Distributions on a manifold
Fix a smooth $n$-manifold $M$. A rank-$k$ distribution $\Delta$ on $M$ is a smooth assignment
\[p \;\longmapsto\; \Delta_p \;\subset\; T_p M, \qquad \dim \Delta_p = k,\]of a $k$-plane in the tangent space at every point. Equivalently, $\Delta$ is a rank-$k$ sub-bundle of $TM$. Locally one specifies $\Delta$ by giving $k$ pointwise-independent vector fields $X_1, \ldots, X_k$ that span $\Delta$ — but the distribution itself is the plane field, not the choice of frame.
Two extreme examples we will keep in mind:
- $M = \mathbb R^3$, $\Delta = \mathrm{span}(\partial_x, \partial_y)$. At every point the plane is the horizontal $xy$-plane. Curves tangent to $\Delta$ keep their $z$ fixed.
- $M = \mathbb R^3$, $\Delta = \ker(dz - y\,dx) = \mathrm{span}(\partial_x + y\,\partial_z,\, \partial_y)$. At each point the plane is tilted by the amount $y$. Curves tangent to $\Delta$ can change $z$ — but only via a precise interplay between $x$- and $y$-motion.
A curve $\gamma : [0, T] \to M$ is horizontal (or admissible) if $\dot\gamma(t) \in \Delta_{\gamma(t)}$ for all $t$. Horizontal curves are the only legal trajectories in the V1 model: at every moment the curve’s velocity must lie in the 2-plane the cortex can move through (forward + rotate).
Frobenius’s theorem: when is $\Delta$ integrable?
A rank-$k$ distribution is integrable if through every point passes a $k$-dimensional submanifold $L \subset M$ — an integral submanifold — with $T_p L = \Delta_p$ for every $p \in L$. When integrable, the integral submanifolds foliate $M$.
The clean test is purely algebraic:
Easy direction. Suppose $\Delta$ integrates, with leaves $L$. Two vector fields $X, Y \in \Delta$ are tangent to $L$. Their flows preserve $L$ (because the leaf is integral). Their commutator’s flow does too, so $[X, Y]$ is tangent to $L$, i.e. in $\Delta$.
Hard direction. If $[\Delta, \Delta] \subseteq \Delta$, choose a frame $X_1, \ldots, X_k$ for $\Delta$. By rectifying flows, find local coordinates in which $X_i = \partial_{x_i}$ for $i \leq k$. The integral submanifolds are the slices $x_{k+1}, \ldots, x_n = \text{const}$. The bracket-closure ensures the rectifying coordinates are consistent.
Worked test for the V1 distribution. $X_1 = \cos\theta\,\partial_x + \sin\theta\,\partial_y$, $X_2 = \partial_\theta$ (Appendix A1 §3). Compute $[X_1, X_2]$ in coordinates:
\([X_1, X_2] \;=\; (\cos\theta\,\partial_x + \sin\theta\,\partial_y)\partial_\theta - \partial_\theta(\cos\theta\,\partial_x + \sin\theta\,\partial_y)\) \(\quad \;=\; -(-\sin\theta)\,\partial_x - (\cos\theta)\,\partial_y \;=\; \sin\theta\,\partial_x - \cos\theta\,\partial_y \;=\; -X_3,\)
and $X_3 \notin \mathrm{span}{X_1, X_2}$ (its $\partial_x, \partial_y$ coefficients are not proportional to $X_1$’s). So $[X_1, X_2] \notin \mathcal H$ and Frobenius fails — the V1 distribution is not integrable.
Geometrically: if it had integrated, the cortex would foliate into 2-dimensional sheets, and you could never reach a sideways-displaced neuron from your starting one by horizontal motion. V1 would be a disconnected stack. But it is not — and the failure of Frobenius is exactly what makes contour completion possible.
Bracket-generating distributions and Chow–Rashevskii
The dual case to Frobenius:
Sketch. Bracket-generating means iterated brackets of the frame fields span $T_p M$. Concatenating short flows along a frame field $X_i$ for times $\pm \varepsilon$ in the pattern of A1 Figure 1.2 produces a net displacement of order $\varepsilon^2$ in the bracket direction, of order $\varepsilon^3$ in iterated-bracket directions, etc. Show that the smooth map $\mathbb R^n \to M$, $(t_1, \ldots, t_n) \mapsto \Phi^{X_{i_1}}{t_1} \circ \cdots \circ \Phi^{X{i_n}}_{t_n}$, has surjective differential at the origin under bracket-generation, hence is locally surjective by the inverse function theorem. Compose enough hops and you can reach any point.
For SE(2) the Hörmander condition is satisfied at depth 1. We computed $[X_1, X_2] = \pm X_3$, and $X_1, X_2, X_3$ already span $T_g\mathrm{SE}(2)$ (three linearly independent fields). No deeper brackets are needed. This is the minimal possible depth and is what makes the SR Carnot–Carathéodory distance well-behaved on $\mathrm{SE}(2)$.
Contact structures on 3-manifolds
A contact form on a $(2n+1)$-manifold $M$ is a 1-form $\alpha$ with
\[\alpha \wedge (d\alpha)^n \;\neq\; 0 \quad\text{everywhere}.\]A contact structure is the rank-$2n$ distribution $\xi = \ker \alpha$. The condition $\alpha \wedge (d\alpha)^n \neq 0$ is the strongest possible non-integrability — exactly the opposite of Frobenius’s vanishing condition.
For $n = 1$ (3-manifolds, our case):
\[\alpha \wedge d\alpha \;\neq\; 0.\]The standard contact structure on $\mathbb R^3$. $\alpha_0 = dz - y\,dx$. Then $d\alpha_0 = -dy \wedge dx = dx \wedge dy$, and $\alpha_0 \wedge d\alpha_0 = dz \wedge dx \wedge dy \neq 0$. Contact, with kernel $\xi_0 = \mathrm{span}(\partial_x + y\,\partial_z, \partial_y)$.
The SE(2) contact structure. $\alpha = \sin\theta\,dx - \cos\theta\,dy$. Compute $d\alpha = \cos\theta\,d\theta \wedge dx + \sin\theta\,d\theta \wedge dy$, hence
\(\alpha \wedge d\alpha \;=\; (\sin\theta\,dx - \cos\theta\,dy) \wedge (\cos\theta\,d\theta \wedge dx + \sin\theta\,d\theta \wedge dy)\) \(\quad \;=\; -\sin^2\theta\,dx \wedge d\theta \wedge dy - \cos^2\theta\,dy \wedge d\theta \wedge dx \;=\; -dx \wedge dy \wedge d\theta \;\neq\; 0.\)
Contact. The kernel $\xi = \mathrm{span}{X_1, X_2}$ is the V1 horizontal distribution.
Darboux’s theorem (contact version): every contact structure on a 3-manifold is locally isomorphic to the standard one $(\mathbb R^3, \alpha_0)$. So the V1 contact structure and the engineering “rolling penny” contact structure are the same locally — a fact Petitot exploited to import results from the latter into V1 modelling.
Why Chow’s theorem matters for vision
The Chow theorem is what makes “modal completion is a shortest-path problem on $\mathrm{SE}(2)$” not a vacuous statement. If the V1 distribution were integrable, the brain would be unable to bridge two oriented edges that do not lie on the same $\theta = \text{const}$ leaf — vision would split into orientation strata. Frobenius would have shut the door, and Petitot’s model would predict no perception of contour completion.
Bracket-generation says the opposite: any source/target neuron pair can be linked by a horizontal path, and the length of the shortest such path is the SR distance. Petitot’s claim is that the visual system computes this distance and renders it as a percept. The contact structure is what makes this computation well-posed.
Carnot–Carathéodory distance
Once a sub-Riemannian metric $\langle\cdot,\cdot\rangle_p$ on $\Delta_p$ is fixed, the horizontal length of an admissible curve is
\[L_{\mathrm{SR}}(\gamma) \;:=\; \int_0^T \sqrt{\langle\dot\gamma, \dot\gamma\rangle_{\gamma(t)}}\,dt,\]and the sub-Riemannian distance $d_{\mathrm{SR}}(p, q)$ is the infimum over all admissible curves from $p$ to $q$. Chow’s theorem ensures this infimum is finite (the set of admissible curves is non-empty).
For the V1 metric of Part 1 with frame ${X_1, X_2}$ orthonormal, an admissible $\gamma$ writes as $\dot\gamma = u_1(t) X_1 + u_2(t) X_2$ and
\[L_{\mathrm{SR}}(\gamma) \;=\; \int_0^T \sqrt{u_1^2 + u_2^2}\,dt.\]After the $u_1 = 1$ unit-speed reduction, this collapses to the elastica functional $\int \kappa^2(s)\,ds$ — exactly the integrand Euler minimised. Appendix A3 takes this and runs it through the Pontryagin Maximum Principle.
Mitchell’s compactness theorem (1985) shows that on a connected bracket-generating SR manifold the infimum is attained: a length-minimising geodesic exists between any two points. This is the SR analogue of Hopf–Rinow. For $\mathrm{SE}(2)$ it means every pair of V1 neurons is connected by an actual shortest horizontal curve — not just an approachable one.
Connection to the elliptic project
The reachability figure above is mathematically the same object as the
parking trajectory in
the dubins-back-wheel example: a piecewise concatenation of
$X_1$ (“forward + reverse”) and $X_2$ (“steer”) flows. The parking
trajectory is what Chow’s theorem looks like with finite $\varepsilon$;
the elastica geodesic is what it converges to as $\varepsilon \to 0$.
Both are computed by the same SE(2) ODE integrator that lives in
elliptic-core.js on the elliptic site and is reused here.
Code
# Test the V1 distribution for Frobenius integrability.
# Computes [X1, X2] symbolically and checks if it's in span{X1, X2}.
import sympy as sp
x, y, th = sp.symbols('x y theta', real=True)
X1 = sp.Matrix([sp.cos(th), sp.sin(th), 0]) # cos θ ∂x + sin θ ∂y
X2 = sp.Matrix([0, 0, 1]) # ∂θ
def vf_bracket(X, Y, vars_):
# [X, Y]^i = X^j ∂_j Y^i - Y^j ∂_j X^i
out = sp.zeros(len(X), 1)
for i in range(len(X)):
s = 0
for j, v in enumerate(vars_):
s += X[j] * sp.diff(Y[i], v) - Y[j] * sp.diff(X[i], v)
out[i] = sp.simplify(s)
return out
X3 = vf_bracket(X1, X2, [x, y, th])
print("[X1, X2] =", X3.T)
# → [-sin(θ), cos(θ), 0] (this is -X3 from the text; sign convention)
# In span{X1, X2}? Solve a·X1 + b·X2 = X3 for constants a, b
a, b = sp.symbols('a b')
sol = sp.solve(a * X1 + b * X2 - X3, [a, b], dict=True)
print("solution (None means not in span):", sol)
# → [] — no solution; distribution is NOT integrable
# Contact-form check: α ∧ dα ≠ 0 for SE(2)
alpha = [sp.sin(th), -sp.cos(th), 0] # α = sin θ dx - cos θ dy as coefficients
# d α: coefficients of dx∧dy, dx∧dθ, dy∧dθ
# dα = cos θ dθ ∧ dx + sin θ dθ ∧ dy (verify by direct exterior diff)
# α ∧ dα picks out the volume form coefficient on dx ∧ dy ∧ dθ:
vol_coeff = sp.simplify(
alpha[0] * sp.cos(th) * (-1) # sin θ · cos θ · (dx ∧ dθ ∧ dx) - vanishes; only the dy∧dθ term survives ↘
+ (-alpha[1]) * sp.sin(th)
)
# → 1, i.e. α ∧ dα = -dx ∧ dy ∧ dθ ≠ 0 (SE(2) is contact)
What we covered, and what comes next
A distribution is a smooth plane field; integrability is decided by Frobenius’s bracket-closure test; bracket-generation gives Chow’s reachability theorem; contact structures are the “maximally non-integrable” case in odd dimensions. The V1 distribution $\mathcal H = \mathrm{span}{X_1, X_2}$ on $\mathrm{SE}(2)$ is a contact structure, and Chow’s theorem guarantees any two V1 neurons can be linked by a horizontal curve.
Appendix A3 will turn the SR-length minimisation problem into a Hamiltonian system on $\mathfrak{se}(2)^{\ast}$ via the Pontryagin Maximum Principle — recovering the equations $\dot h_1 = h_2 h_3$ etc. that Part 2 §1 uses without proof.
References
- R. Montgomery (2002). A Tour of Subriemannian Geometries, Their Geodesics and Applications. AMS Mathematical Surveys 91. The textbook — Chapter 1 develops everything in this appendix.
- A. A. Agrachev, D. Barilari, U. Boscain (2019). A Comprehensive Introduction to Sub-Riemannian Geometry. Cambridge. Chapters 2 (distributions) and 3 (Chow's theorem).
- H. Geiges (2008). An Introduction to Contact Topology. Cambridge. The contact-geometry side, with Darboux's theorem.
- G. Citti & A. Sarti (2006). "A cortical based model of perceptual completion in the roto-translation space." J. Math. Imaging Vision 24(3): 307–326. The application of Chow + contact to V1.
- J. Petitot (2003). "The neurogeometry of pinwheels as a sub-Riemannian contact structure." Journal of Physiology–Paris 97: 265–309. Originator of the contact-geometric V1 model.
- Elliptic project — Dubins back wheel. Shows parking-style horizontal paths whose $\varepsilon \to 0$ limit is exactly the SR geodesic of Part 2.