Sub-Riemannian geometry

In our daily lives, we are constantly navigating a world of constraints. A car cannot instantly slide sideways, a rolling coin is bound by its orientation, and even our own body's movements are not entirely free. While these limitations seem restrictive, they give rise to a surprisingly rich and counter-intuitive geometric structure. Sub-Riemannian geometry is the mathematical framework that studies these very systems, moving beyond simple straight-line distances to explore the true nature of paths in a constrained world. This article addresses a central paradox: how can systems with limited instantaneous motion still achieve full freedom of movement, and what does this mean for our understanding of distance, dimension, and control?

This exploration is divided into two parts. In the first chapter, ​​Principles and Mechanisms​​, we will demystify the "magic" of constrained motion, introducing the Lie bracket as the mathematical tool for generating new directions and the Carnot-Carathéodory distance as the natural way to measure paths in this crooked world. We will uncover how these principles lead to a startling re-evaluation of spatial dimension itself. Following this, the chapter on ​​Applications and Interdisciplinary Connections​​ will journey through diverse fields—from robotics and control theory to probability and neuroscience—revealing how this powerful geometry provides a unifying language for describing everything from a parallel-parking robot to the functional architecture of our own visual system. Let us begin by examining the core principles that allow us to overcome the impossible.

Imagine you're trying to navigate a vast, open warehouse floor, but with a peculiar set of rules. Perhaps you're on an ice skate, which can glide forward and backward with ease, and pivot on the spot, but absolutely cannot slide sideways. Or maybe you're driving a car, which can only move in the direction its wheels are pointing. In both cases, your movement is ​​constrained​​. You can't just move in any direction you please at any given moment. The set of allowed velocity vectors at any point forms a limited subspace—in these cases, a 2D plane within the 3D space of all possible positions and orientations. In the language of geometry, this set of allowed directions at every point is called a ​​distribution​​.

You might think that if you're only ever allowed to move within these 2D planes, you'd be stuck forever gliding along some 2D surface carved out of the larger 3D space. If you start on a giant sheet of paper, you can never leave it. Sometimes, this is true. If the planes of allowed motion fit together perfectly smoothly, like the fibers in a sheet of wood, they are called ​​integrable​​. In this case, your intuition is correct: you are trapped on a lower-dimensional "leaf" within the larger space, and there are places you can see but never reach.

But more often than not, something truly magical happens.

Think about parking a car. A car cannot drive directly sideways into a parking spot. The wheels simply don't allow it. Yet, every licensed driver knows how to perform a parallel parking maneuver: a sequence of moving forward, turning, moving backward, and turning again. Each individual movement respects the constraints, but the combination of these movements produces a net motion in a "forbidden" direction—sideways!

This is the central miracle of sub-Riemannian geometry. Even when your instantaneous movements are strictly limited, you can combine them to generate motion in every direction. The constraints are not as confining as they first appear. The distribution is ​​non-integrable​​. The planes of allowed motion are twisted in such a way that they don't lock you onto a single surface. Instead, by cleverly "wiggling" back and forth between the allowed directions, you can effectively "push off" the constraints to explore the entire space.

But how does this work? And how can we describe this "wiggling" mathematically? The answer lies in a beautiful concept from differential geometry: the Lie bracket.

Let's represent our allowed motions as vector fields. A vector field is simply an arrow at every point in space that tells you the direction and speed of a possible motion. For our unicycle example, we have two fundamental controls: "roll forward" (let's call its vector field $f_1$) and "turn on the spot" ($f_2$).

If these two motions were independent in a simple, commuting way—like moving east and moving north on a grid—then combining them would be straightforward. Go east, then north, and you end up in the same place as going north, then east. The order wouldn't matter.

But rolling and turning don't commute. Try it: roll forward, then turn left. You end up in a different spot and orientation than if you first turn left, then roll forward. The ​​Lie bracket​​, denoted $[f_1, f_2]$, is a mathematical tool that precisely measures this failure to commute. It's the ghost of the displacement that emerges from the sequence: "a little bit of $f_1$", then "a little bit of $f_2$", then "a little bit of $-f_1$" (backward), then "a little bit of $-f_2$" (turning back). If the motions commuted, this sequence would bring you back to your exact starting point. But because they don't, you end up slightly displaced. And the direction of that tiny net displacement is exactly the direction of the Lie bracket vector, $[f_1, f_2]$.

For the unicycle model, a beautiful calculation confirms our intuition. Let the state be $(x, y, \theta)$, where $(x,y)$ is the position and $\theta$ is the heading.

• The "roll forward" vector field is $f_1 = (\cos\theta, \sin\theta, 0)$.
• The "turn in place" vector field is $f_2 = (0, 0, 1)$.

Neither of these vectors can move the unicycle sideways relative to its current orientation. A direct calculation of their Lie bracket gives a new vector field: $[f_1, f_2] = (\sin\theta, -\cos\theta, 0)$ This new vector represents a pure translation in the $xy$-plane. A quick check reveals that it's orthogonal to the forward direction $f_1$ and points directly along the unicycle's "sideways" body axis. We have mathematically generated the parallel parking maneuver! We've created a new, "virtual" control out of thin air, just by combining the ones we already had. This new direction was not in our original distribution, but it was hiding in the way the allowed motions interacted. The basic calculation, which relies on the rates of change of the vector fields, is a cornerstone of this field.

This leads to a profound question. We started with two directions, $f_1$ and $f_2$, and generated a third, $[f_1, f_2]$. What if we now take the Lie bracket of our new direction with one of the originals, say $[f_1, [f_1, f_2]]$? Do we get yet another new direction of motion?

This process of generating new directions by repeatedly taking Lie brackets is the key to understanding controllability. The ​​Chow-Rashevsky Theorem​​ gives a stunningly powerful answer. It states that if, by continuing this process of taking brackets, you can eventually generate a set of vectors that spans every possible direction in your tangent space at every point, then you can get from any point to any other point in the manifold. This is known as the ​​bracket-generating​​ or ​​Hörmander condition​​.

In our unicycle example, the three vectors $f_1$, $f_2$, and $[f_1, f_2]$ are linearly independent at every point. They form a basis for the entire 3D space of position and orientation. Therefore, the unicycle is fully controllable. Despite being constrained to only ever roll and turn, there is no place it cannot go. The constraints have been overcome entirely by the geometry of non-commuting motions.

So, we can get anywhere. But what is the shortest way to get there? In Euclidean space, the shortest path is a straight line. But we are forbidden from moving in a straight line if it isn't one of our allowed directions. We must follow a "horizontal" path—a curve whose velocity vector always lies within our allowed distribution.

The length of the shortest such horizontal path between two points defines a new kind of distance, the ​​Carnot-Carathéodory distance​​, denoted $d_{\mathrm{CC}}$. This distance is the natural way to measure separation in a world of constraints, and it can be very different from the familiar Euclidean distance. A path that is short in $d_{\mathrm{CC}}$ might look very long and convoluted to a Euclidean observer. To travel a short distance sideways, a car must execute a series of forward and backward arcs, covering a much larger path length than the net displacement it achieves. Calculating this distance often involves solving a complex optimization problem, as seen in the classic example of the Heisenberg group. This new way of measuring distance is the foundation of ​​sub-Riemannian geometry​​.

This strange new distance warps the very fabric of space, leading to bizarre and counter-intuitive geometric properties. The most profound of these is that the dimension of the space is not what it seems.

Imagine dropping a spot of ink in water. It spreads out, and its concentration at the center decreases over time. On a 2D sheet, it scales with time as $t^{-2/2} = t^{-1}$. In a 3D volume, it scales as $t^{-3/2}$. The exponent is always tied to the dimension of the space. In mathematical terms, this is described by the ​​heat kernel​​, and its short-time behavior on a standard Riemannian manifold of dimension $m$ is always proportional to $t^{-m/2}$.

What happens in a sub-Riemannian space? The diffusion is constrained, so you might think it would spread more slowly. But the opposite is true in a sense. The heat spreads, but the concentration drops faster than the topological dimension would suggest. The scaling is not governed by the manifold's dimension $m$, but by a new quantity called the ​​homogeneous dimension​​, $Q$. The heat kernel scales as $t^{-Q/2}$.

This homogeneous dimension $Q$ is calculated from the structure of the Lie brackets. We assign a "cost" to each new direction we generate. The original allowed directions (level 1) have a cost of 1. Directions generated by a single Lie bracket (level 2) have a cost of 2. Directions generated by brackets of brackets (level 3) have a cost of 3, and so on. The homogeneous dimension is a weighted sum: $Q = \sum_{j=1}^{s} j \cdot (\text{number of new directions at level } j)$ where $s$ is the total number of bracket levels needed to span the whole space. Since directions at level $j>1$ exist, it is always true that $Q > m$ (unless the space was Riemannian to begin with, where $s=1$ and $Q=m$). For our 3D unicycle problem, the growth vector is (2, 3), meaning we have 2 directions at level 1 and 1 new direction at level 2. The homogeneous dimension is $Q = 1 \cdot 2 + 2 \cdot 1 = 4$.

This means that from the perspective of diffusion, or the volume of small balls, this 3-dimensional space behaves as if it were ​​4-dimensional​​! The volume of a ball of radius $r$ (measured in the $d_{\mathrm{CC}}$ distance) grows not as $r^3$, but as $r^Q = r^4$.

This strange scaling arises because the very notion of a "tangent space"—the local linear approximation of our manifold—is no longer the familiar Euclidean space. Instead, it's a more exotic structure called a ​​Carnot group​​ or a nilpotent Lie group. These are spaces with an inherent, anisotropic scaling. Moving a distance $r$ in a "hard" direction (generated by a long bracket) is equivalent to moving a distance of order $r^{1/j}$ in a base direction, where $j$ is the bracket level. It is this fundamental, non-Euclidean local structure that dictates the "anomalous" dimension and the strange behavior of distance and diffusion.

Sub-Riemannian geometry, therefore, is not just a mathematical curiosity. It is the natural geometry of constrained motion, revealing a world where wiggling can overcome any obstacle, and where the effective dimension of space itself depends on how hard you have to work to move in a particular direction.

Alright, so we’ve spent some time in the previous chapter building this rather peculiar world of sub-Riemannian geometry. We learned that by forbidding motion in certain directions, we didn’t cripple ourselves. Instead, we were forced to invent a new, subtler way to move and a new, more interesting way to measure distance—the Carnot-Carathéodory distance. You might be thinking, "This is a fine mathematical game, but what is it for? Where in the real world do we find these strange rules?" The answer, and this is the wonderful thing about it, is everywhere. From the way you park your car to the way a quantum computer calculates, from the random dance of a microscopic particle to the very structure of our own visual perception, this geometry shows up again and again. It is a unifying language for describing constrained motion, and its applications are a testament to the surprising power of abstract ideas. Let's take a tour.

Our first stop is perhaps the most intuitive: the world of things that move. Imagine a simple robot, maybe just a single disk rolling on a plane without slipping. The rules are simple: the wheel can roll forward or backward, and it can pivot to change its heading. But it cannot, under any circumstances, slide sideways. Its velocity vector is always confined to the direction it's pointing. This is a classic nonholonomic constraint—a rule on velocities, not positions.

So, here's a puzzle. If the disk can't move sideways, how can you possibly get it to a spot that is directly to its left or right? How do you parallel park this thing? You can’t just turn the wheel 90 degrees and drive over; that would be a sideways skid. The solution is a beautiful little dance, a sequence of wiggles. You roll a tiny bit forward, pivot a little, roll backward, and pivot back. What have you done? You've returned to your original orientation, but lo and behold, you've shifted sideways!

This maneuver, which seems like a bit of a magic trick, is the physical embodiment of a Lie bracket. You’re using a sequence of "allowed" motions (rolling and pivoting) to generate a net movement in a "forbidden" direction. The space of all possible motions is not just the directions you can instantaneously move in, but also all the directions you can generate by wiggling back and forth using these infinitesimal commutators. Sub-Riemannian geometry tells us that if the Lie brackets of our allowed motions are rich enough to span all possible directions—a condition known as the Lie Algebra Rank Condition (LARC)—then the system is completely controllable. From any starting point, you can reach any other point and orientation, just by executing the right sequence of wiggles. This principle, formalized by the Chow-Rashevsky theorem, is the cornerstone of nonholonomic control theory.

This isn't just about rolling disks. It applies to mobile robots, satellites reorienting themselves with limited thrusters, and even surgical needles that can be steered through tissue. Finding the best way to perform these maneuvers is an optimal control problem. For a system like the famous "Dubins car," which moves at a constant speed and has a maximum turning rate, the time-optimal paths are precisely the geodesics of the underlying sub-Riemannian geometry. These paths are beautiful concatenations of straight lines and circular arcs, and our geometry gives us the tools to calculate them. The abstract notion of a "shortest path" in our strange space becomes the very real "fastest route" for the robot. This idea extends to far more exotic systems, such as controlling the state of a quantum system, which can be modeled as finding geodesics on Lie groups like $SU(3)$.

But here comes a wonderful paradox, one that reveals the subtlety of this geometry. Just because you can get anywhere doesn't mean it's easy to stay put. Consider the famous "Brockett integrator," a canonical model of a nonholonomic system. It is fully controllable; its Lie brackets fill all of 3D space. You can drive it from the origin to any nearby point. But a famous theorem by Brockett shows that there is no continuous, "reflexive" steering law that can stabilize it at the origin. You can't write a simple function of your position that tells you how to steer to smoothly return to and stay at home. Any attempt to do so will find that there are always some directions from which you simply can't generate the necessary velocity to push you back towards the goal. You can plan a path to get there, but you can't create a simple, automatic "cruise control" for staying there. The very non-integrable structure that gives you the freedom to reach everywhere takes away the simplicity of stabilization.

Let's now turn from the deliberate world of control to the chaotic world of chance. Imagine a particle being buffeted by random forces, a tiny "drunken sailor" trying to walk on a ship's deck. But this sailor is peculiar: their random stumbles are constrained. They can only be pushed forward/backward or left/right by the chaotic waves, never diagonally. This is a stochastic differential equation (SDE) whose noise only acts in certain directions.

What does the world look like to this sailor? Where can they end up? Here, sub-Riemannian geometry provides a stunningly beautiful answer in the form of the Stroock-Varadhan support theorem. In essence, the theorem says that the set of all possible places the random process can wander to is exactly the set of places a clever, deterministic controller could drive to, using the same set of allowed vector fields. The random wiggling of noise effectively explores all the directions made available by the Lie brackets, just as our parallel-parking robot did. The unity is profound: the geometry of controllability is also the geometry of stochastic possibility.

If the Lie algebra of the noise vector fields fails to span the whole space, the random particle is forever trapped. Its possible locations are confined to the closure of a lower-dimensional submanifold, and its probability distribution has no hope of filling the space.

But what if the bracket-generating condition is satisfied? Then something wonderful happens. This idea, first formalized in Lars Hörmander’s celebrated theorem, is one of the pillars of modern analysis. Even though the noise (and thus the "heat") is only injected along a few preferred directions, the bracketing mechanism allows it to spread and smooth out in all directions. The resulting diffusion process has a perfectly smooth probability density, a property known as hypoellipticity.

This is the analytic heart of sub-Riemannian geometry. The operator that governs this diffusion, a "sub-Laplacian," is not elliptic—it looks degenerate. Yet, Hörmander's condition guarantees that its solutions are smooth. More than that, the entire behavior of this diffusion is dictated by the sub-Riemannian structure we built. The probability of the particle traveling from point $x$ to point $y$ in a small time $t$ doesn't depend on the straight-line Euclidean distance between them. Instead, it decays exponentially with the square of the Carnot-Carathéodory distance $d_{SR}(x,y)^2$. The difficult, winding path is what dictates the probability.

Furthermore, the diffusion doesn't behave as if it's in an $n$-dimensional space. For small times, the volume of the reachable set of points—and thus the scaling of the heat kernel—is governed by a different number, the "homogeneous dimension" $Q$, which is determined by how many brackets you need to generate all directions. This "fractal-like" dimension is intrinsic to the sub-Riemannian geometry and is always greater than or equal to the manifold's dimension. The geometry we invented out of mechanical necessity turns out to be the natural stage for these fundamental physical processes of diffusion and heat flow.

Perhaps the most astonishing and speculative appearance of sub-Riemannian geometry is in the workings of our own brain. When you look at an image, the first stage of processing happens in the primary visual cortex, V1. Neurons in V1 are sensitive to lines and edges at specific positions, orientations, and scales. The mathematical space describing all these features is the group of positions and orientations, $SE(2)$.

Remarkably, pioneering research has proposed that the functional architecture of V1 is, in fact, a sub-Riemannian geometry on this $SE(2)$ group. The system is constrained; a neuron can most easily "pass a message" to a nearby neuron that represents a similar orientation. Generating the perception of a long, curved contour from these local edge detections is equivalent to finding a "path of least effort" connecting the initial neuronal responses. And what are these paths of least effort? They are precisely the geodesics of this neural sub-Riemannian geometry. This suggests that our brain might be hard-wired to perceive the world by tracing out optimal paths in a nonholonomic space, a beautiful and humbling thought that connects the highest levels of cognition to the geometry of a rolling wheel.

So, what have we seen? We started with a simple mechanical constraint and discovered a rich geometric world. We saw how this world provides the language for controlling robots, for finding their optimal paths, and for understanding the paradoxical limits of their stability. We then saw this very same geometry emerge from the heart of probability theory, describing the random wanderings of particles and the flow of heat in constrained systems, dictating everything from the support of the process to the shape of the heat kernel. We even caught a glimpse of it in the deep architecture of the brain's visual system.

From the tangible to the abstract, from the deliberate to the random, sub-Riemannian geometry reveals a hidden unity. It teaches us that constraints are not just limitations; they are the source of a new and beautiful structure, a structure that nature seems to have discovered and put to use time and time again. The "impossible" sideways motion of a wheel and the perception of a curved line in our mind's eye may just be two sides of the same geometric coin. And that is a truly wonderful thing to realize.