Lie Symmetry Group

• Lie symmetry groups are smooth transformation groups that map solutions of differential equations to solutions, preserving the underlying geometric or physical structures.
• They utilize infinitesimal generators and prolongation techniques to derive determining equations, enabling systematic symmetry reduction and classification of invariant solutions.
• Applications extend from classic differential equation analysis to advanced fields such as symmetry-informed machine learning, geometric modeling, and topological phases.

A Lie symmetry group is a Lie transformation group acting smoothly on a manifold or on the space of independent and dependent variables of a differential equation, with the defining property that the action preserves the relevant structure. In the differential-equation setting, a symmetry group maps solutions to solutions; in infinitesimal form, it is encoded by vector fields whose prolongations annihilate the equation on its solution manifold. The same Lie-theoretic language also appears in orbit theory, geometric models, measurement invariance, topological phases, and symmetry-informed machine learning, where infinitesimal generators, invariant distributions, and group actions remain the organizing concepts (Kunzinger, 2015, Güngör, 2019, Shaw et al., 13 May 2025).

1. Formal structure of a Lie symmetry group

A Lie group $G$ acts as a Lie transformation group on a manifold $M$ when there is a smooth surjection

$\Phi: G \times M \to M$

such that

$\Phi(g,\Phi(h,m))=\Phi(gh,m), \qquad \Phi(e,m)=m .$

Writing $g\cdot m:=\Phi(g,m)$ gives the usual action notation. The action is effective if the only element acting identically is the identity, and free if only the identity fixes points in $M$. For $m\in M$, the orbit $G\cdot m$ is the set of points reachable from $m$, and the isotropy group is $G_m=\{g\in G\mid g\cdot m=m\}$; orbits are immersed submanifolds diffeomorphic to $M$0 (Kunzinger, 2015).

The infinitesimal side is obtained from the Lie algebra $M$1 of $M$2. Each $M$3 induces a vector field on $M$4,

$M$5

and the span of these vector fields forms the Killing algebra $M$6. The correspondence $M$7 is a Lie algebra homomorphism,

$M$8

This is the basic passage from a global symmetry group to its infinitesimal generators (Kunzinger, 2015).

The orbit picture is not restricted to constant-rank situations. The Stefan-Sussmann theory, as presented in the lecture notes, extends Frobenius-type integrability to distributions of non-constant rank and shows that the orbits of pseudogroups generated by vector fields are immersed submanifolds and maximal integral manifolds of the smallest invariant distribution containing those fields. In symmetry analysis, this places orbit structure and symmetry reduction into a geometric framework that remains valid even for singular distributions (Kunzinger, 2015).

2. Infinitesimal generators, prolongation, and determining equations

For a system

$M$9

a symmetry group is a local transformation group acting on the space of independent and dependent variables and mapping solutions to solutions. Its infinitesimal generator has the local form

$\Phi: G \times M \to M$0

Because differential equations involve derivatives, the action must be prolonged to jet space. The $\Phi: G \times M \to M$1-th prolongation $\Phi: G \times M \to M$2 acts on derivatives up to order $\Phi: G \times M \to M$3, and the infinitesimal symmetry criterion is

$\Phi: G \times M \to M$4

Solving the resulting overdetermined linear PDE system for the coefficients $\Phi: G \times M \to M$5 yields the determining equations for the Lie symmetry algebra (Güngör, 2019).

For first-order ODEs,

$\Phi: G \times M \to M$6

the infinitesimal generator is

$\Phi: G \times M \to M$7

and the first prolongation is

$\Phi: G \times M \to M$8

The invariance condition

$\Phi: G \times M \to M$9

reduces to the linearized symmetry condition

$\Phi(g,\Phi(h,m))=\Phi(gh,m), \qquad \Phi(e,m)=m .$0

The paper on first-order ODEs explicitly states that there is a one-to-one correspondence between infinitesimal generators and Lie symmetries in this setting (Mwanzia et al., 2013).

This infinitesimal formalism also underlies more recent vector-field formulations. For a one-parameter Lie group action generated by

$\Phi(g,\Phi(h,m))=\Phi(gh,m), \qquad \Phi(e,m)=m .$1

its action on a function $\Phi(g,\Phi(h,m))=\Phi(gh,m), \qquad \Phi(e,m)=m .$2 is

$\Phi(g,\Phi(h,m))=\Phi(gh,m), \qquad \Phi(e,m)=m .$3

and $\Phi(g,\Phi(h,m))=\Phi(gh,m), \qquad \Phi(e,m)=m .$4 is invariant under the flow if $\Phi(g,\Phi(h,m))=\Phi(gh,m), \qquad \Phi(e,m)=m .$5 everywhere. This is the same annihilation condition that appears in classical symmetry analysis, recast in a form suited to data-driven settings (Shaw et al., 13 May 2025).

3. Reduction, invariant solutions, and optimal systems

Once a Lie symmetry algebra is known, it can be used to reduce differential equations. For ODEs, a one-dimensional symmetry group permits reduction of order by one through canonical coordinates. For PDEs, one seeks group-invariant or similarity solutions by solving the characteristic equations associated with a chosen generator and substituting the resulting ansatz into the original system. Differential invariants satisfy

$\Phi(g,\Phi(h,m))=\Phi(gh,m), \qquad \Phi(e,m)=m .$6

and they organize invariant equations and reductions (Güngör, 2019).

A central classification device is the optimal system of subalgebras. The adjoint representation,

$\Phi(g,\Phi(h,m))=\Phi(gh,m), \qquad \Phi(e,m)=m .$7

is used to classify subalgebras up to conjugacy, so that one obtains a minimal list of non-equivalent symmetry reductions. This is the standard mechanism behind preliminary classification of group-invariant solutions in the radiation natural convection flow equation and analogous problems (Nadjafikhah et al., 2011).

For the general short pulse equation

$\Phi(g,\Phi(h,m))=\Phi(gh,m), \qquad \Phi(e,m)=m .$8

the one-dimensional optimal system is

$\Phi(g,\Phi(h,m))=\Phi(gh,m), \qquad \Phi(e,m)=m .$9

while the two-dimensional optimal system is

$g\cdot m:=\Phi(g,m)$0

These are obtained from the adjoint representation of the three-dimensional Lie algebra of point symmetries and provide the canonical starting point for symmetry reduction and invariant-solution classification (Nadjafikhah, 2012).

The same strategy appears in the radiation natural convection flow system. After computing the Lie algebra and its commutators, the adjoint action reduces a general one-dimensional subalgebra to an optimal list: $g\cdot m:=\Phi(g,m)$1 Characteristic equations then yield explicit invariant forms, such as solutions depending only on $g\cdot m:=\Phi(g,m)$2 for $g\cdot m:=\Phi(g,m)$3, or similarity variables like $g\cdot m:=\Phi(g,m)$4 for $g\cdot m:=\Phi(g,m)$5 (Nadjafikhah et al., 2011).

4. Finite-dimensional, higher-order, and infinite-dimensional symmetry algebras

Concrete equations exhibit markedly different symmetry types. For the short pulse equation

$g\cdot m:=\Phi(g,m)$6

the maximal Lie point symmetry algebra is three-dimensional, generated by

$g\cdot m:=\Phi(g,m)$7

with commutation relations

$g\cdot m:=\Phi(g,m)$8

The same analysis shows that there are no new non-point contact symmetries, but generalized symmetry analysis produces two new third-order local symmetry generators,

$g\cdot m:=\Phi(g,m)$9

$M$0

The paper also states that there are no non-trivial second- or fourth-order local symmetries, so the first genuinely new local symmetries occur at third order (Nadjafikhah, 2012).

Finite-dimensional Lie point symmetry groups also arise in fluid and transport models. For the radiation natural convection flow system, the computed Lie algebra is spanned by

$M$1

with corresponding one-parameter groups involving translations in $M$2 and $M$3, and anisotropic scalings of $M$4 (Nadjafikhah et al., 2011).

By contrast, integrable $M$5-dimensional systems can have infinite-dimensional Lie symmetry algebras. For the dispersionless Davey-Stewartson system, the infinitesimal generator has the form

$M$6

where $M$7 are arbitrary smooth functions. The algebra has a Kac-Moody-Virasoro structure,

$M$8

with the Witt relation

$M$9

and Kac-Moody-type commutators for $m\in M$0. The full symmetry group splits into connected and discrete parts,

$m\in M$1

where $m\in M$2 and $m\in M$3 is an infinite-dimensional connected component (Güngör et al., 2021).

The general Liénard-type equation

$m\in M$4

illustrates a different phenomenon: autonomy always yields the symmetry $m\in M$5, but the existence of a second Lie point symmetry is exceptional and is completely characterized by explicit conditions on the coefficient functions $m\in M$6. When those conditions hold, the symmetry algebra is two-dimensional; otherwise it is one-dimensional (Figula et al., 2019).

These examples correct two recurrent oversimplifications. First, Lie symmetry groups need not be finite-dimensional; the dDS system has an infinite-dimensional Kac-Moody-Virasoro algebra (Güngör et al., 2021). Second, enlarging the ansatz from point to contact symmetries does not automatically produce new symmetries; for the short pulse equation all contact symmetries reduce to point symmetries (Nadjafikhah, 2012).

5. Geometric, algebraic, and physical extensions

Outside the direct solution theory of differential equations, Lie symmetry groups appear as symmetry actions on algebraic or geometric models. In the Lie model of the triangle, the completed differential graded Lie algebra

$m\in M$7

admits an action of the symmetric group $m\in M$8 compatible with the geometric symmetries of the triangle. The differential satisfies

$m\in M$9

with $G\cdot m$0 the Baker-Campbell-Hausdorff product, and the model contains a $G\cdot m$1-invariant Maurer-Cartan element interpreted algebraically as the barycenter. The same construction generalizes to circuits with $G\cdot m$2 vertices, where one obtains Maurer-Cartan elements invariant under the automorphism group $G\cdot m$3 (Buijs et al., 2018).

In homogeneous Riemannian geometry, the relevant objects are Killing fields and invariant distributions. For a homogeneous manifold $G\cdot m$4, the distribution of symmetry is

$G\cdot m$5

and its rank is the index of symmetry. For naturally reductive nilpotent Lie groups, the distribution of symmetry coincides with the invariant distribution induced by the fixed vectors of the isotropy representation; the leaves of symmetry are Euclidean spaces (Reggiani, 2017). For three-dimensional unimodular and solvable Lie groups with left-invariant metrics, the index of symmetry is completely classified, and in both settings the cited papers emphasize that the index is never equal to $G\cdot m$6 (Reggiani, 2016, May, 2021).

Lie symmetry groups also govern invariance questions in applied domains. In the matrix Lie group model of measurement symmetries, score transformations are represented by

$G\cdot m$7

with

$G\cdot m$8

Under exactly these symmetry conditions, the standardized mean difference is invariant across measures; when symmetry is broken, the paper reports that score distribution symmetry and effect-size comparability are broken as well (Nugent, 18 Dec 2025).

In the theory of uniform discrete defective crystals, the discrete material structure is modeled by a uniform discrete subgroup $G\cdot m$9 of an ambient Lie group $m$0. Symmetries are changes of generators preserving the discrete set of points, and automorphisms of $m$1 that extend uniquely to automorphisms of $m$2 are interpreted as elastic symmetries; other symmetries are classified as inelastic (Nicks, 2013).

In $m$3d topological phases, gauging a compact, connected Lie group symmetry $m$4 is formulated through a central extension

$m$5

and a two-step construction

$m$6

Consistency imposes compatibility conditions between symmetry fractionalization and Hall conductance, and failure of those conditions signals an 't Hooft anomaly with a corresponding $m$7d topological term (Cheng et al., 2022).

6. Symmetry discovery, enforcement, and computational reformulations

Recent work reframes Lie symmetry groups in terms of learnable infinitesimal generators. One approach considers a one-parameter subgroup

$m$8

with generator $m$9 and action

$G_m=\{g\in G\mid g\cdot m=m\}$0

From paired data $G_m=\{g\in G\mid g\cdot m=m\}$1, neural models jointly learn the generator of the symmetry transformation and the per-sample parameters $G_m=\{g\in G\mid g\cdot m=m\}$2, either directly in ambient space or through an autoencoder latent space. The method is designed to recover both the structure of the one-parameter subgroup and the distribution of transformation parameters without presupposing whether the symmetry is a rotation, translation, or another affine transformation (Gabel et al., 2023).

A related vector-field formulation searches for all

$G_m=\{g\in G\mid g\cdot m=m\}$3

such that

$G_m=\{g\in G\mid g\cdot m=m\}$4

This is implemented by estimating derivatives of $G_m=\{g\in G\mid g\cdot m=m\}$5, parameterizing the coefficients $G_m=\{g\in G\mid g\cdot m=m\}$6, and solving a constrained optimization problem of the form $G_m=\{g\in G\mid g\cdot m=m\}$7. The same paper proposes symmetry enforcement by adding a regularizer penalizing deviations from $G_m=\{g\in G\mid g\cdot m=m\}$8,

$G_m=\{g\in G\mid g\cdot m=m\}$9

and further studies restriction of the search space to infinitesimal isometries satisfying the Killing equation

$M$00

This formulation explicitly extends symmetry discovery to non-affine symmetries and to functions defined implicitly by neural networks (Shaw et al., 13 May 2025).

The Lie algebra convolutional network goes further by replacing explicit group discretization with infinitesimal generators. Its core layer is

$M$01

and the paper states that multilayer L-conv can approximate group convolutional layers, with CNNs and graph convolutional networks arising as special cases. The same framework connects invariant loss functions to field-theoretic Lagrangians, relates robustness to Euler-Lagrange equations, and associates equivariance with conservation laws and Noether currents (Dehmamy et al., 2021).

In physics-informed neural networks, Lie symmetry information can also be embedded directly into the collocation process. For Burgers’ equation, the modified loss augments the standard initial, boundary, and residual terms with a symmetry term evaluated at infinitesimally transformed collocation points,

$M$02

where

$M$03

The paper presents this as a use of infinitesimal generators not for classical symmetry reduction but for generating additional constraints during training (Shah et al., 30 Sep 2025).

Taken together, these computational developments preserve the classical emphasis on infinitesimal generators, invariance conditions, and orbit directions. A plausible implication is that the modern machine-learning literature is not replacing Lie symmetry groups so much as reparameterizing them in operator, vector-field, and optimization language. The underlying mathematical objects remain the same: smooth actions, infinitesimal generators, and invariance constraints (Shaw et al., 13 May 2025, Dehmamy et al., 2021).