Symmetries and Continuous Transformations

The audit you ran in worked example 2 is now a design principle in machine learning. The most accurate interatomic force fields of the mid-2020s — the equivariant neural network potentials powering molecular dynamics and materials discovery — are architecturally forced to respect the Euclidean symmetries: their internal features transform in irreducible representations of rotations, so a rotated molecule provably produces rotated forces, exactly and by construction rather than approximately and by training data. The measured payoff of hard-wiring the symmetry is one to two orders of magnitude in data efficiency, which is why "E(3)-equivariant" appears in the title of half the field's benchmark papers. The same logic runs through simulation itself: geometric integrators that respect a system's symmetries conserve the corresponding momenta exactly in floating point — a discrete shadow of the symmetry-conservation pairing — which is why solar-system and plasma codes hold their angular momentum for billions of steps. At the other end of the scale, symmetry audits remain physics' sharpest probe of new law. Tests of Lorentz and CPT invariance — clock-comparison experiments, antihydrogen spectroscopy at CERN — are precision searches for a failing audit, a $\delta L \neq \mathrm{d}F/\mathrm{d}t$ term in nature's Lagrangian, with bounds now at the $10^{-18}$ level. The working assumption of all of it is the pairing this article set up: find the symmetry group, and you have found what the world must conserve.

A continuous transformation comes with a dial. Set the dial to zero and nothing happens — the identity. Turn it a little and every point of the system's history shifts a little; turn it more and the shift grows, smoothly, with no jumps. A rotation has a dial (the angle), a translation has a dial (the distance), a boost has a dial (the relative velocity). A mirror reflection does not: there is no such thing as "reflecting by ten percent." That single geometric fact — dial or no dial — will decide, in the next article, which symmetries pay out conserved quantities and which merely sort solutions into classes.

What should "symmetry" mean, once you have a dial? Not "the trajectory maps to itself" — a rotated Kepler ellipse is a different ellipse. The right demand is that the dial maps solutions to solutions: every image of a legal motion is again a legal motion, so the transformed world is dynamically indistinguishable from the original. Since the equations of motion descend from the action, the clean sufficient criterion lives there: turning the dial must not change the action in any way the variational principle can see. There is a beautifully visual way to hold the strict-invariance case. Draw the level sets of the potential. A transformation that slides points along the level sets changes nothing about the energy landscape — it is a symmetry. A circular bowl is swept into itself by rotations: every point moves, every level set maps onto itself. Squash the bowl into an ellipse and the same rotation now carries points across contour lines — uphill on one diagonal, downhill on the other — and rotational symmetry is gone, along with (as the next article will prove and worked example 2 will quantify) the conservation of angular momentum.

But strict invariance cannot be the whole story, and the gliding train proves it. Boost a free particle and its kinetic energy — its entire Lagrangian — plainly changes. Yet a boosted straight line is still a straight line: solutions map to solutions perfectly. The resolution is that the change in $L$ is a total time derivative, which piles up only at the endpoints of the action integral where the variational principle never looks. Transformations with this property are called quasi-invariances, and admitting them is not a loophole but a necessity: reject them and you reject relativity's entire nonrelativistic shadow.

Fix a system with generalized coordinates $q = (q_1, \dots, q_n)$. A one-parameter family of transformations assigns to each value of a real dial $\varepsilon$ a smooth map of configurations (and possibly of time),

where: $\varepsilon = 0$ is the identity, $\xi_i$ is the infinitesimal generator of the coordinate change and $\tau$ that of the time change. The generator is simply the velocity of the family as the dial starts to turn: $\xi_i = \mathrm{d}q_i'/\mathrm{d}\varepsilon\,\big|_{\varepsilon=0}$. Two facts make generators the whole story. Successive small turns compose by adding generators, so the finite transformation is recovered by "exponentiating" the infinitesimal one — a rotation by a finite angle is many tiny rotations compounded. And every question about symmetry can be answered at first order in $\varepsilon$, which is why the calculus below never needs the finite maps at all. The standard dials of mechanics, in generator form: a spatial translation along $\hat{\mathbf n}$ has $\boldsymbol\xi_a = \hat{\mathbf n}$ for every particle $a$ and $\tau = 0$; a time translation has $\xi_i = 0$, $\tau = 1$; a boost along $\hat{\mathbf n}$ has $\boldsymbol\xi_a = \hat{\mathbf n}\,t$, $\tau = 0$. For a rotation by angle $\theta$ about the axis $\hat{\mathbf n}$, the finite map is the rotation matrix $R(\hat{\mathbf n}, \theta)$ and the generator statement is

where: the second relation follows by differentiating the first — velocities rotate along with positions, a fact used constantly below. Note the cross product's geometry: each point moves perpendicular both to the axis and to its own position vector, which is exactly "sliding along the circle." Now the criterion. A family (1) is a symmetry of the system when the Lagrangian transforms as

where: $F(q, t)$ is the boundary function, and the factor $\mathrm{d}t'/\mathrm{d}t$ accounts for the stretched time measure when $\tau \neq 0$. Strict invariance is the case $F = 0$; quasi-invariance is $F \neq 0$. The action then shifts by $\varepsilon\,[F]_{t_1}^{t_2}$ — pure endpoint bookkeeping. Why endpoint terms are dynamically invisible is the business of Hamilton's principle (fixed-endpoint variations never see them), and the fact that Lagrangians differing by $\mathrm{d}F/\mathrm{d}t$ generate identical Euler–Lagrange equations is treated in full in the article on Lagrangian gauge freedom — here we simply use both facts. The working test in practice: compute $\delta L$ to first order in the dial; if what you get can be written as $\varepsilon$ times a total time derivative of some function of $q$ and $t$, the audit passes.

Quasi-invariance is not a boost-only curiosity. Translate a charged particle in a uniform magnetic field (gauge $\mathbf A = (0, Bx, 0)$) along $\hat x$ and the Lagrangian shifts by a total derivative — the audit passes only in the quasi mode, and the associated conserved quantity is the magged-up momentum $p_y = m\dot y + qBx$ whose consequences the cyclic-coordinates article works out for the cyclotron orbit. Whenever a Lagrangian contains velocity-linear terms, expect symmetries to pass with $F \neq 0$.

Which dial does what

One more structural fact separates continuous symmetries from a mere checklist: they compose, and the composition is itself constrained. Apply two infinitesimal transformations in both orders and subtract — if each order is a symmetry, so is the residue, and the residue is again infinitesimal. The set of generators therefore closes under the commutator,

where: $[a, b]$ denotes the bracket of generators — the defining structure of a Lie algebra, the tangent-space skeleton from which the whole continuous group can be rebuilt. This is why classifying the symmetries of a mechanical system is a finite, algebraic task rather than an endless search: for a closed nonrelativistic system the answer is always some subalgebra of the ten-generator Galilean algebra (three translations, three rotations, three boosts, one time translation).

1 · Auditing a boost with actual numbers (confidence builder)

A free particle of mass $m = 2\ \mathrm{kg}$ travels at $\dot x = 3\ \mathrm{m/s}$ for $t \in [0, 2]\ \mathrm{s}$. Boost by $w = 0.5\ \mathrm{m/s}$. Compute the action before and after, and verify that the change is exactly the boundary function of Derivation B — not approximately, exactly.

Step 1. Action on the original path: $S = \int_0^2 \tfrac12 (2)(3)^2\,\mathrm{d}t = 9 \times 2 = \boxed{18\ \mathrm{J\cdot s}}$.

Step 2. The boosted path is $x'(t) = 3t + 0.5t = 3.5t$ — still a straight line, so still a solution (symmetries map solutions to solutions). Its action: $S' = \int_0^2 \tfrac12(2)(3.5)^2\,\mathrm{d}t = 12.25 \times 2 = \boxed{24.5\ \mathrm{J\cdot s}}$. The action visibly changed: $\Delta S = 6.5\ \mathrm{J\cdot s}$.

Step 3. Now the audit. The finite-boost boundary function from Derivation B, Step 5, is $F = mwx + \tfrac12 mw^2 t$. Evaluate at the endpoints, using $x(0) = 0$, $x(2) = 6\ \mathrm{m}$: $F(t_2) = (2)(0.5)(6) + \tfrac12(2)(0.25)(2) = 6 + 0.5 = 6.5$, and $F(t_1) = 0$. So $F(t_2) - F(t_1) = \boxed{6.5\ \mathrm{J\cdot s}} = \Delta S$ exactly.

Step 4. Interpret. Along the path, $\delta L = mw\dot x + \tfrac12 mw^2 = 3.25\ \mathrm{J}$ at every instant — never zero, never small. Yet all $6.5\ \mathrm{J\cdot s}$ of accumulated change is endpoint data: freeze the endpoints, as every variation in Hamilton's principle does, and the transformed action ranks paths identically. That is quasi-invariance seen in the wild: the dynamics is blind to what the Lagrangian plainly shows.

2 · The lopsided trap: a broken symmetry, quantified (exam level)

A particle of mass $m = 0.5\ \mathrm{kg}$ moves in the anisotropic trap $V = \tfrac12 k(x^2 + 4y^2)$ with $k = 8\ \mathrm{N/m}$. (a) Run the rotation audit and find $\delta L$. (b) Show the failure is not a hidden quasi-invariance. (c) At the instant when $(x, y) = (0.3, 0.2)\ \mathrm{m}$ and $(\dot x, \dot y) = (1, -1)\ \mathrm{m/s}$, compute $L_z$ and its decay rate, and cross-check against the torque.

Step 1. Audit. Under $\delta x = -\theta y$, $\delta y = \theta x$ the kinetic term dies as always (Derivation A, Step 2). The potential does not: $\delta V = kx\,\delta x + 4ky\,\delta y = kx(-\theta y) + 4ky(\theta x) = 3k\theta\,xy$. So $\delta L = -3k\theta\,xy$ — the mismatch is proportional to the anisotropy (the "3" is the gap between the spring constants) and to the product $xy$, vanishing only on the axes where the ellipse locally mimics a circle.

Step 2. Could $-3kxy$ be $\mathrm{d}F/\mathrm{d}t$ for some $F(q, t)$? No: a total derivative of $F(x, y, t)$ is $F_x\dot x + F_y\dot y + F_t$, linear in the velocities — but $-3kxy$ contains no velocities at all and is not identically zero, so no $F$ exists. The audit fails outright; rotation is not a symmetry of this system in any mode, strict or quasi.

Step 3. Quantify the instant. Angular momentum: $L_z = m(x\dot y - y\dot x) = 0.5\,[(0.3)(-1) - (0.2)(1)] = \boxed{-0.25\ \mathrm{kg\,m^2/s}}$.

Step 4. Decay rate from the broken audit. The general rate-of-change theorem (proved via Noether's machinery next article, or directly from the equations of motion) says $\mathrm{d}L_z/\mathrm{d}t$ equals the torque: with $\mathbf F = -\nabla V = (-kx, -4ky)$, $\tau_z = xF_y - yF_x = x(-4ky) - y(-kx) = -3k\,xy$ — precisely the coefficient the audit produced in Step 1, with $\theta$ stripped. Numerically: $\tau_z = -3(8)(0.3)(0.2) = \boxed{-1.44\ \mathrm{N\cdot m}}$, so at this instant $L_z$ is draining at $1.44$ units per second from an inventory of $-0.25$ — the "conserved" quantity turns over on a tenth-of-a-second timescale. The broken-symmetry coefficient is the non-conservation rate; auditing $\delta L$ and computing the torque are the same calculation wearing two costumes.

Step 5. Sanity checks. Set the anisotropy to zero (spring constants equal): $\delta V \propto (1-1)xy = 0$, the audit passes, and $\tau_z = 0$ — the isotropic oscillator conserves $L_z$, as its closed elliptical orbits suggest. And dimensionally $[3kxy] = \mathrm{N\cdot m}$, a torque, as it must be.

References. Goldstein, Poole & Safko, Classical Mechanics (3rd ed.), Ch. 2 and Sec. 13.7 (Noether's theorem) · Landau & Lifshitz, Mechanics (Vol. 1, 3rd ed.), Sec. 6-9 (conservation laws from homogeneity and isotropy) · Morin, Introduction to Classical Mechanics, Ch. 5 (Conservation laws and Noether's theorem) · Taylor, Classical Mechanics, Ch. 7 (Lagrange's equations and conservation laws) · Arnold, Mathematical Methods of Classical Mechanics, Ch. 4 (symmetry groups and Noether)

Provenance. Symmetry condition and Noether charge derivation verified against Goldstein, Poole & Safko Sec. 13.7 and Arnold Ch. 4. Energy-function result $\mathrm dh/\mathrm dt=-\partial L/\partial t$ from Goldstein Ch. 2. Galilean-boost total-derivative argument and the derivation of $T=\tfrac12 mv^2$ from homogeneity/isotropy follow Landau & Lifshitz Mechanics Secs. 4-8. Central-force cyclic-coordinate example follows Morin Ch. 5 and Taylor Ch. 7. A verifier should check: (1) the Noether charge reduces to $p_k$ for pure coordinate shifts and to $-h$ for pure time translation; (2) $\dot{\mathbf G}=0$ for the boost; (3) $p_\phi=mr^2\dot\phi$ for the central-force azimuth. No fabricated numerical constants are used. Worked-example numbers on this page (boost audit: $S = 18 \to 24.5$ J·s with $\Delta S = F(t_2)-F(t_1) = 6.5$ J·s for $m=2$ kg, $\dot x = 3$ m/s, $w = 0.5$ m/s; lopsided trap: $\delta L = -3k\theta xy$, $L_z = -0.25$ kg·m²/s and $\tau_z = -1.44$ N·m at the quoted state) are illustrative arithmetic recomputed from the displayed formulas.

Builds on: The Lagrangian Function · The Action Functional · The Euler-Lagrange Equations · Generalized Coordinates and Configuration Space · Hamilton's Principle of Stationary Action Leads to: Cyclic Coordinates and Conservation of Conjugate Momenta · Generalized Forces and Generalized Momenta · The Energy Function and Jacobi Integral