Lie-Group-Informed Dilation Symmetry Breaking

• Lie-Group-Informed Dilation-Symmetry-Breaking formalism is a framework that employs one-parameter Lie groups to model scale invariance and systematically introduce controlled departures via learned parameters and regularization.
• It uses explicit infinitesimal generators, Lie derivatives, and crossover ansätze to adapt symmetry conditions for applications in machine learning, turbulence modeling, and representation theory.
• The formalism unifies symmetry enforcement, discovery, and breaking to enable interpretable extraction of scale-invariance features across diverse data-driven scientific and engineering problems.

The Lie-group-informed dilation-symmetry-breaking formalism denotes a class of constructions in which dilation is treated as a one-parameter Lie symmetry, its infinitesimal generator is represented explicitly, and departures from exact scale invariance are introduced in a controlled manner through learned parameters, regularization terms, crossover ansätze, or intertwining operators. In the cited literature, this structure appears in at least four technically distinct settings: neural discovery of one-parameter symmetry transformations, Lie-derivative-based symmetry enforcement and promotion in machine learning, multilayer turbulence modeling via local dilation invariance of order functions, and representation-theoretic symmetry breaking between induced modules of reductive Lie groups (Gabel et al., 2023, Otto et al., 2023, She et al., 2011, Kobayashi, 2013, Bi et al., 29 Aug 2025).

1. Mathematical primitives of the dilation formalism

In the neural symmetry-discovery formulation of Gabel et al., the starting point is a one-parameter subgroup

$$g(\theta)=\exp(\theta A)\in G\subset GL(n,\mathbb{R}),$$

with infinitesimal generator $A\in\mathfrak g$, acting on $x\in\mathbb{R}^n$ by

$g(\theta)x=e^{\theta A}x.$

Equivalently, the generator defines the ODE

$$\frac{dT(\theta)}{d\theta}=A\,T(\theta),\qquad T(0)=x,$$

whose solution is $T(\theta)=e^{\theta A}x$ (Gabel et al., 2023).

For the isotropic dilation group on $\mathbb{R}^n$, the finite action is

$g(\theta)x=e^\theta x,$

and the infinitesimal generator is

$A_d=I_n.$

The corresponding Lie algebra is one-dimensional, $\operatorname{span}\{A_d\}$, the bracket is trivial,

$A\in\mathfrak g$0

and the exponential map closes in elementary form,

$A\in\mathfrak g$1

(Gabel et al., 2023).

Otto et al. formulate the same symmetry on a vector bundle $A\in\mathfrak g$2 over $A\in\mathfrak g$3, with dilation group

$A\in\mathfrak g$4

The Lie algebra is $A\in\mathfrak g$5, with generator $A\in\mathfrak g$6 satisfying $A\in\mathfrak g$7. The base action is

$A\in\mathfrak g$8

and the fiber-linear lift is

$A\in\mathfrak g$9

where $x\in\mathbb{R}^n$0 may be $x\in\mathbb{R}^n$1 for invariant outputs or $x\in\mathbb{R}^n$2 for homogeneous outputs of degree $x\in\mathbb{R}^n$3 (Otto et al., 2023).

The infinitesimal generator on the base is the Euler vector field

$x\in\mathbb{R}^n$4

and the Lie derivative of a section $x\in\mathbb{R}^n$5 is

$x\in\mathbb{R}^n$6

In coordinates,

$x\in\mathbb{R}^n$7

(Otto et al., 2023).

These formulations exhibit a common structure: dilation symmetry is encoded by a one-parameter group, the Euler-type infinitesimal generator, and an explicit condition for exact invariance. A plausible implication is that the formalism is best understood not as a single model, but as a reusable symmetry scaffold adaptable to data analysis, PDE modeling, turbulence, and representation theory.

2. Data-driven discovery of dilation symmetry

Gabel et al. study paired data

$x\in\mathbb{R}^n$8

under the assumption

$x\in\mathbb{R}^n$9

for an unknown one-parameter subgroup $g(\theta)x=e^{\theta A}x.$0. The learning objective is twofold: to infer the infinitesimal generator $g(\theta)x=e^{\theta A}x.$1 and to infer the per-datum parameter $g(\theta)x=e^{\theta A}x.$2 together with its sampling distribution $g(\theta)x=e^{\theta A}x.$3 (Gabel et al., 2023).

The generator is parametrized in a fixed basis $g(\theta)x=e^{\theta A}x.$4 as

$g(\theta)x=e^{\theta A}x.$5

while an encoder network

$g(\theta)x=e^{\theta A}x.$6

maps each pair $g(\theta)x=e^{\theta A}x.$7 to an estimated group parameter $g(\theta)x=e^{\theta A}x.$8. Training enforces the reconstruction-by-transformation relation

$g(\theta)x=e^{\theta A}x.$9

through the loss

$$\frac{dT(\theta)}{d\theta}=A\,T(\theta),\qquad T(0)=x,$$0

augmented by a regularizer $$\frac{dT(\theta)}{d\theta}=A\,T(\theta),\qquad T(0)=x,$$1, such as $$\frac{dT(\theta)}{d\theta}=A\,T(\theta),\qquad T(0)=x,$$2, yielding

$$\frac{dT(\theta)}{d\theta}=A\,T(\theta),\qquad T(0)=x,$$3

(Gabel et al., 2023).

For the dilation group, the specialization is particularly simple. One fixes $$\frac{dT(\theta)}{d\theta}=A\,T(\theta),\qquad T(0)=x,$$4, writes

$$\frac{dT(\theta)}{d\theta}=A\,T(\theta),\qquad T(0)=x,$$5

and trains with

$$\frac{dT(\theta)}{d\theta}=A\,T(\theta),\qquad T(0)=x,$$6

After training, the parameter is rescaled according to

$$\frac{dT(\theta)}{d\theta}=A\,T(\theta),\qquad T(0)=x,$$7

so that $$\frac{dT(\theta)}{d\theta}=A\,T(\theta),\qquad T(0)=x,$$8 matches the observed scales. The empirical histogram of $$\frac{dT(\theta)}{d\theta}=A\,T(\theta),\qquad T(0)=x,$$9 then estimates $T(\theta)=e^{\theta A}x$0 (Gabel et al., 2023).

The key contribution of this formulation is that the symmetry-breaking stage can be informed by a learned scale distribution rather than by an externally imposed prior. In the language used later in the same exposition, this yields a coherent pipeline

$T(\theta)=e^{\theta A}x$1

(Gabel et al., 2023).

3. Soft symmetry breaking in machine learning and variational models

Otto et al. organize symmetry use in machine learning into three tasks: enforcing known symmetry, discovering unknown symmetry, and promoting symmetry while allowing data-supported breaking. For the dilation group, exact symmetry is enforced by the linear constraint

$T(\theta)=e^{\theta A}x$2

while discovery is phrased as the computation of the nullspace of the linear map

$T(\theta)=e^{\theta A}x$3

namely

$T(\theta)=e^{\theta A}x$4

The paper states that enforcing and discovering symmetry are dual under the bilinearity $T(\theta)=e^{\theta A}x$5 (Otto et al., 2023).

The exact codimension of symmetry is

$T(\theta)=e^{\theta A}x$6

but Otto et al. replace it by the convex surrogate

$T(\theta)=e^{\theta A}x$7

In the dilation case, $T(\theta)=e^{\theta A}x$8, so the nuclear norm reduces to

$T(\theta)=e^{\theta A}x$9

Minimizing this quantity promotes a large nullspace of $\mathbb{R}^n$0, and when $\mathbb{R}^n$1, one has

$\mathbb{R}^n$2

(Otto et al., 2023).

A concrete example is the one-hidden-layer network

$\mathbb{R}^n$3

on $\mathbb{R}^n$4 with invariant output, for which $\mathbb{R}^n$5 and $\mathbb{R}^n$6. The Lie derivative becomes

$\mathbb{R}^n$7

Full invariance imposes

$\mathbb{R}^n$8

which forces $\mathbb{R}^n$9, leaving $g(\theta)x=e^\theta x,$0 arbitrary. By contrast, the regularized objective adds

$g(\theta)x=e^\theta x,$1

so increasing the hyperparameter $g(\theta)x=e^\theta x,$2 pushes the model toward scale invariance without imposing it exactly (Otto et al., 2023).

The variational version of soft breaking, sketched from the one-parameter subgroup discovery framework, uses the learned dilation generator $g(\theta)x=e^\theta x,$3 and the learned scale distribution $g(\theta)x=e^\theta x,$4 to define

$g(\theta)x=e^\theta x,$5

where $g(\theta)x=e^\theta x,$6 is strictly scale-invariant under

$g(\theta)x=e^\theta x,$7

and

$g(\theta)x=e^\theta x,$8

The weight $g(\theta)x=e^\theta x,$9 is derived from the learned $A_d=I_n.$0, for instance

$A_d=I_n.$1

while $A_d=I_n.$2 is a local potential such as $A_d=I_n.$3 or $A_d=I_n.$4. In this construction, $A_d=I_n.$5 sets the direction in theory-space along which scale breaking is applied, $A_d=I_n.$6 localizes the breaking, and $A_d=I_n.$7 controls its amplitude (Gabel et al., 2023).

A recurrent theme across these formulations is that symmetry breaking is not treated as arbitrary violation. It is parameterized either by a Lie-derivative penalty or by a weight derived from a learned scale distribution, so that deviation from exact invariance remains structured and inspectable.

4. Dilation symmetry and order functions in wall-bounded turbulence

In the turbulence framework of She et al., the primary symmetry object is not the mean velocity $A_d=I_n.$8 but an order function, most prominently the stress-length function

$A_d=I_n.$9

together with the shear-induced eddy length

$\operatorname{span}\{A_d\}$0

where $\operatorname{span}\{A_d\}$1, $\operatorname{span}\{A_d\}$2, $\operatorname{span}\{A_d\}$3, and $\operatorname{span}\{A_d\}$4 is the dissipation (She et al., 2011).

The wall breaks translation normal to the wall, rotational isotropy that tilts into the wall, and shifts in $\operatorname{span}\{A_d\}$5. The remaining continuous symmetry is a one-parameter dilation acting on the wall-normal coordinate and the order function. In wall units, the framework uses the group invariants

$\operatorname{span}\{A_d\}$6

and invariant ansätze for $\operatorname{span}\{A_d\}$7 and $\operatorname{span}\{A_d\}$8 generate the layerwise scaling laws (She et al., 2011).

The simplest case is $\operatorname{span}\{A_d\}$9, which yields

$A\in\mathfrak g$00

The exponent $A\in\mathfrak g$01 identifies the layer: $A\in\mathfrak g$02 and, in outer units for channel and pipe flows,

$A\in\mathfrak g$03

The diagnostic

$A\in\mathfrak g$04

shows plateaus at $A\in\mathfrak g$05, $A\in\mathfrak g$06, $A\in\mathfrak g$07, and $A\in\mathfrak g$08 in DNS data (She et al., 2011).

In the bulk zone, the symmetry of $A\in\mathfrak g$09 itself is broken while its derivative remains nearly invariant. Taking $A\in\mathfrak g$10 leads to the defect law

$A\in\mathfrak g$11

with $A\in\mathfrak g$12 for channel and zero-pressure-gradient turbulent boundary layer, and $A\in\mathfrak g$13 for pipe flow (She et al., 2011).

Smooth transitions between adjacent power-law layers are obtained from a generalized invariant relation between $A\in\mathfrak g$14 and $A\in\mathfrak g$15, producing

$A\in\mathfrak g$16

where $A\in\mathfrak g$17 controls transition sharpness. These layer solutions are then combined through a multiplicative composite rule into a single analytic expression for $A\in\mathfrak g$18, with parameter classes given by scaling exponents, layer thicknesses, and transition sharpnesses (She et al., 2011).

This turbulence literature is central to the modern meaning of dilation-symmetry breaking because it makes the breaking physically interpretable. Exact dilation invariance is not assumed globally; it is recovered locally within layers and then broken, repaired, or interpolated according to the dominant balance of production, dissipation, and transport.

5. Non-equilibrium pressure-gradient turbulent boundary layers

Bi et al. extend the symmetry-based turbulence program to non-equilibrium pressure-gradient turbulent boundary layers through what they explicitly call a Lie-group-informed dilation-symmetry-breaking formalism. In wall units, the total shear stress is

$A\in\mathfrak g$19

and in zero-pressure-gradient equilibrium TBLs the leading inner and outer balances admit the dilation symmetry

$A\in\mathfrak g$20

with infinitesimal generator

$A\in\mathfrak g$21

(Bi et al., 29 Aug 2025).

An adverse pressure gradient introduces a new length scale $A\in\mathfrak g$22, defined as the height where the pressure-gradient-induced shear stress $A\in\mathfrak g$23 becomes $A\in\mathfrak g$24. The appearance of the ratio $A\in\mathfrak g$25 breaks the original dilation symmetry. To recover quasi-invariance, the formalism uses the crossover ansatz

$A\in\mathfrak g$26

which interpolates between two power laws and solves the renormalization-group-type ODE

$A\in\mathfrak g$27

(Bi et al., 29 Aug 2025).

Successive application of this ansatz yields the multilayer defect scaling of the total shear stress. For equilibrium APG TBLs, the two-layer form is

$A\in\mathfrak g$28

with

$A\in\mathfrak g$29

The first factor encodes the inner-layer pressure-gradient linear law, while the second imposes the outer $A\in\mathfrak g$30 defect (Bi et al., 29 Aug 2025).

For gradually varying APG, a third layer is added through a delay function at $A\in\mathfrak g$31,

$A\in\mathfrak g$32

The paper states that one typically sets $A\in\mathfrak g$33, $A\in\mathfrak g$34 for gradually strengthening APG, preserving the near-wall linear law while delaying or overshooting the outer response depending on the sign of $A\in\mathfrak g$35 (Bi et al., 29 Aug 2025).

Abrupt pressure-gradient changes produce a stronger form of symmetry breaking because the inner and outer regions adapt on different timescales,

$A\in\mathfrak g$36

The consequence is a dual-boundary-layer decomposition,

$A\in\mathfrak g$37

with an internal-boundary-layer stress $A\in\mathfrak g$38 and a history-dependent outer residual stress $A\in\mathfrak g$39. The four physically interpretable parameters

$A\in\mathfrak g$40

then characterize the decoupled three-layer TSS (Bi et al., 29 Aug 2025).

A further element is the wall-normal-dependent velocity scale $A\in\mathfrak g$41, introduced to collapse all non-equilibrium TSS profiles to

$A\in\mathfrak g$42

The authors validate this framework on wing-section TBLs over NACA4412 and NACA0012, a relaxing TBL over a Gaussian bump, and converging-diverging channel flow, and they propose the resulting parameterizations as structural priors for machine learning of complex wall turbulence (Bi et al., 29 Aug 2025).

6. Symmetry-breaking operators in representation theory

Kobayashi’s F-method places symmetry breaking in a representation-theoretic setting. Let $A\in\mathfrak g$43 be a real reductive linear Lie group, $A\in\mathfrak g$44 a parabolic subgroup, and $A\in\mathfrak g$45 a finite-dimensional representation of $A\in\mathfrak g$46. One forms the homogeneous bundle

$A\in\mathfrak g$47

whose smooth sections carry the principal-series representation

$A\in\mathfrak g$48

For a one-dimensional character $A\in\mathfrak g$49, this becomes

$A\in\mathfrak g$50

(Kobayashi, 2013).

The problem is to classify $A\in\mathfrak g$51-intertwining operators

$A\in\mathfrak g$52

between induced representations of $A\in\mathfrak g$53 and a reductive subgroup $A\in\mathfrak g$54. Kobayashi gives geometric criteria for finiteness. For continuous operators, $A\in\mathfrak g$55 for all $A\in\mathfrak g$56 if and only if there is an open $A\in\mathfrak g$57-orbit in $A\in\mathfrak g$58. For differential operators, finiteness follows when $A\in\mathfrak g$59 is $A\in\mathfrak g$60-compatible (Kobayashi, 2013).

The F-method passes to the big Bruhat cell, identifies the distribution kernel of an intertwiner with a tempered distribution on the nilpotent model, and applies the algebraic Fourier transform

$A\in\mathfrak g$61

The transformed kernel satisfies a finite system of linear PDEs expressing $A\in\mathfrak g$62-equivariance and $A\in\mathfrak g$63-annihilation. In this way, branching problems are converted into solvable Fourier-side equations (Kobayashi, 2013).

For the conformal pair

$A\in\mathfrak g$64

the principal series $A\in\mathfrak g$65 on densities of weight $A\in\mathfrak g$66 is realized on $A\in\mathfrak g$67, and there is an integral symmetry-breaking operator

$A\in\mathfrak g$68

with kernel

$A\in\mathfrak g$69

up to an explicit $A\in\mathfrak g$70-factor. When $A\in\mathfrak g$71, the meromorphic continuation has a simple pole, and the residue is a local differential operator

$A\in\mathfrak g$72

often called a Juhl operator (Kobayashi, 2013).

This literature uses the phrase “symmetry breaking operator” in a stricter sense than the data-driven or turbulence literatures. Here the breaking is not a penalty or crossover ansatz; it is an intertwiner between different representation spaces, frequently controlled by meromorphic parameters and residue calculus.

7. Scope, interpretations, and common misconceptions

One recurrent misconception is that dilation-symmetry breaking means unconstrained violation of scale invariance. The cited constructions do not support that interpretation. In the data-driven subgroup framework, the breaking term is weighted by the learned distribution $A\in\mathfrak g$73, so under-represented scales can be penalized more heavily through $A\in\mathfrak g$74 (Gabel et al., 2023). In the Lie-derivative framework, the hyperparameter $A\in\mathfrak g$75 explicitly trades off data fit against the nuclear-norm surrogate for symmetry codimension (Otto et al., 2023). In the turbulence literature, breaking enters through specific crossover scales such as $A\in\mathfrak g$76, $A\in\mathfrak g$77, $A\in\mathfrak g$78, and $A\in\mathfrak g$79, not through unrestricted empirical correction terms (Bi et al., 29 Aug 2025).

A second misconception is that the symmetry must act on the observable of immediate interest, such as the mean velocity profile. She et al. emphasize the opposite strategy: the stress length $A\in\mathfrak g$80 and related order functions are the primary symmetry objects, and the mean velocity is reconstructed only after the invariant or defect structure of the order function has been established (She et al., 2011). This shift from the field itself to an order function is one of the defining innovations of the wall-turbulence branch of the subject.

A third misconception is that enforcing, discovering, and breaking symmetry are unrelated tasks. Otto et al. make them parts of a single framework centered on the Lie derivative: exact invariance is the condition $A\in\mathfrak g$81, discovery is the computation of $A\in\mathfrak g$82, and soft breaking is controlled by minimizing $A\in\mathfrak g$83 rather than setting it identically to zero (Otto et al., 2023). The Gabel et al. program is compatible with this viewpoint, because it first identifies the generator and the parameter distribution and only then uses them to build downstream symmetry-breaking terms (Gabel et al., 2023).

The available literature also suggests that there is no single canonical object universally called the Lie-group-informed dilation-symmetry-breaking formalism. Rather, the phrase designates a family of methods sharing the same backbone: a one-parameter dilation group, an explicit infinitesimal generator, and a structured mechanism for departing from exact scale invariance. Across machine learning, turbulence, and representation theory, the main differences lie in the object being acted upon—data pairs, bundle sections, order functions, or induced representations—and in the mechanism by which the breaking is encoded—distribution-weighted potentials, convex regularization, multilayer defect laws, or intertwining operators (Gabel et al., 2023, Otto et al., 2023, She et al., 2011, Kobayashi, 2013, Bi et al., 29 Aug 2025).