Applied Conformal &
Quasiconformal Geometry

Consider two rectangles in the complex plane:

Question: Does there exist a conformal map $f \colon R_{1} \to R_{2}$ that maps the four vertices of $R_{1}$ to the corresponding four vertices of $R_{2}$?

Before demanding strict conformality, let us weaken the condition and ask if there exists a diffeomorphism satisfying the identical boundary conditions.

Observe that the simple linear transformation

is trivially a valid diffeomorphism from $R_{1}$ to $R_{2}$ mapping the vertices to the corresponding vertices.

Using complex coordinates $z = x + iy$, we can express this affine stretch mapping as:

The map $f$ is conformal if and only if the coefficient of $\bar{z}$ entirely vanishes. This holds if and only if $a'/a = b'/b$.

We next ask if there exists any conformal mapping from $R_{1}$ to $R_{2}$ satisfying the boundary conditions.

Suppose such a conformal map $f$ exists. By the Schwarz reflection principle, $f$ extends across the boundaries of $R_{1}$ via repeated reflections to a conformal automorphism of the entire complex plane $\mathbf{C}$.

Every conformal automorphism of $\mathbf{C}$ possesses the exact linear form $f(z) = \alpha z + \beta$ for complex constants $\alpha, \beta$ with $\alpha \neq 0$.

Assuming without loss of generality that $f(0) = 0$, we have $f(z) = \alpha z$. Since $f$ maps the vertices of $R_{1}$ to the corresponding vertices of $R_{2}$, it must satisfy $f(a) = a'$ and $f(ib) = ib'$.

Evaluating the boundary maps yields $\alpha a = a'$ and $\alpha i b = i b'$, which implies:

Thus, a conformal map mapping the vertices exists if and only if the aspect ratios match.

When $a'/a \neq b'/b$, a conformal mapping satisfying these boundary conditions absolutely does not exist.

For accurate facial tracking, we must ensure exact mapping correspondence at numerous landmark points (e.g., boundaries of the eyes, tip of the nose, and corners of the mouth).

Typically, we have $N \gg 3$ such landmarks. We want a mapping $f$ such that $f(z_k) = w_k$ for $k = 1, \dots, N$.

If we require exact mapping correspondence at $N \gg 3$ landmark points, purely conformal mappings are vastly too rigid.

To successfully force alignment across all $N$ landmarks, we must relax the strict conformality condition and seek a quasiconformal mapping.

The objective becomes finding a mapping that satisfies all landmark constraints while minimizing geometric distortion. This is the modern analog of solving Grötzsch's extremal problem for discrete complex structures!

Let $D \subseteq \mathbf{C}$ be a domain and $f \colon D \to \mathbf{C}$ a $C^{1}$-diffeomorphism. The total differential of $f$ in complex coordinates is expressed as:

where the Wirtinger derivatives are $f_{z} = \frac{1}{2}(\partial_{x}f - i \partial_{y}f)$ and $f_{\bar{z}} = \frac{1}{2}(\partial_{x}f + i \partial_{y}f)$.

A direct algebraic computation shows that the Jacobian determinant of the mapping $f$ is given by:

We assume $f$ is strictly orientation-preserving. This means $J_{f} > 0$ everywhere, implying $|f_{z}| > |f_{\bar{z}}|$ at every point in $D$.

One could always ask: what is the image of a circle under a general real-linear transformation?

Consider $T(z) = a z + b \bar{z}$. Evaluating on the unit circle $z = e^{i\theta}$:

By factoring out a phase and applying Euler's formula, we can rewrite this image curve.

The parametrization reduces to:

This describes an ellipse with major axis $|a| + |b|$ and minor axis $|a| - |b|$.

By demanding that this geometric distortion is bounded everywhere, we obtain the analytic definition.

The infimum over all such bounds $K$ is the precise quasiconformal constant $K_f$. A mapping is strictly $1$-quasiconformal if and only if it is conformal.

The geometric distortion $K_f(z)$ is related by $K_{f}(z) = \frac{1 + |\mu_{f}(z)|}{1 - |\mu_{f}(z)|}$. Geometrically, $f$ transforms infinitesimal circles into ellipses with eccentricity determined by the magnitude $|\mu_{f}|$.

To understand quasiconformality globally, we analyze its effect on the module of quadrilaterals.

We define the module of the quadrilateral as the ratio $M(Q) = a/b$.

Conformal mappings preserve the module of any quadrilateral. Quasiconformal mappings, by contrast, distort this module, but by a bounded amount.

This aligns with our analytic definition!

Let us examine a concrete, non-trivial quasiconformal mapping: the radial stretch map.

where $\alpha > 0$ and $\alpha \neq 1$. This map preserves the origin and scales points radially depending on their distance from the origin. Is this mapping quasiconformal?

We utilize the identity $|z| = (z \bar{z})^{1/2}$. Consequently, $f(z) = z (z \bar{z})^{(\alpha - 1)/2}$. Computing the Wirtinger derivatives using the product and chain rules yields:

The Beltrami coefficient $\mu_f(z)$ is the direct ratio of these derivatives:

Taking the magnitude, we find $|\mu_f(z)| = \left| \frac{\alpha - 1}{\alpha + 1} \right|$.

Because $|\mu_f(z)|$ is bounded by a constant less than 1, the mapping is quasiconformal!

Using the formula $K = \frac{1+|\mu|}{1-|\mu|}$, the maximal distortion is :

This illustrates how a non-conformal distortion manifests mathematically.

Armed with the theory of quasiconformal mappings, we can now make precise mathematical sense of Grötzsch's problem: we seek the extremal quasiconformal mapping $f \colon R_{1} \to R_{2}$ that minimizes the maximal distortion $K_{f}$. Assume without loss of generality that $a/b \leq a'/b'$.

Any such mapping must map the left and right vertical edges of $R_1$ to the corresponding edges of $R_2$. Consequently, the length of the image of any horizontal segment $[0, a] \times \{y\}$ must be at least $a'$.

We can formalize this geometric fact by directly integrating the total differential along a horizontal line:

Next, we integrate this inequality with respect to the vertical coordinate $y$ from $0$ to $b$:

This provides a strict lower bound for the integral of the maximal stretch factor over the entire rectangle.

To proceed, we apply the continuous Cauchy-Schwarz inequality for integrals. By carefully separating the integrand into two factors:

The second integral is the Jacobian determinant $J_f$. Integrating the Jacobian over $R_1$ yields the area of the image rectangle $R_2$, which is $a'b'$.

The integrand of the first factor is precisely the local geometric distortion $K_f(z)$. Because we assumed $f$ is a $K$-quasiconformal mapping, $K_f(z) \leq K$ everywhere.

Replacing the integrals with these rigorous bounds reduces the inequality to $(a' b)^{2} \leq (K \cdot ab)(a' b')$. Rearranging gives:

Defining the modules $m = a/b$ and $m' = a'/b'$, we get $m'/m \leq K$. This lower bound is attained by the affine transformation introduced earlier, proving it minimizes the maximal distortion!

We have established that every quasiconformal mapping possesses a bounded Beltrami coefficient $\mu_f$. The converse question asks: given a valid coefficient $\mu$, does there exist a quasiconformal mapping $f$ producing it?

This requires solving the Beltrami Equation:

This PDE is definitively solved by the Measurable Riemann Mapping Theorem.

The Measurable Riemann Mapping Theorem guarantees that Beltrami coefficients $\mu$ (the unit ball in $L^{\infty}$) parameterize all quasiconformal deformations of complex structures.

For a given topological surface $S$, the Teichmüller space $\mathcal{T}(S)$ parameterizes all possible complex structures on $S$ up to isotopy, constructed by taking this infinite-dimensional space of Beltrami coefficients modulo the action of diffeomorphisms isotopic to the identity.

Generalizing the Grötzsch problem for rectangles to arbitrary Riemann surfaces, Teichmüller's Theorem provides a powerful existence and uniqueness result.

A central problem in computer vision and graphics is mapping one complex 3D surface to another—for example, registering a 3D scan of a specific human face onto a standardized character model.

Because these two surfaces possess fundamentally different complex structures, a strictly conformal (distortion-free) mapping between them is mathematically impossible. The geometry must be stretched.

Unlike volumetric data (e.g., MRI scans) which require 3D voxel grids, a human face is a 2D manifold embedded in $\mathbb{R}^3$.

Consequently, we do not discretize it using voxels. Instead, we represent the surface as a triangular mesh $\mathcal{M} = (V, E, F)$, composed of vertices $V$, edges $E$, and triangular faces $F$.

To apply complex analysis to $\mathcal{M}$, we must define a discrete complex structure.

Every Euclidean triangle in $F$ can be mapped isometrically to the complex plane. This induces a piecewise-linear complex coordinate chart across the mesh, allowing us to evaluate angles and Beltrami coefficients at the discrete level.

On a triangular mesh, we assume the mapping $f$ is piecewise linear, implying its derivatives $f_z$ and $f_{\bar{z}}$ are constant on each face $T \in F$.

Thus, the Beltrami coefficient $\mu(z)$ becomes a discrete constant $\mu_T$ per triangle. Finding the map $f$ given a target $\mu$ reduces to solving a sparse linear system via Finite Element Methods (FEM).

Every unique discrete facial topology corresponds to a distinct point in the finite-dimensional Teichmüller space $\mathcal{T}(S)$.

When an algorithm registers one face to another, it is computing a mapping between two points in this parameter space. Since there are infinitely many ways to stretch a surface, a rigorous numerical criterion is required.

Teichmüller's theorem guarantees the existence of exactly one extremal quasiconformal mapping that minimizes the maximum geometric distortion.

This theorem provides numerical registration algorithms with a strictly convergent mathematical target, ensuring the resulting facial mapping is uniquely defined and as distortion-free as theoretically possible.