This is a continuation of The Shared State of Mathematics. After bouncing between Silverman’s Rational Points on Elliptic Curves, Diamond–Shurman’s A First Course in Modular Forms, and Ratcliffe’s Foundations of Hyperbolic Manifolds, I kept arriving at the same space — the upper half-plane \(\mathbb{H}\) — wearing three different costumes. So I sat down to actually work through the geometry side, slowly, end to end.

The deeper motivation, which I should be honest about: I want to one day be able to read the work of maryam mirzakhani. Her thesis and her papers — on the dynamics of geodesics on moduli spaces of Riemann surfaces, on Weil–Petersson volumes, on counting simple closed geodesics on hyperbolic surfaces — every one of those topics sits on top of the upper half-plane. There is no shortcut. To even understand the statement of her theorems I need hyperbolic geometry, the modular surface, Teichmüller space, and the moduli space \(\mathcal{M}_g\) as a hyperbolic-geometric object. This post is the a baby step towards that.

This post is what I have understood so far. It is closer to: if I had to explain hyperbolic geometry to me from six months ago, what would have actually helped me get traction.

For most of school, “geometry” means Euclid:

  • lines are straight,
  • parallel lines never meet,
  • triangles have angle sum \(\pi\),
  • the shortest path between two points is a straight segment,
  • the plane is flat.

It took mathematicians roughly 2000 years to work out that this is not the only consistent geometry. The shock - which we now associate with Gauss, Bolyai, and Lobachevsky in the early 1800s, and later Beltrami, Klein, and Poincaré — is that you can cleanly negate Euclid’s parallel postulate and get a perfectly internally consistent universe.

That universe is hyperbolic geometry. Triangles have angle sum strictly less than \(\pi\). Through any external point there are infinitely many parallels. Circles grow exponentially with radius. And the natural arena, at least the one that haunts number theory, is the upper half-plane

\[\mathbb{H} = \{ z \in \mathbb{C} : \operatorname{Im}(z) > 0 \}\]

equipped with the metric

\[ds^2 = \frac{dx^2 + dy^2}{y^2}.\]

Two formulas. Most of the rest of this post is just consequences.

Tree-like comparison of Euclidean and hyperbolic geometry

Image source: Anil Zenginoglu on X. The picture is the cleanest visual I have seen of why hyperbolic space has “more room” than Euclidean space — the same tree that has to crowd itself flat in Euclidean geometry spreads out comfortably in hyperbolic geometry, because circumference grows exponentially with radius.

1. The parallel postulate, and why it is the special one

Euclid’s Elements is built on five postulates. The first four are uncontroversial — you can join two points by a line, you can extend a finite line, all right angles are equal, and so on. The fifth one is uncomfortable:

Given a line \(L\) and a point \(p \notin L\), there is exactly one line through \(p\) that does not meet \(L\).

For two millennia mathematicians tried to derive it from the other four. Saccheri got the furthest. In 1733 he assumed the parallel postulate was false and tried to derive a contradiction. He produced theorem after theorem with no contradiction in sight, and at the end of his book declared the result “repugnant to the nature of the straight line” and stopped. He had effectively built hyperbolic geometry while refusing to believe in it.

Gauss worked it out privately around 1816 and never published, allegedly afraid of what he called the “uproar of the Boeotians”. Bolyai and Lobachevsky published independently in the late 1820s and 1830s. The proof of consistency — that the new geometry is not self-contradictory — came later, from Beltrami in 1868, by building an explicit model inside Euclidean space. Klein and Poincaré followed with cleaner models, the upper half-plane being one of them.

The trichotomy that emerges is clean:

  • Through \(p\) there is exactly one parallel to \(L\) — Euclidean.
  • Through \(p\) there is no parallel to \(L\) — spherical (elliptic).
  • Through \(p\) there are infinitely many parallels to \(L\) — hyperbolic.

These are not three flavors of the same geometry. They are three genuinely different geometries, distinguished by a single number: the curvature.

The lesson took me a while to absorb: an axiom is not a fact, it is a choice. School frames axioms as self-evident truths — “of course parallel lines never meet” — and you have to actively unlearn that. Negating the parallel postulate does not break geometry. It selects a different one.

2. Curvature is the hidden parameter

Riemannian geometry teaches that local curvature — the failure of parallel transport to commute around an infinitesimal loop — is the invariant that controls everything. For a 2-dimensional surface, sectional curvature collapses to a single function \(K\), the Gaussian curvature.

Geometry Curvature Triangle angle sum Parallels through \(p\)
Spherical \(K > 0\) \(> \pi\) none
Euclidean \(K = 0\) \(= \pi\) exactly one
Hyperbolic \(K < 0\) \(< \pi\) infinitely many

The Gauss–Bonnet theorem makes this quantitative. For a geodesic triangle \(T\) on a surface,

\[\int_T K\, dA \;=\; (\alpha + \beta + \gamma) - \pi.\]

Specializing to constant curvature \(K \equiv -1\), this becomes

\[-\operatorname{Area}(T) \;=\; (\alpha + \beta + \gamma) - \pi \quad\Longrightarrow\quad \operatorname{Area}(T) \;=\; \pi - (\alpha + \beta + \gamma).\]

This is the formula I find most beautiful in the whole subject. In Euclidean geometry, knowing the angles of a triangle tells you nothing about its size — similar triangles exist, scaling is free. In hyperbolic geometry, the angles are the size. Geometry has become rigid.

A consequence: there is no scale-free hyperbolic geometry. Negative constant curvature picks out an absolute unit of length. The map \(z \mapsto \lambda z\) is not an isometry for \(\lambda \neq 1\) except in the very specific way we will see shortly.

The other consequence I never get over: the maximum possible area of a hyperbolic triangle is \(\pi\), achieved when \(\alpha = \beta = \gamma = 0\), i.e. when all three vertices are pushed out to the boundary at infinity. This is an ideal triangle. Three vertices at infinite distance from each other, finite area equal to \(\pi\). Hyperbolic geometry is full of this kind of contradiction-resistant arithmetic.

3. “More room”: exponential volume growth

Negative curvature makes geodesics diverge. Two geodesics shot from a point at slightly different angles spread apart faster than they would in flat space. The space has, in a precise sense, more room.

For Euclidean and hyperbolic 2-space respectively,

\[C_{\mathrm{Euc}}(r) \;=\; 2\pi r, \qquad C_{\mathrm{Hyp}}(r) \;=\; 2\pi \sinh r.\]

For large \(r\), \(\sinh r \sim e^r / 2\), so circumference (and the area of a hyperbolic disk of radius \(r\)) grows exponentially in \(r\).

This is also why every regular tree of degree \(\geq 3\) behaves “morally” like a hyperbolic space: the number of vertices at distance \(\leq r\) from any root grows exponentially in \(r\). Gromov made this precise with \(\delta\)-hyperbolicity — a metric space is hyperbolic in his sense if every geodesic triangle is \(\delta\)-thin, and exponential ball growth is then automatic. Geometric group theory lives here.

It is also the reason hyperbolic embedding is now a topic in machine learning: hierarchical, tree-like data wants exponentially growing volume to embed into without collisions, and Euclidean space simply does not have it.

4. The upper half-plane model

There are several equivalent models of the hyperbolic plane: Klein, Poincaré disk, Poincaré upper half-plane, hyperboloid in Minkowski space. They are mutually isometric, but each highlights different features.

For number theory the right model is the upper half-plane

\[\mathbb{H} \;=\; \{ z = x + iy \in \mathbb{C} : y > 0 \}\]

with metric

\[ds^2 \;=\; \frac{dx^2 + dy^2}{y^2},\]

because the \(SL_2(\mathbb{R})\) Möbius action stares you in the face. A few things to internalize about this metric:

  • It is conformal. It scales the Euclidean metric pointwise by \(1/y^2\) but does not deform angles. Hyperbolic angles equal Euclidean angles. This is enormously useful — you can read angles off pictures without worrying about the distortion.
  • It is incomplete from the bottom. As \(y \to 0\), distances blow up. The real axis is infinitely far away. It is not part of the space. It is the boundary at infinity, often denoted \(\partial \mathbb{H} = \mathbb{R} \cup \{\infty\}\).
  • It is invariant under horizontal translation. The metric does not depend on \(x\), so \(z \mapsto z + a\) is an isometry for any \(a \in \mathbb{R}\).
  • It is invariant under positive vertical scaling. Under \(z \mapsto \lambda z\) for \(\lambda > 0\), both $$ dz \(and\)y\(scale by\)\lambda\(, so\) dz /y$$ is unchanged. This is the only “scaling” that survives, and it is the analogue of translation along a hyperbolic geodesic.

These three facts already pin down a lot of the structure. Everything else follows.

5. Geodesics, by way of the metric

A geodesic in a Riemannian manifold is a locally length-minimizing curve, satisfying the geodesic equation

\[\ddot{x}^k + \Gamma^k_{ij}\, \dot{x}^i \dot{x}^j \;=\; 0.\]

You can grind out the Christoffel symbols for \(ds^2 = (dx^2 + dy^2)/y^2\) and solve, but the cleaner argument uses isometries.

Claim. Vertical Euclidean half-lines \(\{ x = c \}\) are geodesics.

For a vertical curve \(\gamma(t) = (c, y(t))\) from height \(a\) to height \(b > a\), the length is

\[L(\gamma) \;=\; \int \frac{\sqrt{0 + \dot{y}^2}}{y}\, dt \;=\; \int_a^b \frac{dy}{y} \;=\; \log(b/a).\]

For any other curve \(\eta = (x(t), y(t))\) between the same two heights,

\[L(\eta) \;=\; \int \frac{\sqrt{\dot{x}^2 + \dot{y}^2}}{y}\, dt \;\geq\; \int \frac{|\dot{y}|}{y}\, dt \;\geq\; \log(b/a).\]
So vertical lines are length-minimizing. Now use isometries to find every other geodesic. The map \(z \mapsto -1/z\) is an isometry (we will verify this in §7). It fixes the unit semicircle $$ z = 1\(setwise, sends the imaginary axis to itself, and more usefully sends a generic vertical line to a Euclidean semicircle perpendicular to the real axis. Composing with horizontal translations\)z \mapsto z + a\(gives every Euclidean semicircle whose center sits on\)\mathbb{R}$$.

So the geodesics of \(\mathbb{H}\) are exactly:

  1. Vertical Euclidean half-lines \(\{ x = c, \; y > 0 \}\).
  2. Euclidean semicircles whose centers lie on the real axis.

Why must the semicircles meet \(\mathbb{R}\) at right angles? Because the metric is conformal: hyperbolic angles equal Euclidean angles. A geodesic must approach the boundary perpendicularly because of the symmetry under \(z \mapsto -1/z\), which swaps the two ends of the geodesic and forces the limit angle at \(\partial \mathbb{H}\) to be \(\pi/2\).

This is also where Euclid’s fifth postulate visibly fails. Take the vertical line \(L = \{ x = 0 \}\) and a point \(p = 5 + i\). Every Euclidean semicircle through \(p\) that does not meet \(L\) — and there are infinitely many of them — is a hyperbolic line through \(p\) “parallel” to \(L\) in the sense that they never intersect. Parallels are everywhere.

6. Distance, with a derivation

The hyperbolic distance between two points \(z, w \in \mathbb{H}\) is

\[d(z, w) \;=\; \operatorname{arcosh}\!\left( 1 + \frac{|z - w|^2}{2 \operatorname{Im}(z) \operatorname{Im}(w)} \right).\]

The cleanest way to derive this is the cross-ratio formula. Let \(\gamma\) be the geodesic through \(z\) and \(w\), and let \(z^*, w^* \in \partial \mathbb{H}\) be its ideal endpoints, with \(z^*\) on the \(z\)-side. Then

\[d(z, w) \;=\; \log [z^*, z, w, w^*] \;=\; \log \frac{|z - w^*| \, |w - z^*|}{|z - z^*| \, |w - w^*|}.\]

The cross-ratio is a Möbius invariant, and Möbius transformations are the isometries (next section), so this expression is forced — it is the unique Möbius-invariant function that reduces to the right local Euclidean distance near the imaginary axis at \(i\).

Specializing to two points on the imaginary axis \(z_1 = ai, z_2 = bi\) with \(0 < a < b\), the geodesic is the imaginary axis itself, with ideal endpoints \(0\) and \(\infty\), and the formula collapses to

\[d(ai, bi) \;=\; \log \frac{b}{a}.\]

This is the entire story of vertical motion: hyperbolic height is logarithmic Euclidean height. Doubling Euclidean height costs a fixed hyperbolic distance \(\log 2\) regardless of where you start. Hyperbolic geometry is the geometry of multiplicative scale — it converts ratios into distances, the same way the logarithm does.

7. Isometries: \(SL_2(\mathbb{R})\) and how it acts

The orientation-preserving isometry group of \(\mathbb{H}\) is

\[\operatorname{Isom}^+(\mathbb{H}) \;\cong\; PSL_2(\mathbb{R}) \;=\; SL_2(\mathbb{R}) / \{\pm I\}.\]

A matrix \(g = \begin{pmatrix} a & b \\ c & d \end{pmatrix}\) with \(ad - bc = 1\) acts by the Möbius transformation

\[g \cdot z \;=\; \frac{az + b}{cz + d}.\]

Two checks need to be done.

It preserves \(\mathbb{H}\). A direct computation gives

\[\operatorname{Im}\!\left( \frac{az + b}{cz + d} \right) \;=\; \frac{(ad - bc)\, \operatorname{Im}(z)}{|cz + d|^2} \;=\; \frac{\operatorname{Im}(z)}{|cz + d|^2}.\]

So if \(\operatorname{Im}(z) > 0\), the image has positive imaginary part too.

It preserves the metric. Differentiating the Möbius map,

\[\frac{d}{dz}\!\left( \frac{az + b}{cz + d} \right) \;=\; \frac{1}{(cz + d)^2}.\]

So under \(g\),

\[|dz| \;\mapsto\; \frac{|dz|}{|cz + d|^2}, \qquad y \;\mapsto\; \frac{y}{|cz + d|^2},\]
and both numerator and denominator of $$ dz /y\(pick up the same factor\) cz + d ^{-2}$$. The ratio is preserved. The metric is invariant. Done.

\(SL_2\) vs \(PSL_2\): why the quotient by \(\pm I\)

This is a small distinction that confused me for a while, so let me write it out.

The two groups are

\[SL_2(\mathbb{R}) \;=\; \{ g \in M_2(\mathbb{R}) : \det g = 1 \}, \qquad PSL_2(\mathbb{R}) \;=\; SL_2(\mathbb{R}) \,/\, \{ \pm I \}.\]

So \(PSL_2(\mathbb{R})\) identifies \(g\) and \(-g\). Why is that the right group for the geometry?

Because the Möbius action does not see the sign. If you negate every entry of \(g\),

\[\frac{(-a)z + (-b)}{(-c)z + (-d)} \;=\; \frac{-(az + b)}{-(cz + d)} \;=\; \frac{az + b}{cz + d}.\]

So \(g\) and \(-g\) act on \(\mathbb{H}\) by the same transformation. The group of distinct Möbius transformations preserving \(\mathbb{H}\) is therefore \(SL_2(\mathbb{R}) / \{\pm I\} = PSL_2(\mathbb{R})\), not \(SL_2(\mathbb{R})\) itself. \(SL_2(\mathbb{R})\) acts on \(\mathbb{H}\) with kernel \(\{\pm I\}\) — it is a double cover of the actual isometry group.

A short comparison:

Property \(SL_2(\mathbb{R})\) \(PSL_2(\mathbb{R})\)
Definition \(\det = 1\) \(SL_2(\mathbb{R}) / \{\pm I\}\)
Center \(\{\pm I\}\) trivial
Acts on \(\mathbb{H}\) yes, with kernel \(\{\pm I\}\) yes, faithfully
Simple as a Lie group? not quite — has center \(\{\pm I\}\) yes, simple
Fundamental group \(\mathbb{Z}\) \(\mathbb{Z}\) (same Lie algebra)
Useful for spin / metaplectic / cohomology with sign the actual isometries of \(\mathbb{H}\)

So when do you use which?

  • If you only care about transformations of \(\mathbb{H}\), use \(PSL_2\). The sign is invisible.
  • If you care about automorphy factors like \((c\tau + d)^k\) in the definition of a modular form of weight \(k\), you have to use \(SL_2\), because the sign does flip the factor when \(k\) is odd. This is exactly why modular forms of odd weight exist for some congruence subgroups but not for \(SL_2(\mathbb{Z})\) — at level 1, \(-I \in SL_2(\mathbb{Z})\) forces \(f(\tau) = (-1)^k f(\tau)\), which kills all odd-weight forms.
  • If you care about spin structures, metaplectic covers, or representations that are projective on \(PSL_2\) but linear on \(SL_2\), you need \(SL_2\).

The same distinction applies to integer coefficients:

\[SL_2(\mathbb{Z}) \;=\; \{ g \in M_2(\mathbb{Z}) : \det g = 1 \}, \qquad PSL_2(\mathbb{Z}) \;=\; SL_2(\mathbb{Z}) \,/\, \{\pm I\}.\]

\(PSL_2(\mathbb{Z})\) is the actual group of distinct Möbius transformations of \(\mathbb{H}\) with integer coefficients. It is what acts faithfully on the modular surface. But for level structures, modular forms, Hecke operators, and anything involving how matrices multiply with signs intact, you keep the cover and work in \(SL_2(\mathbb{Z})\).

When I write \(\mathbb{H} / SL_2(\mathbb{Z})\) in §9, I mean the quotient by the action — which is the same set as \(\mathbb{H} / PSL_2(\mathbb{Z})\), because \(\pm I\) acts trivially. The notational sloppiness is standard. Just remember: the geometry sees \(PSL\), the algebra often sees \(SL\).

The action is transitive: any point can be sent to any other. The stabilizer of \(i\) turns out to be \(SO(2)\), so we get the homogeneous space description

\[\mathbb{H} \;\cong\; PSL_2(\mathbb{R}) \,/\, PSO(2).\]

There is a sharper decomposition called the Iwasawa decomposition, \(SL_2(\mathbb{R}) = NAK\), with

\[N \;=\; \left\{ \begin{pmatrix} 1 & x \\ 0 & 1 \end{pmatrix} : x \in \mathbb{R} \right\}, \quad A \;=\; \left\{ \begin{pmatrix} \sqrt{y} & 0 \\ 0 & 1/\sqrt{y} \end{pmatrix} : y > 0 \right\}, \quad K \;=\; SO(2).\]

\(N\) acts on \(\mathbb{H}\) by horizontal translation, \(A\) by vertical scaling, and the product \(NA\) acts simply transitively on \(\mathbb{H}\) — the matrix in \(NA\) corresponding to \(z = x + iy\) is precisely the one that sends \(i\) to \(z\). The upper half-plane is the homogeneous space of the solvable Borel subgroup \(NA\), with \(K\) as the stabilizer.

This same Iwasawa decomposition is what powers the theory of automorphic forms. It is not a coincidence that it turns up here.

8. Classification of isometries by trace

Möbius transformations come in three flavors, distinguished by the trace of any representative matrix in \(SL_2(\mathbb{R})\). Let $$t = \operatorname{tr}(g) $$.
  • Elliptic \((t < 2)\): one fixed point in \(\mathbb{H}\), no fixed points on the boundary. Acts as a hyperbolic rotation about the interior fixed point. Example: \(z \mapsto -1/z\) has trace \(0\) and fixes \(i\).
  • Parabolic \((t = 2)\): no fixed points in \(\mathbb{H}\), exactly one fixed point on the boundary \(\partial \mathbb{H}\). Acts by sliding along horocycles — curves that look like horizontal lines (when the fixed point is \(\infty\)) or Euclidean circles tangent to the real axis (otherwise). Example: \(z \mapsto z + 1\) has trace \(2\) and fixes \(\infty\).
  • Hyperbolic \((t > 2)\): no fixed points in \(\mathbb{H}\), two fixed points on the boundary. Acts as translation along the unique geodesic connecting those two boundary points — pushing every other point away from one fixed point and toward the other. Example: \(z \mapsto \lambda z\) for \(\lambda > 0, \lambda \neq 1\) fixes \(0\) and \(\infty\) on the boundary, and acts by translation along the imaginary axis.

This trichotomy is the geometric face of the trace. It generalizes to all real Lie groups (where it is the Jordan decomposition), but on \(\mathbb{H}\) it has a vivid meaning: elliptic = rotation, parabolic = horocycle flow, hyperbolic = geodesic flow.

For arithmetic groups like \(SL_2(\mathbb{Z})\), this classification controls the local structure of the quotient. Elliptic fixed points become orbifold cone points. Parabolic fixed points become cusps. Hyperbolic conjugacy classes correspond to closed geodesics on the quotient surface — and their lengths are exactly the data the Selberg trace formula reads.

9. \(SL_2(\mathbb{Z})\) and the modular surface

Inside \(SL_2(\mathbb{R})\) sits the discrete subgroup

\[SL_2(\mathbb{Z}) \;=\; \left\{ \begin{pmatrix} a & b \\ c & d \end{pmatrix} : a, b, c, d \in \mathbb{Z}, \; ad - bc = 1 \right\}.\]

It is generated by

\[T \;=\; \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix} \quad (z \mapsto z + 1), \qquad S \;=\; \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix} \quad (z \mapsto -1/z).\]

These satisfy \(S^2 = (ST)^3 = -I\), so in \(PSL_2(\mathbb{Z})\) we get the presentation

\[PSL_2(\mathbb{Z}) \;\cong\; \langle S, T \mid S^2 = (ST)^3 = e \rangle \;\cong\; \mathbb{Z}/2 \,*\, \mathbb{Z}/3.\]

A free product. So the modular group is, abstractly, very simple — a free product of two finite cyclic groups. All the richness comes from how it acts on \(\mathbb{H}\).

The standard fundamental domain for the action is

\[\mathcal{F} \;=\; \left\{ \tau \in \mathbb{H} : |\tau| \geq 1, \; -\tfrac{1}{2} \leq \operatorname{Re}(\tau) \leq \tfrac{1}{2} \right\}.\]

Every \(SL_2(\mathbb{Z})\)-orbit on \(\mathbb{H}\) meets \(\mathcal{F}\), and meets it exactly once except on the boundary identifications. The proof is a tidy two-step:

  1. Pick any \(\tau \in \mathbb{H}\). Use a power of \(T\) to move \(\tau\) into the strip $$ \operatorname{Re}(\tau) \leq 1/2$$.
  2. If $$ \tau < 1\(, apply\)S\(. Since\)\operatorname{Im}(-1/\tau) = \operatorname{Im}(\tau)/ \tau ^2\(, this strictly increases the imaginary part. Repeat steps 1 and 2. The imaginary part is bounded above on any\)SL_2(\mathbb{Z})\(-orbit (because\) c\tau + d ^2\(over integer pairs\)(c, d)\(has a positive minimum), so the process terminates inside\)\mathcal{F}$$.

The hyperbolic area of \(\mathcal{F}\) is

\[\operatorname{Area}(\mathcal{F}) \;=\; \int\!\!\int_{\mathcal{F}} \frac{dx\, dy}{y^2} \;=\; \frac{\pi}{3}.\]

This number — \(\pi/3\) — keeps showing up. It is the volume of \(SL_2(\mathbb{Z}) \backslash SL_2(\mathbb{R})\) up to a normalization, the leading constant in the Selberg trace formula for the modular group, and the residue at \(s = 1\) of the relevant Eisenstein series. Once you see it once, you keep seeing it.

The quotient \(\mathbb{H} \,/\, SL_2(\mathbb{Z})\) is the modular surface, technically an orbifold with two cone points (at \(i\) of order 2, fixed by \(S\); at \(\rho = e^{2\pi i / 3}\) of order 3, fixed by \(ST\)) and one cusp at \(\infty\). The cone points are the elliptic fixed points. The cusp is the parabolic fixed point of \(T\).

This space is the moduli of complex elliptic curves. That is the entire reason any of this matters for number theory.

10. Why this is the moduli space of elliptic curves

Brief sketch, since it’s the bridge from this post to whatever I write about modular forms.

A complex torus is \(\mathbb{C} / \Lambda\) where \(\Lambda \subset \mathbb{C}\) is a rank-2 lattice. Up to scaling we can write any lattice as \(\Lambda_\tau = \mathbb{Z} + \mathbb{Z}\tau\) for some \(\tau \in \mathbb{H}\) (choose a basis \(\omega_1, \omega_2\) with \(\operatorname{Im}(\omega_2/\omega_1) > 0\), then divide by \(\omega_1\)).

A change of basis is a matrix in \(GL_2(\mathbb{Z})\), but to keep orientation we restrict to \(SL_2(\mathbb{Z})\). A basis change

\[\begin{pmatrix} \omega_1' \\ \omega_2' \end{pmatrix} \;=\; \begin{pmatrix} d & c \\ b & a \end{pmatrix} \begin{pmatrix} \omega_1 \\ \omega_2 \end{pmatrix}\]

translates, after dividing through by the new \(\omega_1'\), into

\[\tau' \;=\; \frac{a\tau + b}{c\tau + d}.\]

So the same lattice corresponds to many different \(\tau\)’s, all related by \(SL_2(\mathbb{Z})\). We get the bijection

\[\{\text{complex tori up to isomorphism}\} \;\;\longleftrightarrow\;\; \mathbb{H} \,/\, SL_2(\mathbb{Z}).\]

Hyperbolic geometry is what governs the parameter space of elliptic curves over \(\mathbb{C}\). The j-invariant is the function on this quotient that kills all remaining ambiguity. Modular forms are sections of certain line bundles over the same space. None of this language even makes sense without the hyperbolic structure underneath.

This is also where the cusp matters. The cusp at \(\infty\) corresponds to \(\tau \to i\infty\), i.e. \(q = e^{2\pi i \tau} \to 0\), i.e. the lattice \(\mathbb{Z} + \mathbb{Z}\tau\) degenerating into something not a lattice. So the cusp is the boundary of the moduli space — it is what compactifies \(Y(1) = \mathbb{H}/SL_2(\mathbb{Z})\) into the modular curve \(X(1)\).

11. The Poincaré disk model and the Cayley transform

The upper half-plane is the right model when you want to do Fourier expansion at the cusp \(\infty\). But for actually visualizing the geometry — fundamental domains of cocompact groups, Escher’s Circle Limit, ideal triangles — the Poincaré disk $$\mathbb{D} = { w \in \mathbb{C} : w < 1 }$$ with metric
\[ds^2 \;=\; \frac{4 \, (du^2 + dv^2)}{(1 - u^2 - v^2)^2}\]

gives the cleaner picture. The boundary is now a finite circle. Geodesics are circular arcs perpendicular to the unit circle (or diameters). Rotational symmetry around the origin is manifest. The whole hyperbolic plane fits inside a finite Euclidean disk, with the boundary at infinity collapsed to the unit circle.

The two models are isometric via the Cayley transform

\[w \;=\; \frac{z - i}{z + i}, \qquad z \;=\; i \, \frac{1 + w}{1 - w}.\]

This sends \(i \in \mathbb{H}\) to \(0 \in \mathbb{D}\) and the boundary \(\mathbb{R} \cup \{\infty\}\) to the unit circle. It also conjugates \(PSL_2(\mathbb{R})\) acting on \(\mathbb{H}\) to \(PSU(1, 1)\) acting on \(\mathbb{D}\). Pick whichever model makes your problem easier; the geometry is the same.

I keep both pictures in my head. The disk for shapes — fundamental polygons, ideal triangles, the way infinitely many triangles tile a finite-looking region. The half-plane for arithmetic — \(q\)-expansions, cusps, the Iwasawa decomposition, anything where being “near \(\infty\)” is a useful idea.

12. Ideal triangles and area \(= \pi\) in coordinates

It is worth doing the area-of-an-ideal-triangle calculation by hand once, because it is the cleanest sanity check that the metric and Gauss–Bonnet agree.

Take the ideal triangle with vertices at \(0, 1, \infty\). Its three sides are:

  • the vertical line \(x = 0\) (geodesic from \(0\) to \(\infty\)),
  • the vertical line \(x = 1\) (geodesic from \(1\) to \(\infty\)),
  • the upper semicircle $$ z - 1/2 = 1/2\(, i.e.\)y = \sqrt{x(1 - x)}\(for\)0 \leq x \leq 1\((geodesic from\)0\(to\)1$$).

The interior is \(\{ (x, y) : 0 < x < 1, \; y > \sqrt{x(1-x)} \}\), so

\[\operatorname{Area} \;=\; \int_0^1 \!\! \int_{\sqrt{x(1-x)}}^{\infty} \frac{dy}{y^2}\, dx \;=\; \int_0^1 \frac{dx}{\sqrt{x(1-x)}}.\]

Substituting \(u = 2x - 1\) gives

\[\int_{-1}^{1} \frac{du}{\sqrt{1 - u^2}} \;=\; \arcsin(u) \Big|_{-1}^{1} \;=\; \pi.\]

Exactly \(\pi\), as Gauss–Bonnet requires. All three angles are \(0\), all three vertices are infinitely far away, and the triangle has finite area \(\pi\).

The really striking corollary: every hyperbolic triangle, no matter how it is drawn, has area at most \(\pi\). There is a hard upper bound on triangle area, set by the curvature alone. There is no analog of this in Euclidean geometry.

13. The programmer’s mental model

Now that I have done the math, the engineering intuition I keep coming back to is:

Hyperbolic space is what happens when distance is not uniform across the coordinate system.

In code, the only thing the upper half-plane model adds on top of \(\mathbb{R}^2\) is a position-dependent metric tensor

\[g(x, y) \;=\; \frac{1}{y^2} \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}.\]

Inner products at a point \((x, y)\) are

\[\langle v, w \rangle_{(x, y)} \;=\; \frac{v_x w_x + v_y w_y}{y^2}.\]

That single change cascades into everything: lengths, geodesics, isometries, curvature, area. There is no other magic. All the surprising consequences — \(\sinh r\) growth, area equals angle defect, exponential boundary, \(PSL_2(\mathbb{R})\) symmetry — fall out of integrating this one metric. Hyperbolic geometry, from a programmer’s point of view, is just Riemannian geometry with this single tensor.

The deeper observation, which I think is the actual punchline of the upper half-plane story, is the duality between metric and group:

The metric is what makes \(PSL_2(\mathbb{R})\) an isometry group. The group action is what generates the metric. Each determines the other.

That duality — between a geometric structure and the symmetries that preserve it — is Klein’s Erlangen program in miniature. It is also why hyperbolic geometry shows up wherever \(SL_2\) does. Wherever you have \(SL_2\) acting on something, there is a hyperbolic geometry hiding.

14. Where this fits in the bigger picture

Hyperbolic geometry is the natural arena for:

  • the upper half-plane and its compactification (the modular curve \(X(1)\)),
  • congruence covers \(Y_0(N), Y_1(N), Y(N)\) — the modular curves that show up in the Modularity Theorem,
  • modular forms (sections of certain line bundles on these curves) and their Fourier expansions at cusps,
  • Riemann surfaces of genus \(\geq 2\), all of which carry a unique hyperbolic metric by the uniformization theorem,
  • Teichmüller space, with its \(3g - 3\) complex dimensions of hyperbolic structures,
  • Mostow rigidity in dimension \(\geq 3\) (a closed hyperbolic manifold determined up to isometry by its fundamental group),
  • geometric group theory, via Gromov \(\delta\)-hyperbolicity,
  • spectral theory on \(\Gamma \backslash \mathbb{H}\) — Maass forms, Selberg trace formula, Selberg zeta function,
  • and, going wider, the entire automorphic side of the Langlands program.

For my current trajectory — elliptic curves into modular forms into a real understanding of what Wiles actually did — the relevant fragment is just the modular surface and its cusps. But the fact that the same \(\mathbb{H}\) shows up in 3-manifold topology, in Teichmüller theory, in geometric group theory, and in Langlands is what convinced me that this is not a side quest. It is the same space being viewed from different rooms in the same building.

15. Summary

Two formulas to take home:

\[\mathbb{H} \;=\; \{ z \in \mathbb{C} : \operatorname{Im}(z) > 0 \}, \qquad ds^2 \;=\; \frac{dx^2 + dy^2}{y^2}.\]

From them, everything:

  • geodesics are vertical half-lines and Euclidean semicircles perpendicular to \(\mathbb{R}\),
  • triangles have area \(\pi - (\alpha + \beta + \gamma)\), bounded above by \(\pi\),
  • circles have circumference \(2\pi \sinh r\) — exponential growth,
  • the boundary \(\mathbb{R} \cup \{\infty\}\) is at infinite distance,
  • isometries are \(PSL_2(\mathbb{R})\), classified by trace into elliptic, parabolic, hyperbolic,
  • the discrete subgroup \(SL_2(\mathbb{Z})\) has fundamental domain of hyperbolic area \(\pi/3\),
  • and the quotient \(\mathbb{H} / SL_2(\mathbb{Z})\) is the moduli of complex elliptic curves.

The thing I find shocking is how much of the arithmetic geometry of elliptic curves and modular forms is just this — a single Riemannian metric, a single discrete group, and the consequences of those two sharing a space.

I am working through Ratcliffe’s Foundations of Hyperbolic Manifolds alongside Diamond–Shurman’s A First Course in Modular Forms. The first builds the geometry from the ground up. The second uses it. The combination is good.

I love math and systems engineering.