Product Manifolds
Product manifolds
Definition
Let \((\mathcal{M}_1,\dots,\mathcal{M}_k)\) be smooth manifolds. Their Cartesian product
admits a natural smooth-manifold structure called the product manifold.
Charts and dimension
If \((U_i,\varphi_i)\) is a chart on \(\mathcal{M}_i\) with \(\varphi_i:U_i\to\mathbb{R}^{n_i}\), then
defines a chart on \(\mathcal{M}\) via
Hence,
The canonical projections \(\pi_i:\mathcal{M}\to\mathcal{M}_i\), \(\pi_i(x_1,\dots,x_k)=x_i\), are smooth.
Tangent space: direct sum structure
At a point \(x=(x_1,\dots,x_k)\in\mathcal{M}\), the tangent space splits as
Intuitively, an “infinitesimal motion” on the product is just a tuple of infinitesimal motions on each factor. In coordinates, derivatives and Jacobians often become block-structured, mirroring this direct-sum decomposition.
Riemannian product metric
If each \(\mathcal{M}_i\) has a Riemannian metric \(\langle\cdot,\cdot\rangle_i\), a common choice on the product is
Lie-group special case
If each \(\mathcal{M}_i\) is a Lie group \(G_i\), then the product \(G=G_1\times\cdots\times G_k\) is a Lie group with component-wise multiplication:
Its Lie algebra is the direct sum
and exponential/log maps act componentwise:
In particular, since \(\mathbb{R}^n\) is a Lie group under addition, the common state space
is (also) a Lie group under \((R,a)\cdot(S,b)=(RS,a+b)\), with Lie algebra \(\mathfrak{so}(3)\oplus\mathbb{R}^n\).
Engineering view: “mixed” state vectors in flight dynamics
Many GNC/flight-sim states are naturally tuples of heterogeneous components, e.g.
The product-manifold viewpoint says:
You should update each component using the correct geometry.
Euclidean components update by addition; rotational components update by group composition (or a retraction).
A convenient way to formalize this is with “box-plus / box-minus” operators. Let \(\oplus_i\) be the update rule for \(\mathcal{M}_i\) (e.g. \(+\) on \(\mathbb{R}^n\), left-multiplication by \(\exp(\cdot)\) on \(SO(3)\)). Then define for the product:
and similarly for \(\ominus\) component-wise.
This is exactly what error-state EKFs and Lie-group / RKMK integrators exploit: propagate on the product manifold, and linearize in the (direct-sum) tangent space.
Why “separate” \(\mathbb{R}^n\) even though it is a Lie group?
You can treat everything as a Lie group if you want (e.g. \(SO(3)\times\mathbb{R}^n\)). In practice we keep \(\mathbb{R}^n\) as “plain Euclidean” because:
its group operation is already \(+\), so the manifold machinery adds no complexity where it is unnecessary;
linearization and covariance bookkeeping are simplest on Euclidean blocks;
non-Euclidean parts (e.g. \(SO(3)\), \(SE(3)\)) are where constraints and curvature matter (e.g. quaternion normalization, avoiding singular charts).
The product-manifold framing cleanly supports both: treat curved components with proper group/retraction updates, and treat Euclidean ones with standard addition.