28
$\begingroup$

$\DeclareMathOperator\ord{ord}\DeclareMathOperator\Im{Im}$The valence formula for a modular form asserts that if $f: \mathbf{H} \to \mathbf{C}$ is a modular form of weight $k$ on the upper half-plane ${\mathbf H} := \{ z: \Im z > 0 \}$ in the sense that $$ f\left(\frac{az+b}{cz+d}\right) = (cz+d)^k f(z)$$ for all $\begin{pmatrix} a & b \\ c & d \end{pmatrix} \in SL_2({\bf Z})$, and $f$ is holomorphic on all of ${\mathbf H}$ and also bounded as $\Im z \to \infty$ then (if $f$ is not identically zero) one has the identity $$ \sum_\rho \ord_\rho(f) + \ord_\infty(f) + \frac{1}{2} \ord_i(f) + \frac{1}{3} \ord_{e^{\pi i/3}}(f) = \frac{k}{12}$$ where $\rho$ ranges over the zeroes of $f$ away from the (partially) fixed points $i, e^{\pi i/3}, \infty$ of the action, with each orbit of $\operatorname{SL}_2(\mathbf{Z})$ avoiding these fixed points being represented precisely once in this sum, and the order of vanishing at infinity defined in terms of the (squared) nome variable $q = e^{2\pi i z}$.

The standard proof of this identity proceeds by integrating the logarithmic derivative of $f'/f$ on a somewhat complicated contour designed to avoid zeroes, and which looks something like this (image taken from this source):

complicated contour

The claim then follows from a routine application of the residue theorem, after estimating the contribution of the various components of the contour appropriately.

It turns out that this argument, while straightforward, is somewhat difficult to formalize in proof assistant languages; see the discussion in the slides linked previously. Are there other proofs of this formula that do not rely on a tricky contour integration? I have experimented with a Green's formula type approach in which $f'/f$ is integrated against a suitable cutoff function, say on the unit disk in the (squared) nome variable $q = e^{2\pi i z}$, but the computations are quite complicated. One can also obtain this formula from using the $j$-invariant to coordinatize the modular curve, though this is somewhat circular as often the valence formula is needed to establish properties of this invariant. I also considered trying to use general Riemann surface tools such as the Riemann–Hurwitz formula, but among other things I ran into the need to triangularize a Riemann surface, which also would be complicated to formalize I think.

$\endgroup$
3
  • $\begingroup$ Regarding the use of the j-invariant: perhaps it might be possible to avoid circularity by using a weaker form of the valence formula to establish the necessary properties of the j-invariant, and then looping back round to establish the full valence formula from this. E.g. perhaps one could show by other, more elementary means that $j$ has no zeroes on the boundary of $\mathcal{D}$ except the two at the corners, and then one can integrate around a much simpler contour. $\endgroup$ Commented Mar 19 at 8:36
  • 1
    $\begingroup$ One interpretation of the proof of the valence formula is that it is proving a generalized version of the residue theorem where poles are allowed to lie on the contour and the integral is therefore somewhat regularized. Shifting the contour away from poles is one approach to regularizing the integral, but there are others. The key desiderata the regularization needs to satisfy are (1) it agrees with an ordinary integral when there are no poles and (2) one can compute the regularized integral of functions like $1/(z-a)^n$. $\endgroup$
    – Will Sawin
    Commented Mar 19 at 20:27
  • 1
    $\begingroup$ Combining those, one gets the residue theorem by expressing the function as a combination of holomorphic functions and $1/(z-a)^n$s. Then for the valence formula one needs (3) compatibility with the change of variables one uses to cancel the contour integral. Maybe one in fact only needs $n=1$ since one has a logarithmic derivative. There are a lot of approaches to regularizing divergent integrals, and maybe one of them could lead to a simpler-to-formalize proof? $\endgroup$
    – Will Sawin
    Commented Mar 19 at 20:30

2 Answers 2

6
$\begingroup$

Since you cited my talk, I'd like to also mention something that I learnt only yesterday and that might lead to a simpler proof than what I discussed in this talk.

I'm talking about the paper Non-integer valued winding numbers and a generalized Residue Theorem by Hungerbühler and Wasem.

Using the Cauchy principal value, one can define the winding number of a piecewise smooth contour even in cases where the point in question lies directly on the contour. For a geometrical intuition: in the case of a closed simple counter-clockwise loop, the winding number of a point directly on the loop is then the probability that an ε-perturbation of the point in a random direction will be inside the curve. For a square-shaped contour, the four corners have winding number 1/4 and every other point on it has winding number 1/2.

With this, one can then prove versions of the Residue Theorem and the Argument Principle that also work when there are (isolated) singularities directly on the contour. There are some restrictions, but none of them should apply in the case you have.

I think if one could formalise this (and I don't see any obstacles to doing that) the proof of the valence formula should be relatively easy.

Disclaimer: I'm not exactly an expert in complex analysis and I only skimmed the paper.

$\endgroup$
0
5
$\begingroup$

I am recording here a "regularized" version of the standard contour integral argument that avoids detours (though it still needs the standard contour traversing the standard fundamental domain). I first need a regularized version of the residue theorem:

Smoothed residue theorem. If $f: U \to {\bf C}$ is holomorphic and not identically zero on some connected open $U$, and $\psi \in C^\infty_c(U)$, then $$ \sum_\rho \psi(\rho) = -\frac{1}{\pi} \int_U \frac{f'}{f}(x+iy) \partial_{\bar z} \psi(x+iy)\ dx dy$$ where $\partial_{\bar z} = \frac{1}{2} (\partial_x + i \partial_y)$ is the Wirtinger derivative, and $\rho$ ranges over the zeroes of $f$ counted with multiplicity.

Proof We can factor $f(z) = g(z) \prod_\rho (z-\rho)$ where $\rho$ ranges over the zeroes of $f$ in the support of $\psi$ (counting multiplicity), and $g$ is non-vanishing on $\psi$. Then $\frac{f'}{f}(z) = \frac{g'}{g}(z) + \sum_\rho \frac{1}{z-\rho}$, and $\partial_{\bar z} \frac{g'}{g}=0$ by the Cauchy-Riemann equations, so it suffices to show that $$ \psi(\rho) = -\frac{1}{\pi} \int_U \frac{1}{x+iy-\rho} \partial_{\bar z} \psi(x+iy)\ dx dy.$$ Using polar coordinates $x+iy = \rho + re^{i\theta}$, the right-hand side is $$ -\frac{1}{2\pi} \int_0^\infty \int_0^{2\pi} \frac{1}{re^{i\theta}} (e^{i\theta} \partial_r + ie^{i\theta} \partial_\theta) \psi(\rho+re^{i\theta}) \ r dr d\theta,$$ and the claim then follows from two applications of the fundamental theorem of calculus. $\Box$

Now we work on the uncompactified modular curve $Y(1) = \Gamma \backslash {\bf H}$ with $\Gamma := \mathrm{SL}_2({\bf Z})$, which has the standard fundamental domain $F := \{ z \in {\bf H}: |z| \geq 1, |\mathrm{Re} z| \leq 1/2\}$:

Smoothed valence formula for Y(1). If $f$ is a modular form of weight $k$ that is not identically zero, and $\psi \in C^\infty_c(Y(1))$, then $$ \sum_{\Gamma \rho} \psi(\Gamma \rho)/|\mathrm{Stab}(\rho)| = - \frac{1}{\pi} \int_F \frac{f'}{f}(x+iy) \partial_{\bar z} \psi(x+iy)\ dx dy + \frac{ik}{2\pi} \int_{\pi/3}^{\pi/2} \psi(e^{i\theta})\ d\theta,$$ where $\Gamma \rho$ ranges over the elements of $Y(1)$ on which $f$ vanishes (counting multiplicity), and $\mathrm{Stab}(\rho)$ is the stabilizer of the $\Gamma$ action on $\rho$ (thus this is of order two if $\Gamma \rho = \Gamma i$, of order three if $\Gamma \rho = \Gamma e^{\pi i/3}$, and trivial otherwise).

Proof. Using a smooth partition of unity (to localize the support of $\psi$ to individual coordinate charts) we see that we can write $$ \psi(\Gamma z) = \sum_{\gamma \in \Gamma} \varphi(\gamma z)$$ for some $\varphi \in C^\infty_c(\mathbf{H})$. Then we can unfold $$ \sum_{\Gamma \rho} \psi(\Gamma \rho)/|\mathrm{Stab}(\rho)| = \sum_\rho \varphi(\rho)$$ so by the smoothed residue theorem it suffices to show that $$ \int_{\bf H} \frac{f'}{f}(x+iy) \partial_{\bar z} \varphi(x+iy)\ dx dy = \int_F \frac{f'}{f}(x+iy) \partial_{\bar z} \psi(x+iy)\ dx dy - \frac{ik}{2} \int_{\pi/3}^{\pi/2} \psi(e^{i\theta})\ d\theta.$$ We can split the left-hand side using the fundamental domain into $$ \sum_{\gamma \in \Gamma} \int_{\gamma F} \frac{f'}{f}(x+iy) \partial_{\bar z} \varphi(x+iy)\ dx dy $$ which after a change of variables becomes $$ \sum_{\gamma \in \Gamma} \int_{F} \frac{f'}{f}(\gamma z) (\partial_{\bar z} \varphi)(\gamma z)\ \frac{dx dy}{|cz+d|^4}$$ where we write $z=x+iy$ and $\gamma z= \frac{az+b}{cz+d}$. From modularity we can compute $$ \frac{f'}{f}(\gamma z) = (cz+d)^2 \frac{f'}{f}(z) + k c(cz+d)$$ while from the chain rule we have $$ (\partial_{\bar z} \varphi)(\gamma z) = (\overline{cz+d})^2 \partial_{\bar z} (\varphi(\gamma z)).$$ The contributions involving $\frac{f'}{f}$ now line up, and it remains to show that $$ \sum_{\gamma \in \Gamma} \int_{F} \frac{c}{cz+d} (\partial_{\bar z} \varphi(\gamma z))\ dx dy = - \frac{i}{2} \int_{\pi/3}^{\pi/2} \psi(e^{i\theta})\ d\theta.$$ Since $\partial_{\bar z} \frac{c}{cz+d} = 0$ (and $\varphi$ is compactly supported), we can apply Stokes' theorem and the definition of the Wirtinger derivative to write the left-hand side as $$ \frac{-i}{2} \sum_{\gamma \in \Gamma} \int_{\partial F} \frac{c}{cz+d} \varphi(\gamma z)\ dz$$ where $\partial F$ is traversed anticlockwise.

The contribution of the vertical sides of $\partial F$ can be combined as $$ \frac{-i}{2} \int_{e^{i\pi/3}}^{e^{i\pi/3}+i\infty} \sum_{\gamma \in \Gamma} (\frac{c}{cz+d} \varphi(\gamma z) - \frac{c}{c\alpha z+d} \varphi(\gamma \alpha z))\ dz$$ where $\alpha$ is the shift $z \mapsto z-1$. Relabeling $\gamma \alpha$ by $\gamma$ in the second term (which changes $\frac{c}{c\alpha z+d}$ to $\frac{c}{cz+d}$), we see that this expression cancels to zero. Similarly, the contribution of the bottom side of $\partial F$ is $$ \frac{-i}{2} \int_C \sum_{\gamma \in \Gamma} (\frac{c}{cz+d} \varphi(\gamma z) - \frac{1}{z^2} \frac{c}{c\beta z+d} \varphi(\gamma \beta z))\ dz$$ where $\beta$ is the inversion map $z \mapsto -1/z$, and $C$ is the anticlockwise contour on the circle from $e^{\pi i/3}$ to $i$. Relabeling $\gamma \beta$ to $\gamma$ in the second term (which changes $\frac{c}{c\beta z+d}$ to $\frac{d}{-d/z-c}$), the claim then follows after a routine calculation. $\Box$

Now we pass to the compactification $X(1) = Y(1) \cup \{\infty\}$.

Smoothed valence formula for X(1). If $f$ is a modular form of weight $k$ that is not identically zero, and vanishes to order $m$ at infinity, and $\psi \in C^\infty(X(1))$, then $$ \sum_{\Gamma \rho} \psi(\Gamma \rho)/|\mathrm{Stab}(\rho)| + m \psi(\infty) = - \frac{1}{\pi} \int_F \frac{f'}{f}(x+iy) \partial_{\bar z} \psi(x+iy)\ dx dy + \frac{ik}{2\pi} \int_{\pi/3}^{\pi/2} \psi(e^{i\theta})\ d\theta.$$

Proof In view of the smoothed valence formula for $Y(1)$, we may assume that $\psi$ is supported on a neighborhood of infinity that avoids all other zeroes of $f$. Our task is now to show that $$ \int_F \frac{f'}{f}(x+iy) \partial_{\bar z} \psi(x+iy)\ dx dy = - \pi m.$$ Since $(f'/f) \psi$ decays (exponentially fast) to $2\pi i m \psi(\infty)$ and obeys the Cauchy-Riemann equations $\partial_{\bar z} \frac{f'}{f} = 0$, this follows from a routine integration by parts and a limiting argument. $\Box$

Finally, taking $\psi=1$ in the smoothed valence formula for $X(1)$, we recover the classical valence formula.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .