riemanifold 1.0.2

A package for performing geometric computations on Riemannian manifolds

RieManifold is a framework for Riemannian geometry, powered under the hood by PyTorch's computational graph and automatic differentiation system.

Installation

The package is available through the pip package manager. Run the following command to install it on your machine:

pip install riemanifold

Functionality & Usage

[!NOTE]
Because RieManifold uses PyTorch's automatic differentiation for computations, all user-entered numerical operations must be torch-compatible, such as torch.sin for sine. See a comprehensive list of available PyTorch operations here.

Specifying a manifold

With RieManifold, manifolds to perform computations on may be specified either extrinsically, via a concrete embedding into euclidian $m$-space, or intrinsically, by merely specifying the metric tensor field on a coordinate system.

Method 1: Extrinsic Definition

Formally, let $U\subset\mathbb{R}^n$ be a coordinate domain. Then a differentiable map $\varphi : U \to \mathbb{R}^m$ constitutes an embedding of an $n$-manifold into $m$-space. In this case, the manifold is endowed with natural pullback of the euclidian metric on $\mathbb{R}^m$ to define its intrinsic geometry.
The following shows how to construct a 2-sphere embedded into 3-space:

class UVSphere(EmbeddedRiemannianManifold):
    embedding_dim = 3
    def __init__(self, r: float):
        self.r = r

    coordinate_domain = [0.0, 2*torch.pi], [0.0, torch.pi]
    default_subdivisions = 41, 41
    
    def embedded(self, coords):
        u, v = coords
        return (
            self.r * torch.sin(v) * torch.cos(u),
            self.r * torch.sin(v) * torch.sin(u),
            self.r * torch.cos(v)
        )

As shown in the example, these class attributes and methods must be supplied:

Attribute / Method	Type	Meaning
embedding_dim	int	the dimension in which to embed the manifold (2 and 3 supported)
coordinate_domain	$n\times 2$-tuple	the bounds of the coordinate domain as a rectangular subset of $\mathbb{R}^n$
default_subdivisions	$n$-tuple	The amount of subdivisions across each coordinate axis when plotting
embedded	$n$-tuple $\to$ $m$-tuple	the parametric embedding function

[!IMPORTANT]
Both ways of specifying a manifold require the choice of a single coordinate system across the entire manifold. This simplification from the theoretical model of smooth manifolds, which allows a collection of coordinate charts (called atlas), was put in place to avoid excessive complication of the interface. It is important to be aware of, however, that some surfaces such as the 2-sphere do not admit a single coordinate system that is everywhere nondegenerate (e.g. the poles in standard coordinates), which may cause errors with computations involving these points.

Method 2: Intrinsic Definition

It is also possible to define geometric concepts such as lengths, angles and curvature entirely intrinsically, through a mathematical object called a Riemannian metric. Formally, a Riemannian metric on a coordinate domain $U\subset\mathbb{R}^n$ is a smoothly varying covariant 2-tensor field on $U$ representing the local inner product for the tangent space at each point. The inner product at each point $p$ is a bilinear form $g_p : T_pM \times T_pM \to \mathbb{R}$ called the metric tensor and is represented by an $n\times n$ matrix. This matrix, as a function of the coordinates, is what the user must specify for the intrinsic manifold specification method.

The following example shows how to create the same sphere manifold as shown above, but purely intrinsically:

class UVSphere(EmbeddedRiemannianManifold):
    def __init__(self, r: float):
        self.r = r
    
    def g(self, coords):
        u, v = coords
        return (
            (torch.sin(v).square() * (self.r ** 2), 0.0),
            (0.0, self.r ** 2)
        )

Here, the only method that needs to be implemented is the function returning the metric tensor g at the coordinates coords. In the example above, for instance, the method returns the standard metric on the 2-sphere in spherical coordinates (degenerate at the poles, as mentioned above):

  g_{\text{sphere}} = \begin{bmatrix} 
    r^2 & 0 \\ 
    0 & sin^2(r)r^2
  \end{bmatrix}

Specifying curves on the manifold

Curves that represent simple linear interpolations in coordinate space can be constructed with the coord_lerp utility method:

curve = coord_lerp([0.0, torch.pi/8], [-torch.pi/2, torch.pi/2])

Custom curves must be supplied as lambdas/callables accepting a single time argument. By convention, the domain of interest should be the unit interval, so as to match the signature $\lambda : [0, 1] \to U$.

coordinate_path = lambda t: torch.stack([0.25*t, t*torch.pi])

Plotting the manifold and tangential vectors

Embedded manifolds can be plotted, along with geometric objects on them. manifold.show() just displays the manifold, while manifold.show_curve(<curve>), manifold.show_point(<point>) and manifold.show_tangential_vector(<point>, <vector>) can be used to display basic geometry on the embedded manifold. All arguments are interpreted in coordinate space.

Geometric computations

The RieManifold package provides the following additional methods to perform geometric computations:

Method	Purpose
compute_curve_length(curve)	Compute the intrinsic length of a given curve in coordinate space
Gamma(coords)	Compute Christoffel symbols at the given coordinates
parallel_transport(curve, vector)	Paralell transport a vector along a given curve, starting at curve(0.0)
is_geodesic(curve)	Determine if a given curve is a Geodesic on the manifold
source_geodesic(start_coords, initial_veloity)	Create a geodesic by following a straight path on a manifold, given a starting point and an initial velocity vector in coordinate space
R(coords)	Compute the Riemannian curvature tensor at the given coordinates
Ric(coords)	Compute the Ricci curvature tensor at the given coordinates
ric(coords)	Compute the Ricci curvature scalar at the given coordinates
gaussian_curvature(coords)	Compute the Gaussian curvature at the given coordinates (only defined for 2D manifolds)

Additionally, the methods for parallel transport and source geodesics provide show<...>() methods to plot the resulting geometric objects.

Getting started

To get started, there are some out-of-the-box embedded manifolds available to tinker with, which are listed below:

Manifold Name & Parameters	Plot
UVSphere(r) r: the radius of the sphere
Torus(ro, ri) ro: the outer radius ri: the inner radius
MobiusStrip(radius, width) radius: the base radius width: the width of the band
CartesianPlane()
PolarPlane()