Riemannian Metrics

A Riemannian metric is a smoothly varying inner product that turns local velocities into lengths, angles, and shortest paths on curved spaces.

The metric ellipse at each point shows which directions count as long or short

Why Does This Exist?

Once you know the feasible directions (tangents), you still cannot say which step is large: a unit coordinate move can mean a tiny or huge behavior change depending on position. Without a ruler, step sizes are meaningless and shortest paths undefined.

The Riemannian metric installs that ruler, varying smoothly from point to point. It powers geodesics, volume, and the Fisher geometry of statistics. Prerequisites: manifolds, tangent spaces. Application: Fisher metric.

The one idea you need first

An inner product $\langle u, v\rangle$ gives lengths $\|v\| = \sqrt{\langle v,v\rangle}$ and angles. A metric assigns one such product $g_p$ to every tangent plane, changing with $p$ .

Think of It Like This

Wind that changes with altitude

Two identical strides cost different effort depending on altitude: thin air up high, thick air below. A fixed pedometer lies about the workout; a wind-adjusted meter tells the truth.

The metric is that adjustment card, posted at every point. It stretches the ruler where motion is expensive and shrinks it where motion is cheap, so path lengths reflect real cost.

Where the analogy stops: metrics must vary smoothly and stay positive definite, or lengths and angles break.

How It Actually Works

In coordinates, $g$ is a symmetric positive-definite matrix $G(x)$ and $\|v\|^2 = v^T G(x) v$ . Curve length integrates $\int \sqrt{\dot{\gamma}^T G(\gamma)\dot{\gamma}}\,dt$ . The metric also gives volume $d\text{vol} = \sqrt{\det G}\,dx$ and converts covectors to gradients via $G^{-1}$ .

Worked example

Scaled metric on $\mathbb{R}^2$ : $G = \text{diag}(1, 4)$ , so vertical moves cost double:

Vector $(1, 0)$ : length $1$ . Vector $(0, 1)$ : length $\sqrt{4} = 2$ .
Curve A (straight): $(0,0) \to (0,2)$ has length $2 \times 2 = 4$ .
Curve B (detour): $(0,0) \to (2,0) \to (2,2)$ has length $2 + 4 = 6$ . Straight wins here.
Under $G = \text{diag}(9, 1)$ the ranking flips: A costs $2$ , B costs $6 + 2 = 8$ ? No: horizontal legs cost $3$ per unit, so B $= 6 + 2 = 8$ vs A $= 2$ . Metric decides the shortest route, not coordinates.

Watch Out For

Averaging metric matrices like plain numbers

The mean of two positive-definite matrices is positive definite, but naive Euclidean averaging distorts distances and volumes badly under strong anisotropy.

Use affine-invariant or Log-Euclidean means when the metric itself is the object being averaged.

Ignoring the volume factor in sampling

Uniform coordinate sampling oversamples regions where $\sqrt{\det G}$ is small. Density estimates and Monte Carlo integrals come out biased.

Weight by the volume element or parametrize by an isometric chart when sampling manifolds.

The Quick Version

A Riemannian metric puts a smoothly varying inner product on every tangent plane.
Lengths, angles, gradients, and volumes all derive from that one object.
Coordinate-identical vectors can have very different true lengths.
Shortest paths depend on the metric, not on coordinate straightness.
The Fisher matrix is the canonical metric for statistical families.