \left( - \frac{\lvert \quat{R}_{b} \rvert^2} {\lvert \quat{R}_{a} \rvert^2} \right)^{\rho}. \end{align} Typically, this last expression can be evaluated pretty efficiently.

But to make it really fast, accurate, and robust, we have to use some tricks. First of all, the various combinatorical factors are expensive to re-evaluate each time. The obvious way to handle this is just to pre-compute them, and have a function returning the appropriate values with some simple indexing tricks. These functions are implemented in this module’s initialization code.

Now, that sum is essentially a polynomial, and the best way to evaluate a polynomial uses Horner form — which is both faster and more accurate than the naive approach. Also, since the coefficients involve factorials of the summation index, we can factor out the $\rho_{\text{min}}$ binomials, and be left with just a few factorials, which we can then evaluate more efficiently. This also allows us to pull out the lowest-order coefficient of that polynomial: $(-\lvert R_b \rvert / \lvert R_a \rvert)^{2 \rho_{\mathrm{min}}}$. We will see momentarily that being able to do this is very important.

That deals very nicely with the sum, but we also need to deal with the factor in front of the sum. And there are some fairly subtle complications to deal with when evaluating that factor. First of all, complex exponentials are slow. Second, there are cases where those terms can become really huge, only to be multiplied by something really tiny, leaving us reasonable values in the neighborhood of 1. But those exponents can be very large; if we’re looking for matrix elements for $\ell=32$, the exponents will range from $-64$ all the way up to $64$. And since the largest and smallest numbers python can represent are in the neighborhood of $10^{323}$ and $10^{-308}$, respectively, we can easily get nan or 0 for calculations that should actually give us something that is neither of those values — for example, if $\lvert R_b \rvert \approx 10^{-6}$ and $-m’=m=-32$. But these terms only appear due to our separation of powers into the sum. If we now bring out the lowest factor from the sum, we can make those exponents disappear. The problem now is to cancel the exponents in a mixture of exponents of complex numbers and exponents of real numbers, and to do so efficiently.

The solution to this problem is simple and, miraculously, makes the code 2.3 times faster! We decompose $R_a$ and $R_b$ into magnitude and phase, calculate the exponents to which those must be raised, and then produce the result as a complex number from the separate magnitude and phase. There are fast functions in the standard python module cmath that do these operations, and they are directly supported in numba since version 0.15, so this is quite a simple solution to implement, also. The calculation of this prefactor is given in this module here.

But this doesn’t solve all our problems. The same issues can creep up if $\lvert R_a \rvert$ is small (though it requires a smaller magnitude, and occurs for a smaller subset of $m’,m$ values). In this case, it would be better if we could find an expression that sort of reverses the roles of $a$ and $b$. It turns out that this isn’t hard. In deriving Eq. \eqref{eq:DAnalytically}, a choice was made regarding the summation variable. We can simply transform that summation variable as $\rho \mapsto \ell-m-\rho$, and obtain \begin{multline*} \mathfrak{D}^{(\ell)}_{m’,m}(\quat{R}) = \sum_{\rho} \binom{\ell+m’} {\ell-m-\rho}\, \binom{\ell-m’} {\rho}\, (-1)^{\ell-m-\rho}\, \\\\ \times \quat{R}_{a}^{m’+m+\rho}\, \co{\quat{R}}_{a}^{\rho}\, \quat{R}_{b}^{\ell-m’-\rho}\, \co{\quat{R}}_{b}^{\ell-m-\rho}\, \sqrt{ \frac{ (\ell+m)!\, (\ell-m)! } { (\ell+m’)!\, (\ell-m’)! } }. \end{multline*} It’s interesting to note the symmetry with the earlier version of this equation; we’ve essentially just exchanged the labels $a$ and $b$, while also reversing the sign of $m’$, and multiplying by an overall factor of $(-1)^{\ell+m}$.

In any case, we can apply the same simplification to this expression as before: \begin{align} \nonumber \mathfrak{D}^{(\ell)}_{m’,m}(\quat{R}) &= (-1)^{\ell-m} \sqrt{ \frac{ (\ell+m)!\, (\ell-m)! } { (\ell+m’)!\, (\ell-m’)! } } \quat{R}_{a}^{m’+m}\, \quat{R}_{b}^{\ell-m’}\, \co{\quat{R}}_{b}^{\ell-m} \\\
\nonumber &\qquad \times \sum_{\rho} \binom{\ell+m’} {\ell-m-\rho}\, \binom{\ell-m’} {\rho}\, (-1)^\rho\, \quat{R}_{a}^{\rho}\, \co{\quat{R}}_{a}^{\rho}\, \quat{R}_{b}^{-\rho}\, \co{\quat{R}}_{b}^{-\rho} \\\
\nonumber &= (-1)^{\ell-m} \sqrt{ \frac{ (\ell+m)!\, (\ell-m)! } { (\ell+m’)!\, (\ell-m’)! } } \quat{R}_{a}^{m’+m}\, \quat{R}_{b}^{m-m’}\, \lvert \quat{R}_{b} \rvert^{2\ell-2m} \\\
\label{eq:D_RaLeqRb} &\qquad \times \sum_{\rho} \binom{\ell+m’} {\ell-m-\rho}\, \binom{\ell-m’} {\rho}\, \left( - \frac{ \lvert \quat{R}_{a} \rvert^2 } { \lvert \quat{R}_{b} \vert^2 } \right)^{\rho} \end{align} And again, we evaluate this cleverly, as above.

So we get four branches in our logic, with a different expression for $\mathfrak{D}$ in each branch:

  1. When $\lvert \quat{R}_a \rvert \lesssim 10^{-15}$, use Eq. \eqref{eq:D_RaApprox0}.
  2. When $\lvert \quat{R}_b \rvert \lesssim 10^{-15}$, use Eq. \eqref{eq:D_RbApprox0}.
  3. When $\lvert \quat{R}_a \rvert \geq \lvert \quat{R}_b \rvert$, use Eq. \eqref{eq:D_RaGeqRb}.
  4. When $\lvert \quat{R}_a \rvert < \lvert \quat{R}_b \rvert$, use Eq. \eqref{eq:D_RaLeqRb}.

Note that these expressions are valid even for half-integer values of $\ell$, noting that if $\ell$ is half-integer, then so must $m’$ and $m$ be. However, in the interests of fast implementation, the code in this module assumes integer values. (It would be simple for me to implement the more general case. If this is a functionality you need, please feel free to open an issue on this module’s github page to request it.)

Relation to the antiquated form of $\mathfrak{D}$ using Euler angles

I hope I don’t have to repeat my utter disdain for the use of Euler angles. However, it is important to make contact with other literature to be able to compare conventions. As noted previously, the rotation performed by the set $(\alpha, \beta, \gamma)$ of Euler angles (using conventions to agree with Wikipedia’s page on $\mathfrak{D}$ matrices) can be written in quaternion form as \begin{align*} \quat{R}_{(\alpha, \beta, \gamma)} &= e^{\alpha\, \basis{z}/2}\, e^{\beta\, \basis{y}/2}\, e^{\gamma\, \basis{z}/2}, \\\
&= \left( \cos \frac{\alpha}{2}\, \cos \frac{\beta}{2}\, \cos \frac{\gamma}{2} -\sin \frac{\alpha}{2}\, \cos \frac{\beta}{2}\, \sin \frac{\gamma}{2} \right) \\\
&\qquad + \basis{x} \left( \cos \frac{\alpha}{2}\, \sin \frac{\beta}{2}\, \sin \frac{\gamma}{2} -\sin \frac{\alpha}{2}\, \sin \frac{\beta}{2}\, \cos \frac{\gamma}{2} \right) \\\
&\qquad + \basis{y} \left( \cos \frac{\alpha}{2}\, \sin \frac{\beta}{2}\, \cos \frac{\gamma}{2} +\sin \frac{\alpha}{2}\, \sin \frac{\beta}{2}\, \sin \frac{\gamma}{2} \right) \\\
&\qquad + \basis{z} \left( \sin \frac{\alpha}{2}\, \cos \frac{\beta}{2}\, \cos \frac{\gamma}{2} +\cos \frac{\alpha}{2}\, \cos \frac{\beta}{2}\, \sin \frac{\gamma}{2} \right). \end{align*} Taking the complex components of this, we have \begin{equation*} \quat{R}_a = e^{i\,\alpha/2}\, \cos\frac{\beta}{2}\, e^{i\,\gamma/2}, \qquad \quat{R}_b = e^{-i\,\alpha/2}\, \sin\frac{\beta}{2}\, e^{i\,\gamma/2}. \end{equation*} We can plug these values into, e.g., Eq. \eqref{eq:D_RaGeqRb}, and get the standard, hideous, reprehensible form of the $\mathfrak{D}$ matrices in terms of Euler angles.

Again, of course, this is not a form that should be used for calculations, but could be useful for comparing conventions with antiquated research.


  1. This, of course, is not a productive way of thinking about quaternions, but it is a very useful way of calculating with quaternions, since complex numbers are already built in to many languages. That is, this decomposition into two complex components is something that the user probably does not need to worry about. It is, however, an isomorphism between quaternions and the usual (awful) presentation of Pauli spinors as two-component complex vectors, which is closer to the flavor of Wigner’s original derivation. It is also how the code is actually implemented.