Constrained Lagrangian Systems: Lagrange Multipliers and D'Alembert's Principle

Intuition Beginner

Imagine a bead threaded on a stiff wire bent into a hoop. The bead can slide freely along the wire but cannot leave it. The wire is a constraint -- it restricts the bead's possible positions to a one-dimensional curve embedded in two- or three-dimensional space.

An unconstrained particle in three dimensions has three independent coordinates $(x,y,z)$ and three degrees of freedom (DOF). A bead on a wire has only one degree of freedom -- the distance along the wire, or equivalently an angle $\theta$ that parameterises position. The constraint has removed two degrees of freedom.

A simple pendulum is another constrained system. The bob can swing in a plane but cannot move farther from the pivot than the rod length $\ell$. Without the rod, the bob would have two degrees of freedom $(x,y)$ in the plane. The rigid rod imposes the constraint $x^2+y^2={\ell }^2$, reducing the system to one degree of freedom -- the angle $\theta$.

In the Lagrangian framework 09.02.02, the standard approach is to choose coordinates that automatically satisfy the constraints. For the pendulum, you write the Lagrangian in terms of $\theta$ alone: $L=\frac{1}{2}m{\ell }^2{\dot{\theta }}^2+mg\ell \cos \theta$. The constraint is built into the coordinate choice and never appears explicitly.

But sometimes you need the constraint force -- the tension in the pendulum rod, or the normal force the wire exerts on the bead. Sometimes the constraint is complicated enough that eliminating it by a clever coordinate choice is impractical. This is where Lagrange multipliers enter.

The idea: keep all the original coordinates, add a new unknown variable (the multiplier $\lambda$) for each constraint, and solve the extended equations of motion. The multiplier appears exactly where the constraint force would appear in Newton's second law. Solving gives you both the trajectory and the constraint force simultaneously.

The multiplier has a direct physical meaning. For a constraint $f(q)=0$, the constraint force points perpendicular to the constraint surface, in the direction of steepest change of $f$. The multiplier $\lambda$ scales this gradient to produce the correct magnitude. It adjusts dynamically so the trajectory never leaves the constraint surface.

Visual Beginner

Figure: A bead of mass $m$ on a circular wire of radius $R$ in a vertical plane. The bead's position is parameterised by the angle $\theta$ measured from the bottom. Gravity $mg$ acts downward. The constraint force $\mathbf{N}$ acts radially (perpendicular to the wire, toward the centre). At the bottom, $\mathbf{N}$ points upward, opposing gravity and providing centripetal acceleration. At the top, $\mathbf{N}$ points downward, and the bead falls off if gravity exceeds the required centripetal force.

The constraint force is always perpendicular to the wire (normal to the constraint surface). The tangential component of gravity alone determines the acceleration along the wire.

Worked example Beginner

The bead on a hoop

A bead of mass $m$ slides on a vertical circular wire of radius $R$. Find the equation of motion and the constraint force.

Using the built-in coordinate. Parameterise by the angle $\theta$ from the bottom. Position: $(R\sin \theta ,-R\cos \theta )$. The Lagrangian is

$$L=\frac{1}{2}m{R}^2{\dot{\theta }}^2+mgR\cos \theta .$$

The Euler-Lagrange equation gives $m{R}^2\ddot{\theta }=-mgR\sin \theta$, or

$$\ddot{\theta }=-\frac{g}{R}\sin \theta .$$

This is a simple pendulum equation. But it tells you nothing about the normal force.

Using Lagrange multipliers to find the constraint force. Work in Cartesian coordinates with the constraint $f(x,y)=x^2+y^2-R^2=0$. The Lagrangian is $L=\frac{1}{2}m({\dot{x}}^2+{\dot{y}}^2)-mgy$. The extended equations of motion are

$$m\ddot{x}=2\lambda x,m\ddot{y}=-mg+2\lambda y,$$

together with $x^2+y^2=R^2$.

At the bottom of the hoop ($\theta =0$, so $x=0$, $y=-R$): the constraint force is purely radial (vertical). Using the centripetal condition $m\ddot{y}=m{v}^2/R$ (with $v$ the speed along the hoop), the $y$-equation gives $m{v}^2/R=-mg+2\lambda (-R)$, so $2\lambda R=-mg-m{v}^2/R$. The magnitude of the normal force is $|N|=|2\lambda R|=mg+m{v}^2/R$. At rest ($v=0$), the wire supports the full weight: $N=mg$.

At the top ($\theta =\pi$, $x=0$, $y=R$): $m{v}^2/R=mg+2\lambda R$ gives $N=m{v}^2/R-mg$. The bead maintains contact only if $v^2\ge gR$ -- the centripetal requirement exceeds gravity.

Check your understanding Beginner

Formal definition Intermediate+

Holonomic constraints

A holonomic constraint on a mechanical system with configuration coordinates $q^1,\dots ,q^n$ is an equation of the form

$$f_{\alpha }(q^1,\dots ,q^n,t)=0,\alpha =1,\dots ,k.$$

The adjective "holonomic" means the constraint involves only positions and possibly time, not velocities -- or if velocities appear, they can be integrated to a position-only constraint. Each holonomic constraint reduces the number of degrees of freedom by one: $n$ coordinates with $k$ constraints leave $n-k$ independent degrees of freedom.

By contrast, a non-holonomic constraint is one that involves velocities in a way that cannot be integrated to a position constraint. A wheel rolling without slipping is the standard example: the no-slip condition relates linear and angular velocities but cannot be reduced to a constraint on positions alone. Non-holonomic constraints do not reduce the number of degrees of freedom in configuration space (the system can still reach any configuration); they restrict the paths the system can take between configurations.

The extended Euler-Lagrange equations

For a system with Lagrangian $L(q,\dot{q},t)$ and holonomic constraints $f_{\alpha }(q,t)=0$ ($\alpha =1,\dots ,k$), the equations of motion are

$$\boxed{\frac{d}{dt}\frac{\partial L}{\partial {\dot{q}}^i}-\frac{\partial L}{\partial q^i}=\sum _{\alpha =1}^k{\lambda }_{\alpha }\frac{\partial f_{\alpha }}{\partial q^i},i=1,\dots ,n,}$$

together with the constraint equations

$$f_{\alpha }(q,t)=0,\alpha =1,\dots ,k.$$

This is a system of $n+k$ equations for $n+k$ unknowns $(q^1,\dots ,q^n,{\lambda }_1,\dots ,{\lambda }_k)$. The right-hand side is the generalised constraint force:

$$Q_i^{(c)}=\sum _{\alpha }{\lambda }_{\alpha }\frac{\partial f_{\alpha }}{\partial q^i}.$$

Variational derivation

The extended equations arise from extremising the augmented action

$$S={\int }_0^T\left(L(q,\dot{q},t)-\sum _{\alpha =1}^k{\lambda }_{\alpha }(t)f_{\alpha }(q,t)\right)dt,$$

where the ${\lambda }_{\alpha }(t)$ are functions of time (one multiplier per constraint, varying freely in time). Varying $q^i$ gives the extended Euler-Lagrange equations. Varying ${\lambda }_{\alpha }$ gives the constraints $f_{\alpha }=0$. The endpoint conditions $\delta q^i(0)=\delta q^i(T)=0$ are used in the integration by parts for the $q^i$ variation.

D'Alembert's principle

D'Alembert's principle states that the total virtual work of the applied forces minus the inertial terms vanishes for all virtual displacements compatible with the constraints:

$$\sum _i\left(Q_i-\left[\frac{d}{dt}\frac{\partial L}{\partial {\dot{q}}^i}-\frac{\partial L}{\partial q^i}\right]\right)\delta q^i=0$$

for all $\delta q^i$ satisfying ${\sum }_i(\partial f_{\alpha }/\partial q^i)\delta q^i=0$. The multipliers enforce this orthogonality: the "unbalanced force" (the left side of the extended Euler-Lagrange equation) must be a linear combination of the constraint gradients, which is exactly the statement that the constraint force compensates the component of the applied force normal to the constraint surface.

Constraint force identification

For a single particle of mass $m$ in a potential $V(\mathbf{r})$ with constraint $f(\mathbf{r})=0$, the constraint force is ${\mathbf{F}}^{(c)}=\lambda \nabla f$. The multiplier satisfies

$$\lambda =\frac{m(\dot{\mathbf{r}}\cdot \nabla {)}^{2}f-\nabla V\cdot \nabla f}{|\nabla f{|}^2},$$

evaluated on the constraint surface. The numerator has two contributions: a centripetal term from the curvature of the constraint surface and a term from the projection of the applied force onto the surface normal.

Worked example: the double pendulum

Two masses $m_1$ and $m_2$ are connected by rigid rods of lengths ${\ell }_1$ (pivot to $m_1$) and ${\ell }_2$ ($m_1$ to $m_2$), swinging in a vertical plane. This system has two constraints: $x_1^2+y_1^2={\ell }_1^2$ and $(x_2-x_1{)}^2+(y_2-y_1{)}^2={\ell }_2^2$. Four Cartesian coordinates minus two constraints gives two degrees of freedom, which can be taken as the angles ${\theta }_1$ and ${\theta }_2$.

Expressing the positions in terms of angles:

$$x_1={\ell }_1\sin {\theta }_1,y_1=-{\ell }_1\cos {\theta }_1,$$ $$x_2={\ell }_1\sin {\theta }_1+{\ell }_2\sin {\theta }_2,y_2=-{\ell }_1\cos {\theta }_1-{\ell }_2\cos {\theta }_2.$$

The Lagrangian is

$$L=\frac{1}{2}(m_1+m_2){\ell }_1^2{\dot{\theta }}_1^2+\frac{1}{2}{m}_2{\ell }_2^2{\dot{\theta }}_2^2+m_2{\ell }_1{\ell }_2{\dot{\theta }}_1{\dot{\theta }}_2\cos ({\theta }_1-{\theta }_2)+(m_1+m_2)g{\ell }_1\cos {\theta }_1+m_{2}g{\ell }_2\cos {\theta }_2.$$

The Euler-Lagrange equations for ${\theta }_1$ and ${\theta }_2$ give a coupled nonlinear system. In the small-angle limit, this reduces to two coupled linear oscillators, which can be diagonalised into normal modes. The full nonlinear system exhibits chaotic behaviour for large amplitudes.

To find the rod tensions (the constraint forces), one would instead work in Cartesian coordinates with two Lagrange multipliers ${\lambda }_1$ and ${\lambda }_2$. The extended equations give both the motion and the tensions as functions of time.

Worked example: the Atwood machine

Two masses $m_1$ and $m_2$ ($m_1>m_2$) are connected by a massless inextensible string over a frictionless pulley. The constraint is $y_1+y_2=\ell$ (constant string length), giving ${\dot{y}}_1=-{\dot{y}}_2$ and ${\ddot{y}}_1=-{\ddot{y}}_2$.

The Lagrangian with both coordinates is $L=\frac{1}{2}{m}_1{\dot{y}}_1^2+\frac{1}{2}{m}_2{\dot{y}}_2^2+m_{1}g{y}_1+m_{2}g{y}_2$. The extended equations are

$$m_1{\ddot{y}}_1=m_{1}g+\lambda ,m_2{\ddot{y}}_2=m_{2}g+\lambda ,$$

with constraint $y_1+y_2=\ell$. Using ${\ddot{y}}_2=-{\ddot{y}}_1$:

$$m_1{\ddot{y}}_1=m_{1}g+\lambda ,-m_2{\ddot{y}}_1=m_{2}g+\lambda .$$

Subtracting: $(m_1+m_2){\ddot{y}}_1=(m_1-m_2)g$, giving the acceleration ${\ddot{y}}_1=(m_1-m_2)g/(m_1+m_2)$. Back-substituting for the multiplier: $\lambda =m_1{\ddot{y}}_1-m_{1}g=-2m_1{m}_{2}g/(m_1+m_2)$. The constraint force on $m_1$ is $T=\lambda =-2m_1{m}_{2}g/(m_1+m_2)$ (the string tension, directed upward).

Key theorem with proof Intermediate+

Theorem (Constraint forces do no virtual work). For a system with scleronomic (time-independent) holonomic constraints $f_{\alpha }(q)=0$, the constraint forces do zero work along any virtual displacement compatible with the constraints:

$$\sum _i{Q}_i^{(c)}\delta q^i=\sum _i\sum _{\alpha }{\lambda }_{\alpha }\frac{\partial f_{\alpha }}{\partial q^i}\delta q^i=0$$

for all $\delta q^i$ satisfying ${\sum }_i(\partial f_{\alpha }/\partial q^i)\delta q^i=0$ for each $\alpha$.

Proof. A virtual displacement $\delta q^i$ compatible with the constraints satisfies ${\sum }_i(\partial f_{\alpha }/\partial q^i)\delta q^i=0$ for each $\alpha$ (this is the linearisation of the constraint: to first order, $\delta f_{\alpha }=0$). The virtual work of the constraint forces is

$$\sum _i{Q}_i^{(c)}\delta q^i=\sum _{\alpha }{\lambda }_{\alpha }\underset{=0}{\underbrace{\sum _i\frac{\partial f_{\alpha }}{\partial q^i}\delta q^i}}=0.$$

Each term in the outer sum vanishes because the inner sum is the variation $\delta f_{\alpha }$, which is zero by the compatibility condition. $\square$

This is the formal expression of D'Alembert's principle: the constraint forces are orthogonal to the constraint surface, and virtual displacements lie within the constraint surface, so their dot product vanishes. The physical content is that rigid wires, rods, and surfaces do no work on the system -- they redirect motion without adding or removing energy.

Corollary. The work done by constraint forces on an actual trajectory also vanishes for scleronomic constraints. The power is $P={\sum }_i{Q}_i^{(c)}{\dot{q}}^i={\sum }_{\alpha }{\lambda }_{\alpha }{\dot{f}}_{\alpha }$. Since $f_{\alpha }(q)=0$ implies ${\dot{f}}_{\alpha }={\sum }_i(\partial f_{\alpha }/\partial q^i){\dot{q}}^i=0$, the constraint forces do zero work on the actual trajectory. This is consistent with energy conservation: if the Lagrangian has the standard form $L=T-V$ with no explicit time dependence, the total energy $T+V$ is conserved, and the constraint forces contribute nothing to the energy budget.

Worked example: bead on a parabola

A bead of mass $m$ moves in a vertical plane, constrained to the parabola $y=x^2$. The constraint is $f(x,y)=y-x^2=0$.

The Lagrangian in Cartesian coordinates is $L=\frac{1}{2}m({\dot{x}}^2+{\dot{y}}^2)-mgy$. The extended equations:

$$m\ddot{x}=-2\lambda x,m\ddot{y}=-mg+\lambda .$$

The constraint implies $\dot{y}=2x\dot{x}$ and $\ddot{y}=2{\dot{x}}^2+2x\ddot{x}$. Substituting the $\ddot{x}$ equation: $\ddot{y}=2{\dot{x}}^2-4\lambda x^2/m$. Setting equal to the $\ddot{y}$ from the second extended equation: $2{\dot{x}}^2-4\lambda x^2/m=-g+\lambda /m$, giving

$$\lambda =\frac{m(g+2{\dot{x}}^2)}{1+4x^2}.$$

The constraint force is ${\mathbf{F}}^{(c)}=\lambda (-2x,1)$, perpendicular to the parabola. At equilibrium ($\dot{x}=0$, $\ddot{x}=0$): $\lambda =mg/(1+4x^2)$. The only equilibrium is at the vertex $x=0$, where $\lambda =mg$ and the normal force balances gravity.

Exercises Intermediate+

Lean formalization Intermediate+

lean_status: none. The Lagrange multiplier method extends Mathlib's unconstrained Euler-Lagrange machinery to systems with holonomic constraints. The main formalisation gap is the construction of the constrained action functional $S^{\prime }=\int (L-{\sum }_{\alpha }{\lambda }_{\alpha }{f}_{\alpha })dt$ and its variation, together with the identification of ${\lambda }_{\alpha }\nabla f_{\alpha }$ as the constraint force. This would require combining Mathlib's calculus of variations on manifolds with the submanifold geometry of constraint surfaces -- assembling the tangent-bundle geometry of constraint manifolds with the variational framework, and extending the existing Euler-Lagrange machinery to the constrained setting. The Lagrange-d'Alembert equations for non-holonomic constraints are a further gap.

Advanced results Master

Constraint manifolds and the geometry of multipliers

When $k$ holonomic constraints $f_1,\dots ,f_k$ are imposed on an $n$-dimensional configuration space $Q$, the constraint surface $S=\{q\in Q:f_{\alpha }(q)=0,\alpha =1,\dots ,k\}$ is a submanifold of codimension $k$ provided the gradients $\nabla f_{\alpha }$ are linearly independent at every point of $S$. The tangent space $T_{q}S$ at $q\in S$ consists of all vectors $\delta q$ satisfying ${\sum }_i(\partial f_{\alpha }/\partial q^i)\delta q^i=0$ for each $\alpha$. The constraint forces span the normal bundle of $S$: at each $q\in S$, the constraint force lies in the span of $\{\nabla f_1,\dots ,\nabla f_k\}$, which is the orthogonal complement of $T_{q}S$ in $T_{q}Q$.

The multipliers ${\lambda }_{\alpha }$ are determined by the requirement that the trajectory remain on $S$. Differentiating the constraint $f_{\alpha }=0$ twice with respect to time gives ${\ddot{f}}_{\alpha }=0$, which produces $k$ algebraic equations for the $k$ multipliers. This is the content of the constraint-force formula derived in the Formal definition section: the multipliers adjust instantaneously (they are not dynamical variables with their own equations of motion) to enforce the constraints at every instant.

Non-holonomic constraints and the Lagrange-d'Alembert equations

Non-holonomic constraints $g_{\alpha }(q,\dot{q},t)=0$ that depend on velocities in a non-integrable way cannot be handled by the augmented-action method. The correct equations are the Lagrange-d'Alembert equations:

$$\frac{d}{dt}\frac{\partial L}{\partial {\dot{q}}^i}-\frac{\partial L}{\partial q^i}=\sum _{\alpha }{\mu }_{\alpha }\frac{\partial g_{\alpha }}{\partial {\dot{q}}^i}.$$

The crucial structural difference from the holonomic case: the multipliers multiply $\partial g_{\alpha }/\partial {\dot{q}}^i$, not $\partial g_{\alpha }/\partial q^i$. This is because non-holonomic constraints restrict the velocity direction, not the position, so the constraint force acts in velocity space.

The Lagrange-d'Alembert equations cannot be derived from an augmented variational principle $\delta \int L^{\prime }dt=0$. They are instead derived from D'Alembert's principle directly: the virtual work of the constraint forces vanishes for all virtual displacements satisfying ${\sum }_i(\partial g_{\alpha }/\partial {\dot{q}}^i)\delta q^i=0$. This is a weaker foundation than the variational principle for holonomic constraints and is the source of ongoing discussion in the mechanics literature.

Example: the rolling disk. A thin disk of radius $r$, mass $m$, rolling without slipping on a horizontal plane. The configuration is $(x,y,\theta ,\varphi )$ where $(x,y)$ is the contact point, $\theta$ the heading angle, $\varphi$ the rotation angle. The rolling constraints are $g_1=\dot{x}-r\dot{\varphi }\cos \theta =0$ and $g_2=\dot{y}-r\dot{\varphi }\sin \theta =0$. These are linear in velocities but cannot be integrated to position constraints (the path depends on the history of $\theta$). The Lagrange-d'Alembert equations give the friction forces at the contact point as the multipliers ${\mu }_1$ and ${\mu }_2$.

Semi-holonomic constraints

A constraint is semi-holonomic if it can be written in the form $g_{\alpha }(q,\dot{q},t)=0$ and admits an integrating factor -- a function $\mu (q,\dot{q},t)$ such that $\mu g_{\alpha }$ is the total time derivative of some function of $q$ and $t$. Semi-holonomic constraints are effectively holonomic after this integration and can be treated by the augmented-action method. The boundary between semi-holonomic and genuinely non-holonomic constraints is important: Frobenius's theorem provides the integrability condition (the distribution defined by the constraint must be involutive).

Dirac constraints in Hamiltonian mechanics -- preview

The Lagrange multiplier method in the Lagrangian framework has a direct Hamiltonian counterpart. When the Legendre transform 09.04.01 is applied to a constrained system, the constraints may become relations among the phase-space variables $(q,p)$. Dirac's theory of constrained Hamiltonian systems classifies these as:

• Primary constraints: relations among $(q,p)$ that follow directly from the definition $p_i=\partial L/\partial {\dot{q}}^i$. If the constraint $f(q)=0$ is holonomic, the corresponding primary constraint in phase space involves both $q$ and $p$ in a way that depends on the specific Lagrangian.
• Secondary constraints: consistency conditions requiring that primary constraints be preserved under time evolution.
• First-class constraints: those whose Poisson brackets with all other constraints vanish on the constraint surface. These generate gauge transformations.
• Second-class constraints: those that have non-vanishing Poisson brackets with at least one other constraint. These are eliminated by the Dirac bracket, a modified Poisson bracket that respects the constraint structure.

The Lagrange multipliers of the Lagrangian formulation correspond to the undetermined velocities associated with primary constraints in the Dirac theory. The passage from Lagrange multipliers to Dirac brackets is the Hamiltonian counterpart of the material in this unit.

Gauge theory analogy

In gauge field theories, the gauge field $A_{\mu }$ has a Lagrangian (the Yang-Mills Lagrangian $L=-\frac{1}{4}{F}_{\mu \nu }{F}^{\mu \nu }$) that is singular: the Hessian with respect to ${\dot{A}}_{\mu }$ is degenerate because ${\dot{A}}_0$ does not appear. This is a primary constraint in Dirac's sense. The Lagrange multiplier associated with this constraint is the undetermined time component of the gauge field, and the freedom to choose it is the gauge freedom.

The parallel is exact: the Lagrange multiplier method for finite-dimensional systems with holonomic constraints is the finite-dimensional prototype for the gauge-fixing procedures of field theory. The constraint forces (normal to the constraint surface) are the finite-dimensional analogues of the ghost fields and gauge-fixing terms in the Faddeev-Popov formalism. Understanding the finite-dimensional case -- this unit -- is prerequisite for the field-theoretic generalisation 09.02.06.

Rheonomic constraints

Constraints that depend explicitly on time, $f_{\alpha }(q,t)=0$, are called rheonomic. The extended Euler-Lagrange equations are formally identical to the scleronomic case, but the constraint force now does work: the power is $P={\sum }_{\alpha }{\lambda }_{\alpha }{\dot{f}}_{\alpha }={\sum }_{\alpha }{\lambda }_{\alpha }\partial f_{\alpha }/\partial t$, which is generally non-zero. Energy is not conserved for rheonomic systems; the constraint force injects or extracts energy through the time-dependent boundary.

The parametrically driven pendulum (Exercise 6) is the canonical example. The oscillating pivot is a rheonomic constraint, and the work done by the pivot's motor is what drives the parametric resonance.

Synthesis. The Lagrange multiplier method is the foundational device that lets constrained dynamics be formulated without eliminating coordinates. The central insight is that the multiplier is dual to the constraint, encoding the force needed to maintain it. This duality -- kinematic restriction as equation, dynamic enforcement as force, linked by the multiplier -- is the algebraic structure that Dirac exploited in his theory of constrained Hamiltonian systems. Putting these together, the multiplier formulation provides the computational bridge from the Lagrangian coordinate-elimination strategy to the Hamiltonian constraint-classification strategy. The generalisation to field theories with gauge constraints -- where the multipliers become gauge fields -- completes the passage from classical mechanics to modern theoretical physics.

Connections Master

• Action principle 09.02.01 -- the augmented action $S=\int (L-\sum {\lambda }_{\alpha }{f}_{\alpha })dt$ is a constrained variational principle. The multipliers enforce the constraints as Euler-Lagrange equations of the extended Lagrangian.

• Euler-Lagrange equations 09.02.02 -- the extended Euler-Lagrange equations are the standard EL equations with the constraint force added to the right-hand side. The unconstrained EL equations are the special case $k=0$.

• Legendre transform 09.04.01 -- the passage to Hamiltonian mechanics carries the Lagrange multipliers into Dirac's primary constraints. The multiplier $\lambda$ in the Lagrangian setting becomes the undetermined velocity of a primary constraint in the Hamiltonian setting. The Dirac-Bergmann algorithm is the Hamiltonian counterpart of the multiplier method.

• Field Lagrangians 09.02.06 -- gauge field theories use the Lagrange multiplier method at the field-theoretic level. The electromagnetic Gauss-law constraint $\nabla \cdot \mathbf{E}=\rho$ arises as a secondary constraint in the Dirac analysis of the Maxwell Lagrangian, and the multiplier is the time component $A_0$ of the gauge field.

• Noether's theorem 09.03.01 -- symmetries of a constrained system must preserve the constraint surface to generate conserved quantities. The multiplier formalism makes this condition explicit: only symmetries for which $\delta f_{\alpha }=0$ on the constraint surface contribute conserved Noether charges.

Historical & philosophical context Master

Jean-Baptiste le Rond d'Alembert introduced the principle of virtual work in his Traite de dynamique (1743). The key insight was that the forces of constraint do no work under virtual displacements compatible with the constraints -- a principle that converts the problem of constrained motion from a force-balance problem (Newton) into a geometric problem (restricting motion to a submanifold of configuration space). D'Alembert's principle subsumes the statics principle of virtual work (Bernoulli, 1717) into dynamics by including the "inertial force" $-m\mathbf{a}$ alongside the applied forces.

Joseph-Louis Lagrange, in Mechanique analytique (1788), built on d'Alembert's principle to develop the generalised coordinate formulation of mechanics. Lagrange introduced the multiplier method as a systematic algebraic procedure for handling constraints without identifying and eliminating them by coordinate choices. His key recognition was that the multipliers have a direct mechanical interpretation as constraint forces -- they are not merely computational devices but carry physical information.

The distinction between holonomic and non-holonomic constraints emerged gradually in the 19th century. Hertz, in Die Prinzipien der Mechanik (1894), gave the first systematic treatment of non-holonomic constraints and emphasised their fundamentally different character: holonomic constraints reduce the dimension of configuration space, while non-holonomic constraints restrict the possible paths without changing the accessible configurations.

Paul Dirac, in his 1950 paper "Generalized Hamiltonian Dynamics" and his Lectures on Quantum Mechanics (1964), developed the theory of constrained Hamiltonian systems. Dirac's classification of constraints into first-class and second-class, and the construction of the Dirac bracket, transformed the Lagrange multiplier from a computational tool into a structural element with deep physical meaning. First-class constraints generate gauge transformations -- their multipliers represent genuine physical freedom (gauge choice), while second-class constraints are eliminable.

A philosophical observation. The multiplier method reveals that constraints have a dual nature. As equations ($f=0$), they restrict the configuration space. As forces ($\lambda \nabla f$), they act on the system. The multiplier $\lambda$ is the link between these two aspects -- it is determined by the constraint equation but enters the dynamics as a force. This duality between kinematic restriction and dynamic enforcement is a structural feature of all constrained systems, from classical pendulums to gauge field theories. In modern language, it is an instance of duality between a constraint and its Lagrange dual variable, the same mathematical structure that appears in optimisation theory and convex analysis.

Bibliography Master

• D'Alembert, J. le R., Traite de dynamique (1743).
• Lagrange, J.-L., Mechanique analytique (1788).
• Hertz, H., Die Prinzipien der Mechanik (1894).
• Dirac, P. A. M., "Generalized Hamiltonian Dynamics," Canadian Journal of Mathematics 2 (1950), 129-148.
• Dirac, P. A. M., Lectures on Quantum Mechanics (Yeshiva University, 1964), Ch. 1.
• Arnold, V. I., Mathematical Methods of Classical Mechanics, 2nd ed. (Springer GTM 60, 1989), Section 3.
• Taylor, J. R., Classical Mechanics (University Science Books, 2005), Ch. 7.
• Goldstein, H., Poole, C. P. & Safko, J., Classical Mechanics, 3rd ed. (Pearson, 2002), Ch. 2.
• Landau, L. D. & Lifshitz, E. M., Mechanics, 3rd ed. (Course of Theoretical Physics Vol. 1, Pergamon, 1976), Sections 1, 25.
• Marsden, J. E. & Ratiu, T. S., Introduction to Mechanics and Symmetry, 2nd ed. (Springer TAM 17, 1999), Ch. 1.3-1.4.
• Bloch, A. M., Nonholonomic Mechanics and Control (Springer, 2003), Ch. 1-3.
• Henneaux, M. & Teitelboim, C., Quantization of Gauge Systems (Princeton University Press, 1992), Ch. 1-2.