---
title: 'Nonholonomic Constraints: Theory and Control'
url: https://www.emergentmind.com/topics/nonholonomic-constraints
type: topic
---

# Nonholonomic Constraints: Theory and Control

A nonholonomic constraint is a relation—typically involving both configurations and velocities—imposed on a mechanical system that restricts its admissible velocities but is not generally integrable to a configuration-only condition. Unlike holonomic constraints, which reduce configuration space dimension via integrable relations, nonholonomic constraints produce non-integrable distributions in the tangent bundle, leading to fundamentally different geometric, analytic, and control-theoretic properties. Such constraints can arise from physical mechanisms (e.g., wheel rolling without slipping) or be enforced “virtually” via closed-loop control laws.

## 1. Geometric Foundations of Nonholonomic Constraints

A nonholonomic constraint on an $n$-dimensional configuration manifold $Q$ is typically expressed as either:
- A set of $m < n$ independent relations $C(q,\dot q)=0$ homogeneous or nonhomogeneous in velocities, or  
- Equivalently, as a velocity constraint—an $m$-codimensional distribution $D\subset TQ$ locally defined by one-forms $\mu^a(q)$:
  $$
  D_q = \left\{ v_q \in T_qQ \;\middle|\; \mu^a(q)(v_q) = 0,\quad a=1,...,m \right\}.
  $$
A constraint is strictly nonholonomic when $D$ is non-integrable, i.e., $[X,Y] \not\in \Gamma(D)$ for some $X,Y \in \Gamma(D)$. This non-integrability is central, originating from the Frobenius theorem and manifesting in the vanishing of the curvature terms $\mu^a \wedge d\mu^a \neq 0$. The resultant admissible velocity set at any instant forms a non-integrable hyperplane field in $TQ$ [2207.01299][1910.09395][2411.01692][2310.01849].

Constraints may be more general, as in the nonlinear case, taking the form $\Phi_j(q,\dot q)=0$ where $\Phi_j$ may be nonlinear in velocities [1910.09395][2310.01849].

## 2. Dynamical Consequences and Formulations

The dynamics of nonholonomic systems are governed by the Lagrange–d'Alembert principle: the admissible variations $\delta q$ must themselves satisfy the constraint, resulting in the constrained Euler–Lagrange equations:
$$
\frac{d}{dt}\frac{\partial L}{\partial \dot q^i}-\frac{\partial L}{\partial q^i} = \lambda_a\,\mu^a_i(q),\quad
\mu^a_i(q) \dot q^i = 0,
$$
where $\lambda_a$ are Lagrange multipliers enforcing the constraints [1707.03929][1701.04766].

Alternative variational approaches, such as the naive use of Lagrange multipliers in the action functional, are generally incompatible with nonholonomic constraints in dimensions $N\ge3$, unless the constraints are integrable. This incompatibility arises because d'Alembert’s principle restricts only admissible virtual displacements, not arbitrary variations [1002.0533]. Extensions such as vakonomic mechanics or Dirac–Hamilton–Jacobi theory are required for consistent variational treatments [1109.6056].

For constraints nonlinear in velocities, Voronec’s equations provide an intrinsic reduction framework, expressing dynamics directly on the submanifold determined by solving the constraints for the dependent velocities, and generalize beyond the linear case [1910.09395].

## 3. Virtual and Affine Nonholonomic Constraints

A virtual nonholonomic constraint is a velocity-dependent relation not physically imposed on the system but rendered invariant via state-dependent closed-loop control laws. Formally, let a mechanical control system on $Q$ be equipped with control forces $u_a$ via vector fields $Y^a$. A virtual constraint is specified as a controlled-invariant submanifold $M=\phi^{-1}(0) \subset TQ$. A feedback law $u^* : M \to \mathbb{R}^m$ is constructed so that every solution starting in $M$ remains in $M$ for all time—making $M$ controlled-invariant [2310.01849][2411.01692][2207.01299][2301.03890].

Transversality between the constraint distribution and the input distribution (i.e., $T_{v_q}TQ = T_{v_q}M \oplus \mathcal{F}_{v_q}^V$) is essential for the well-posedness and uniqueness of the enforcing control law. The general constraint enforcement law arises from solving the linear system:
$$
A(v_q)\,\tau = b(v_q),
$$
with $A^b_a(v_q) = d\phi^b((Y^a)^V)$ and $b^b(v_q) = -d\phi^b(G(v_q))$, ensuring the closed-loop vector field is tangent to $M$ [2310.01849][2207.01299].

For affine (nonlinear) constraints, the framework extends to controlled-invariant affine distributions associated with affine-connection control systems, requiring similar transversality and construction of a unique smooth feedback [2301.03890][2502.03902].

## 4. Stabilization and Exponential Convergence

Recent developments provide not only invariance but exponential stabilization to the virtual constraint submanifold using a gain-based feedback:
$$
u^*_a(q,\dot q) = (C^{-1})_{ab}(q,\dot q)\left[ -K^{bc}\phi^c(q,\dot q) - G(\phi^b)(q,\dot q)\right],
$$
with $C^{ab}(q,\dot q) = (Y^a)^V(\phi^b)$, and $K$ a positive diagonal gain matrix. This yields closed-loop error dynamics $\dot\phi = -K\phi$, guaranteeing exponential convergence of the constraint violation to zero [2411.01692][2502.03902].

This methodology provides Lyapunov-based guarantees and is demonstrated in scenarios such as multi-agent flocking (velocity alignment via virtual constraints) and drift compensation in unmanned surface vehicles, showing rapid convergence to the desired constraint manifold under the derived control law [2502.03902].

## 5. Examples, Physical Realizability, and Applications

Various canonical models are treated within this geometric and control-theoretic framework:
- The Chaplygin sleigh and vertical rolling disk are analyzed as nonholonomic systems on (possibly homogeneous) configuration spaces [2411.05485][2207.01299][2310.01849][2411.01692].
- Parallel velocity alignment in multi-agent systems and trajectory tracking for underactuated vehicles are implemented via nonlinear virtual constraints and demonstrated by simulation [2502.03902][2310.01849].
- Recent work addresses physical realizability, illustrating that in systems coupled to a thermal environment, naive enforcement of nonholonomic constraints may violate fundamental thermodynamic laws unless stochastic fluctuations at the contact interface are properly modeled (fluctuation–dissipation consistency), setting practical limits for “ideal” constraint enforcement in real systems [2603.14424].

Finally, structure-preserving learning frameworks for nonholonomic dynamics have emerged, employing Gaussian process regression with kernel construction encoding the constraint distribution, thereby ensuring data-driven models that are physically consistent by construction [2603.27580].

## 6. Interplay with Control, Optimality, and Hamiltonian Structures

Optimal control approaches for systems subject to nonholonomic constraints require formulating the problem as constrained second-order variational or Hamiltonian systems—typically using geometric frameworks such as Lie algebroids or distributional two-forms. This allows for Hamilton–Jacobi reductions and geometric numerical integration [1701.04766][2205.01998][1109.6056].

For nonholonomic controlled Hamiltonian systems, one replaces the symplectic structure by a fiberwise non-degenerate distributional two-form on the constraint submanifold. The resulting dynamics are governed by projections of the unconstrained vector field and preserve certain geometric structures, but lack a global symplectic foliation due to the non-integrability of the distribution [2205.01998][1911.05409].

## 7. Methodological and Foundational Issues

A well-documented controversy concerns the distribution of virtual displacements (d'Alembertian) versus generalized variations (vakonomic) in enforcing constraints. In general, only d'Alembert’s principle reproduces the correct nonholonomic equations for genuinely non-integrable constraints in dimension $N\geq 3$. Vakonomic approaches, unless suitably modified (e.g., using Dirac structures or restriction on allowed variations), yield incorrect equations except in the holonomic or $N=2$ cases, where all constraints are effectively holonomic [1002.0533][1010.3902][1109.6056].

Nonlinear and higher-order constraints necessitate generalized approaches, such as those based on Voronec's method, projection techniques, or coordinate-free geometric conditions, emphasizing the importance of explicit construction and proper matching between constraint enforcement, dynamics, and geometric structure [1910.09395][2404.11997].

---

**References**:  
- [2310.01849] On the Geometry of Virtual Nonlinear Nonholonomic Constraints  
- [2411.01692] Geometric stabilization of virtual linear nonholonomic constraints  
- [2502.03902] Geometric Stabilization of Virtual Nonlinear Nonholonomic Constraints  
- [2207.01299] Virtual Nonholonomic Constraints: A Geometric Approach  
- [2301.03890] Virtual Affine Nonholonomic Constraints  
- [2411.05485] Nonholonomic mechanics and virtual constraints on Riemannian homogeneous spaces  
- [2404.11997] Geodesic extensions of mechanical systems with nonholonomic constraints  
- [1910.09395] Nonlinear nonholonomic constraints  
- [1701.04766] A variational-geometric approach for the optimal control of nonholonomic systems  
- [2205.01998] Nonholonomic Controlled Hamiltonian System: Symmetric Reduction and Hamilton-Jacobi Equations  
- [1109.6056] Hamilton-Jacobi Theory for Degenerate Lagrangian Systems with Holonomic and Nonholonomic Constraints  
- [1002.0533] On the compatibility of non-holonomic systems and certain related variational systems  
- [1010.3902] Non-holonomic Constraint Force Postulates  
- [2603.14424] Nonholonomic constraints at finite temperature  
- [2603.27580] Structure-Preserving Learning of Nonholonomic Dynamics  
- [1610.02595] Constraint Control of Nonholonomic Mechanical Systems  
- [1707.03929] Geometric analysis of noisy perturbations to nonholonomic constraints

Source: https://www.emergentmind.com/topics/nonholonomic-constraints