Equivariant Representation Learning | Rotation, Translation, Reflection and Symmetry

When the world moves, a good representation must know what should remain the same, what should move with it, and what should change in a precisely predictable way.

Equivariant representation learning is about preserving the structure of transformations. Rotate a molecule and its energy should remain unchanged while its force vectors rotate. Translate an object and its class should remain the same while its coordinates move. Reflect a chiral molecule and some properties should change because a mirror image can be physically distinct. A symmetry-aware model is therefore not simply “invariant to transformations”. It must encode the correct transformation law for each quantity.

This article belongs to the World Representation & Cognitive Tools library. It extends the advanced representation branch from disentanglement and causal structure into symmetry: how should internal representations transform when the input is acted on by a known family of transformations?

1. The arrow that should move, and the answer that should not

Place a blue arrow on a square board. A camera looks straight down. The arrow points east. Rotate the board ninety degrees counter-clockwise. The blue paint has not changed. The object is still an arrow. But its direction in camera coordinates now points north. A representation that calls both images “the same” has preserved identity and destroyed orientation. A representation that treats every pixel as unrelated has preserved neither identity nor the known transformation.

This simple board separates three cases. A scalar invariant such as object class may stay unchanged. A vector such as direction should rotate. A coordinate position should translate when the board translates. If we ask one feature to behave like another, the representation can become mathematically tidy and semantically wrong.

Equivariance formalises the relationship. Let a transformation g act on an input x through a representation T_g. Let the same transformation induce an action S_g on the model’s output or internal features. A function f is equivariant when

f(T_g x) = S_g f(x)

Invariance is the special case in which S_g is the identity: the output does not change. Equivariance is more general. It says the output changes in a known way. That difference is the core of the subject.

2. Write both sides of the transformation contract

A symmetry claim is incomplete if it names only the input transformation. “Rotation invariant” does not say which quantity is being predicted. An object’s class may be invariant to camera rotation; a force vector is not. “Translation equivariant” must specify whether the translation acts on an image grid, point cloud, molecular coordinates or a continuous field, and what output representation should transform with it.

Write the contract as four pieces: input object, transformation group, feature type and expected output action. For a 2D image classifier: input is an image, transformation is a translation, output is a class score, and the desired final output is usually invariant. For a segmentation map: translating the image should translate the segmentation. For a molecular force predictor: translating or rotating the coordinates should translate or rotate the predicted force vectors appropriately, while total energy remains invariant.

This contract is a scientific boundary. It prevents symmetry from becoming an aesthetic preference. A dataset may contain approximate rather than exact symmetry. A camera may introduce gravity-dependent cues. A boundary may break translation symmetry. A chiral molecule may not be mirror-equivalent. The architecture should reflect the symmetry of the job, not the symmetry we wish the world had.

3. A group is more than a list of augmentations

Mathematically, symmetries are often organised by groups. A group contains transformations with four properties: closure, associativity, an identity element and inverses. Rotations in the plane form a group. Rigid motions combine rotations and translations. Reflections can be included to create larger symmetry groups. The power of the group language is compositional: if the system understands how two transformations combine, it can reason about the entire structured family rather than memorising a bag of augmentations.

Data augmentation and equivariant architecture are therefore not equivalent interventions. Augmentation shows the model transformed examples and encourages it to learn a desired relationship from data. Equivariant architecture constrains the function class so that the relationship holds by construction, subject to implementation details and the stated group. Augmentation can support approximate or broken symmetry; architectural equivariance can provide stronger guarantees and may also impose the wrong constraint if the assumed symmetry is false.

The foundational Group Equivariant Convolutional Networks work generalised convolution beyond ordinary translations by building feature maps indexed over transformation groups. The broader field is surveyed in Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges, which provides a common language for symmetry-aware architectures across structured domains.

4. Scalars, vectors, tensors and handed quantities are different representation types

A scalar such as mass does not acquire a new direction when the coordinate frame rotates. A vector such as velocity does. A second-order tensor such as a stress tensor transforms by a different rule. A pseudoscalar or pseudovector can respond differently under reflection. If a network treats all channels as ordinary scalars, it must relearn transformation behaviour that physics already specifies—or may be unable to express the correct behaviour cleanly.

Representation type is therefore part of meaning. Saying “feature channel 12 encodes direction” is not enough. How does that channel transform? If it is a 2D vector, a ninety-degree rotation should multiply it by the appropriate rotation matrix. If it is an invariant scalar, the same transformation should leave it unchanged. The representation can be tested mechanically against these laws.

This is one reason equivariant models can be more data-efficient. They do not have to rediscover every rotated version of a law independently. The architecture shares information across transformed configurations. But the benefit appears only when the assumed group action matches the task and the implementation preserves it closely enough.

5. Invariance is a special case, not the universal goal

Invariance is often introduced first because classification makes it intuitive. A cat remains a cat if the image is translated slightly. But destroying all transformation information early can make other tasks impossible. If a robot must reach toward the cat, location is not nuisance. If a molecule must predict a force vector, orientation cannot simply be erased. If a map must support navigation, coordinates matter.

A common architecture therefore carries equivariant features through several layers and creates invariance only when the receiver actually requires it. Pooling or integration over the group can produce invariant summaries. Before that point, the model may preserve pose, direction or local frames in structured features.

The design question is similar to the representation-budget problem in disentanglement: what information is nuisance for this receiver, and what is state? Invariance discards transformation information by design. Equivariance keeps it organised.

6. A two-by-two matrix teaches the whole constraint

Consider a linear layer f(x)=Wx acting on 2D vectors. Suppose we require rotational equivariance for a rotation matrix R. Then

W R x = R W x  for every x
therefore W R = R W.

The weight matrix must commute with the rotations in the group. If we require this for every planar rotation, the allowed linear maps are severely constrained: they behave like scaled rotations generated by the commutant of the representation, and if we additionally demand reflection symmetry the possibilities narrow further. The exact algebra depends on the representation type, but the lesson is general: equivariance turns semantic transformation rules into constraints on parameters.

This is why equivariant architecture is not simply “more augmentation”. The hypothesis class itself changes. Some functions become impossible. The benefit is stronger generalisation under the assumed symmetry. The cost is bias: if the true task violates the symmetry, the model may be unable to represent the needed exception.

7. Nonlinearities can break geometry

Suppose a feature transforms as a vector. Applying an ordinary ReLU independently to its Cartesian components is generally not rotation equivariant. Rotate the vector first and threshold each component; the result differs from thresholding first and then rotating. A harmless-looking activation has changed the transformation law.

Symmetry-aware architectures therefore design nonlinear operations around feature type. Scalar channels can often use ordinary pointwise nonlinearities. Vector or higher-order features may use norms, gated nonlinearities, tensor products or representation-specific operations that preserve the group action.

This is a practical warning: equivariance is a property of the whole pipeline, not a sticker attached to one layer. Normalisation, activation, interpolation, padding, discretisation, coordinate preprocessing and output postprocessing can all violate a theorem that was correct for the core operator.

8. Translation equivariance and the boundary that changes the theorem

Ordinary convolution is translation equivariant on an ideal infinite grid or under compatible boundary conditions: translate the input and the feature map translates. In finite images, padding changes the story. Zero padding introduces a special value at the boundary. A translated object that approaches the edge interacts with padding differently from the same object in the centre.

Striding and pooling can also reduce exact translation equivariance, particularly for shifts smaller than the stride. Resampling introduces aliasing. A theoretical claim about convolution should therefore be separated from a claim about an end-to-end image system.

A strong evaluation includes boundary cases and subpixel or sub-stride shifts, not only transformations that line up perfectly with the discretisation. Exact symmetry in continuous mathematics can become approximate symmetry in digital implementation.

9. Averaging over transformations: a proof and a price

One generic way to make a function invariant over a finite group is to average its output over transformed inputs:

F(x) = (1/|G|) Σg h(T_g x).

Under a group transformation k, the set of transformed inputs is merely permuted, so the sum is unchanged. The construction gives invariance by design.

The price is equally clear. Computation grows with the number of transformations, and the averaging discards information about which transformation occurred. For continuous groups such as all rotations, exact averaging requires integration or an approximation. If orientation later matters, early averaging may destroy useful state.

This small proof captures a larger pattern: invariance is often created by quotienting out a symmetry, while equivariance retains the orbit structure. The right choice depends on whether the receiver needs the transformed state or only an identity independent of that state.

10. Steerable features preserve orientation as organised information

Steerable filters and feature fields are designed so that transformed responses can be predicted from known transformation rules. Instead of learning an unrelated detector for every angle, the model represents features in bases that transform according to group representations.

General E(2)-Equivariant Steerable CNNs provide a systematic treatment of planar Euclidean symmetries using representation theory. The practical advantage is parameter sharing across orientations with transformation laws guaranteed by construction under the chosen discretisation and representation types.

The conceptual advantage is deeper. Orientation is not thrown away; it is encoded in how feature components transform. The representation carries an internal geometry instead of reducing the task to “same or different”.

11. Build an equivariant graph update from distances and directions

Graphs and point clouds do not live on a fixed image grid. Nodes have features and, in geometric problems, coordinates. An E(n)-equivariant graph network can build messages from quantities such as squared distances ||x_i-x_j||², which are invariant to global rotations, reflections and translations, while coordinate updates use relative direction vectors x_i-x_j, which transform equivariantly.

A schematic update is

m_ij = φ_e(h_i, h_j, ||x_i - x_j||²)
x_i' = x_i + Σ_j (x_i - x_j) φ_x(m_ij)
h_i' = φ_h(h_i, Σ_j m_ij)

If the scalar functions are constructed appropriately, a rigid transformation of all coordinates produces the corresponding transformed coordinate updates while invariant scalar features remain unchanged. E(n) Equivariant Graph Neural Networks popularised this economical design.

Notice what the architecture does not need to learn: the fact that Euclidean distance is unchanged by rotation, or that a relative vector rotates with the coordinate frame. These are placed in the representation contract.

12. Attention is geometric only when its ingredients are geometric

Attention is not automatically equivariant. If queries, keys, values and positional information transform inconsistently, attention can violate the desired symmetry. Geometry-aware attention must define which quantities are invariant scalars, which carry orientation, and how pairwise relations transform.

SE(3)-Transformers combine attention with equivariant feature types for 3D roto-translation symmetry. Tensor Field Networks, described in the foundational 3D point-cloud paper, use spherical harmonics and tensor products to construct rotation-equivariant features.

The broader lesson applies to transformer-style models in any domain: an attention mechanism does not remove the need to define transformation semantics. Positional encodings are part of the representation. If they break the claimed group action, the network is not saved by the word “attention”.

13. From invariant energy to equivariant force

Physics offers a beautiful example. Suppose a model predicts a scalar potential energy E(x) that is invariant to global translation and rotation of atomic coordinates. Forces can be derived as the negative gradient F_i = -∂E/∂x_i. Under appropriate differentiability and symmetry, the force field transforms equivariantly: rotate the molecule and the forces rotate.

This is more than architectural elegance. It connects representation symmetry to a physical conservation structure. Models such as NequIP, described in Nature Communications, exploit E(3)-equivariant representations for interatomic potentials and demonstrate strong data efficiency in benchmark regimes.

But “energy invariant, force equivariant” is not the whole physical validation. A learned force field may still violate other constraints, extrapolate poorly, or produce unstable trajectories. Symmetry is one necessary structure among several. Conservation tests, force errors, energy errors and out-of-distribution behaviour remain separate evidence.

14. Reflection, chirality and the information a distance matrix cannot contain

Pairwise distances are invariant under rotations, translations and reflections. That makes them powerful descriptors when mirror images are physically equivalent. But a set of distances alone cannot distinguish a chiral configuration from its mirror image. The two have identical distance matrices.

For molecules, handedness can matter profoundly. Enantiomers can interact differently with biological systems. If the task requires chirality, imposing full reflection invariance or using only reflection-invariant descriptors can erase essential information.

The repair is not “never use invariant distances”. It is to state the symmetry group accurately. Perhaps the task should be SE(3)-equivariant—rotations and translations—without identifying reflections. Perhaps pseudoscalar or oriented geometric features are required. The transformation contract must follow the physical semantics.

15. A symmetric input cannot supply a unique direction by itself

Suppose the input is a perfectly rotationally symmetric circle with no mark and the model is asked to predict one distinguished direction. There is no information in the input that selects north over east. An exactly rotation-equivariant deterministic model cannot break the symmetry arbitrarily while remaining consistent.

This is not a model failure. It is an information boundary. To select a direction, the system needs an asymmetry: a marker, an external coordinate frame, stochastic symmetry breaking, a task convention or additional context. The same principle appears throughout representation learning. If the data and assumptions leave several states equivalent, the network cannot justify a unique one without importing information from somewhere else.

This connects equivariance to identifiability. Symmetry explains not only what a model should preserve but also which distinctions the evidence cannot make.

16. Canonicalisation: turn the object once, then remember how it was turned

Another route to handling symmetry is canonicalisation. Estimate a canonical pose, transform the input into that pose, process it with an ordinary model, then transform outputs back as required. Conceptually this separates “where is the object oriented?” from “what is the object?”

Canonicalisation can be computationally attractive but has a hard failure mode: symmetric objects may not have a unique canonical orientation. A sphere has infinitely many equivalent orientations. A square may have four. Near a symmetry boundary, small input noise can cause a canonical frame to jump discontinuously between equivalent choices.

Recent work such as Adaptive Canonicalization with Application to Invariant Anisotropic Geometric Networks continues to explore when and how canonical frames can be learned or adapted. The evaluation should inspect continuity, ambiguity and the stability of the chosen frame under perturbations.

17. Local frames and gauge equivariance

On a curved surface there may be no single global orientation frame that works everywhere without singularities or distortion. A vector in the tangent plane is represented relative to a local basis. Change the local basis and the coordinate components change even though the physical vector has not.

This is the setting for gauge equivariance. The model should produce consistent physical results independent of arbitrary local choices of basis. Gauge Equivariant Convolutional Networks and the Icosahedral CNN develop this idea for signals on manifolds such as the sphere.

The conceptual lesson reaches beyond differential geometry. A representation often contains coordinates chosen by the modeller. If the answer should not depend on an arbitrary convention, the architecture should either remove the convention or transform correctly when the convention changes.

18. Gravity does not rotate merely because an image does

Image augmentation can suggest a symmetry that the world does not possess. Rotate a photograph of a room ninety degrees. Pixel geometry has rotated, but gravity in the original world did not. A cup that appears sideways in the augmented image is not a physically plausible new scene produced by rotating the entire universe relative to gravity; it is a camera/image transformation.

The distinction matters when labels depend on world orientation. “Up”, “standing”, “falling”, “floor” and “gravity-aligned” properties are not generally invariant to arbitrary image rotation. If we impose rotation invariance because object category is rotation-robust, a multi-task model may erase cues needed for physical reasoning.

Always ask what transformed: the object, the sensor, the coordinate frame, or the world? Equivariance is only meaningful relative to that answer.

19. A symmetric law can live inside an asymmetric dataset

The underlying physics may be rotation symmetric while the dataset is not. Cameras may favour one orientation. Molecules in a training corpus may occupy biased conformations. Satellite images may have north consistently at the top. Human poses are shaped by gravity. An equivariant model encodes a law-level prior even when the empirical distribution breaks the symmetry.

This can be beneficial because it distinguishes mechanism from sampling bias. But it can also remove predictive shortcuts that happen to be useful within the observed distribution. Whether that is desirable depends on the deployment. If the model must generalise across arbitrary orientations, the symmetry prior may reduce dependence on sampling bias. If orientation itself is informative, full invariance can be harmful.

The important comparison is between symmetry of the governing relation and symmetry of the observed distribution. They are not the same object.

20. Architecture, augmentation and approximate symmetry are different interventions

There are at least three ways to exploit a symmetry. First, hard-code exact equivariance in the architecture. Second, use data augmentation and let an ordinary architecture learn approximate transformation behaviour. Third, use a hybrid model that permits controlled symmetry breaking or learns which subgroup is relevant.

Exact equivariance is strongest when the symmetry is reliable and the numerical implementation respects it. Augmentation is flexible when the symmetry is approximate or task-dependent, but it requires data and optimisation to learn the relation. Hybrid approaches can preserve useful structure while allowing systematic deviations.

Research such as To Augment or Not to Augment? Diagnosing Distributional Symmetry Breaking and 2026 work on Any-Subgroup Equivariant Networks via Symmetry Breaking reflects this frontier. The question has moved beyond “is symmetry good?” toward “which symmetry, at which scale, and how should deviations be represented?”

21. Design a test that can actually break the claim

An equivariance test should apply transformations independently of the training examples and measure the discrepancy between “transform then model” and “model then transform”. For a feature map f, define an error such as

ε(g,x) = || f(T_g x) − S_g f(x) || / (||S_g f(x)|| + δ).

Evaluate this across transformations, inputs, magnitudes and boundary conditions. Report maximum as well as mean error. A tiny average can hide catastrophic failure near specific rotations or spatial boundaries.

Then evaluate task performance under transformed and symmetry-breaking data. Exact equivariance of a zero-output model is useless. A model must be both equivariant and informative. Conversely, high task accuracy does not prove equivariance if the test distribution never challenges the symmetry relation.

22. The pipeline can break the theorem before the network sees the data

Preprocessing is part of the model’s effective transformation law. Cropping can remove different content after rotation. Interpolation can introduce orientation-dependent blur. Voxelisation can privilege grid axes. Neighbour selection can change discontinuously. Sorting point clouds can invent an order. Data normalisation can use frame-specific statistics.

The output pipeline can also break symmetry. Non-maximum suppression, thresholding, coordinate rounding or an ordinary MLP head can destroy a property preserved by the backbone. End-to-end evaluation must therefore include every deterministic and learned stage that contributes to the final answer.

A theorem about one operator is not automatically a theorem about the deployed system.

23. Exact equivariance, numerical tolerance and the zero-predictor trap

Floating-point arithmetic, discretisation and finite grids mean practical systems often satisfy equivariance only within tolerance. Report the tolerance. Compare float32 and float64 where relevant. Test repeated transformation composition. If four quarter-turns should return to the original state, measure the accumulated error.

Also guard against trivial solutions. A function that always returns zero is perfectly equivariant for many vector representations. Equivariance error alone would call it excellent. The representation must retain task information. Pair symmetry metrics with predictive, reconstructive or physical metrics.

This mirrors representation collapse in self-supervised learning: satisfying one structural constraint can be trivial if the system throws away the content the constraint was meant to organise.

24. Executable laboratory: five tests worth running before training

2D rotation matrix: verify composition and inverse numerically.
Linear commutation test: construct matrices that do and do not commute with a chosen group action.
Nonlinearity break: show componentwise ReLU violating vector equivariance.
Distance invariance: rotate and translate point sets and verify pairwise distances are unchanged, then reflect a chiral configuration and show the same distances fail to encode handedness.
Energy-to-force test: define an invariant synthetic energy from pairwise distances, differentiate it, and verify force vectors rotate with the coordinates.

These calculations are deliberately small. Their purpose is to make the transformation law inspectable. Neural benchmarks should come after algebraic sanity checks, not replace them.

25. Diagnostic clinic for geometric failures

Wrong group: the model treats reflections as equivalent although chirality matters.
Wrong output type: a vector is trained as an invariant scalar.
Boundary leak: padding breaks translation behaviour.
Activation leak: a pointwise nonlinearity violates feature transformation rules.
Discretisation bias: voxel or grid orientation creates artificial anisotropy.
Canonical-frame instability: symmetric inputs cause frame jumps.
Gauge inconsistency: results depend on arbitrary local basis choices.
Symmetry overreach: the world contains systematic symmetry breaking the architecture cannot express.
Trivial equivariance: outputs collapse while equivariance error remains excellent.
Metric mismatch: average error hides worst-case failures.
Pipeline break: preprocessing or postprocessing violates the backbone guarantee.
Coordinate confusion: sensor rotation is mistaken for physical-world rotation.

26. What recent research changes

The frontier increasingly treats symmetry as discoverable, conditional or partially broken rather than a single fixed group supplied by the modeller. AtlasD: Automatic Local Symmetry Discovery investigates local symmetry discovery. Work on symmetry breaking allows networks to exploit a larger group’s structure while representing subgroups or asymmetric tasks. Recent research on reducing unintended symmetry increase warns that an architecture can become invariant to more transformations than the task justifies.

These developments do not make classical group-equivariant networks obsolete. They clarify their boundary. Exact symmetry is a powerful inductive bias when the domain supplies it. When the symmetry is local, approximate, data-dependent or broken by context, the representation needs a way to express that fact rather than forcing every example through one global rule.

The scientific responsibility is the same as elsewhere in the representation library: state what is assumed, what is learned, what is exact, what is approximate and which tests could show the claim failing.

27. Connection to sets, objects, causality and tokenisation

Permutation equivariance is the symmetry of unordered sets. Reorder input elements and the corresponding per-element outputs should reorder rather than change semantically. This connects directly to Set and Permutation Representation.

Object-centric representation introduces entities whose positions and poses transform while identities persist. Equivariance supplies a transformation law for the object attributes. Causal representation asks which variables survive interventions; equivariance asks how those variables should transform under known symmetry operations. Tokenisation determines the units on which any of these laws operate.

The layers should remain distinct. A point cloud can be tokenised into points, represented as an unordered set, processed with permutation-equivariant message passing and given rotation-equivariant geometric features. Each design decision solves a different problem.

28. Application brief: molecular modelling

Receiver job: predict energies and forces for atomic configurations. Desired symmetry: translation and rotation should not change total energy; forces should rotate. Reflections require care if chirality-sensitive properties are modelled. Permuting identical atoms should not change physical predictions.

Tests: rigidly transform held-out configurations, measure energy invariance and force equivariance, compare force consistency with energy gradients, inspect extrapolation to new conformations and chemical environments, and verify permutation behaviour. Symmetry metrics do not replace chemical validation.

29. Application brief: robotics and embodied perception

Receiver job: estimate object state and actions in changing camera and world frames. Desired symmetry: object identity may be invariant to camera pose, while position, orientation and action vectors must transform between frames. Gravity defines a privileged world direction, so arbitrary 3D rotation invariance may be inappropriate.

Tests: move the camera without moving objects, move objects without moving the camera, change robot base pose, check frame conversions explicitly, and evaluate whether predicted actions remain correct in the physical frame. The highest-level safety rule is that representation symmetry does not grant authority to execute an action; control constraints and confirmation remain separate layers.

30. Application brief: medical imaging

Receiver job: classify or segment anatomical and pathological structures. Desired symmetry: small translations and some orientation changes may be nuisance, while left-right laterality can be clinically meaningful. Acquisition protocol, patient positioning and anatomy create systematic asymmetries.

Tests: distinguish scanner-coordinate transformations from physically meaningful laterality, measure performance across acquisition sites, and ensure that imposed symmetry does not erase evidence. In high-stakes domains, the transformation contract should be reviewed by domain experts rather than inferred from computer-vision convention.

31. Worked problem: construct, break and repair an equivariant layer

Let an input vector v rotate by Rv. Define f(v)=a v for scalar a. Then f(Rv)=aRv=R(av)=Rf(v); the map is equivariant. Now define f(v)=(ReLU(v₁),ReLU(v₂)). Choose v=(-1,1) and a ninety-degree rotation. Evaluating the two sides shows they generally disagree.

Repair the nonlinearity by using an invariant norm to gate the vector: f(v)=σ(||v||)v. Because ||Rv||=||v||,

f(Rv)=σ(||Rv||)Rv = σ(||v||)Rv = Rf(v).

The exercise illustrates a general method: identify the representation type, construct invariant scalar quantities, then use them to modulate equivariant quantities without breaking their transformation law.

32. How to read an equivariance paper

What is the input domain?
What group acts on it?
Is the claimed property invariance or equivariance?
What representation acts on each feature type?
Is the property exact, approximate or empirical?
Which layer or whole pipeline satisfies it?
What boundary conditions or discretisation assumptions apply?
Does the dataset obey the same symmetry as the governing law?
Can the architecture express systematic symmetry breaking?
Which task metrics improve, and on which distributions?
Does the model outperform matched-capacity non-equivariant baselines?
What computation or memory cost is added?
What transformations were withheld for evaluation?
What would falsify the symmetry claim?

This checklist prevents the word “equivariant” from functioning as a quality label. Equivariance is a precise relationship. A model can satisfy it and still be inaccurate, slow or wrong for the domain.

33. Teaching symmetry as meaning attached to change

Start with physical objects. Rotate an arrow, a ball and a letter. Ask what stayed the same and what changed. Then rotate a coordinate axis without moving the object and ask the question again. Learners quickly see that “rotation” can describe different operations.

Next classify quantities: scalar mass, vector velocity, orientation, class label, left-right handedness. Give each a transformation rule. Only then introduce groups and the equation f(T_gx)=S_gf(x). The mathematics becomes a compact way of preserving distinctions already understood.

Use counterexamples early. Show zero-output equivariance. Show boundary padding. Show a chiral mirror pair with identical distances. Show a componentwise nonlinearity breaking vector behaviour. Symmetry becomes a habit of checking semantics rather than a theorem memorised in isolation.

34. The transformation is part of the meaning

The arrow returns us to the central principle. Its identity did not change when the board rotated. Its direction did. A good representation preserves both facts: sameness where sameness is meaningful, change where change is meaningful.

Equivariance is powerful because it places known structure inside the model. It can reduce sample complexity, improve generalisation and make geometric quantities easier to reason about. But every imposed symmetry is also an assumption. Reflections may destroy chirality. Rotation may conflict with gravity. Translation can be broken by boundaries. Local coordinate frames can be arbitrary. Real datasets can contain lawful asymmetry.

The mature question is therefore not “should the model be invariant?” It is “what is the transformation contract for this quantity, under this operation, in this world?”

A representation respects a symmetry when it does not merely ignore change, but carries the right change forward without losing the meaning attached to it.

Advanced Technical Expansion | Representation Theory, Broken Symmetry and Falsifiable Equivariance

The earlier sections establish the practical transformation contract: name the input action, name the feature type, and test whether the output transforms as promised. This expansion pushes that contract into the mathematical machinery that makes modern equivariant networks work. The aim is not to turn a learning guide into a pure representation-theory textbook. It is to show enough of the machinery that a reader can distinguish a genuine equivariance guarantee from a convenient augmentation story, understand where a guarantee can fail in implementation, and design experiments that reveal whether a symmetry prior helps the receiver job.

A. A group action is the bridge between an abstract symmetry and the actual data

A group by itself is an abstract collection of transformations with composition rules. To become relevant to a dataset, it needs an action. An action tells us how each group element changes an object. The rotation group can act on coordinates, images, spherical functions, point clouds, vector fields or internal feature channels, and the action is different in each case.

Write an input action as T_g and an output action as S_g. Equivariance is the intertwining relation f T_g = S_g f. This equation is stronger than saying the model has “seen rotations”. It states that the model commutes with the group action. If the action is incorrectly specified, the guarantee is about the wrong transformation.

A camera rotation, an object rotation and a change of coordinate basis can all produce rotated numbers while representing different physical operations. The group element may be numerically identical, yet its semantic action differs. A rigorous system stores both the mathematical action and the interpretation of the action.

B. Orbits tell us which states a symmetry declares equivalent

Choose an object x. Its orbit is the set {T_g x : g∈G}. If the final task is invariant to G, the model is being asked to assign the same task-relevant answer across the orbit. In image classification under translation, all translated copies of one object can belong to one orbit. In molecular energy prediction, all global rotations and translations of the same atomic configuration can belong to one physical orbit.

Orbit language makes invariance concrete. It also exposes overreach. If mirror images of a chiral molecule lie in the same orbit under the group we impose, the architecture is declaring them equivalent before seeing evidence. If the physical receiver distinguishes enantiomers, the group is too large.

A useful audit samples several orbit elements and several non-orbit near-neighbours. The model should preserve the required relation inside an orbit without collapsing genuinely distinct states outside it. This is the geometric version of testing both invariance and discrimination.

C. Stabiliser subgroups explain why canonical poses become ambiguous

The stabiliser of an object is the subgroup of transformations that leave it unchanged. A perfectly round unmarked disc is fixed by every planar rotation. A square is fixed by quarter-turn rotations. A generic asymmetric shape may have only the identity in its stabiliser.

Canonicalisation becomes difficult when stabilisers are nontrivial. If several orientations are physically indistinguishable, there is no unique canonical pose selected by the object alone. A learned canonicaliser can jump between equivalent orientations as noise changes. The discontinuity is not merely optimisation instability; it can follow from the geometry of the quotient space.

When canonicalisation is used, record the stabiliser structure of important object classes. Evaluate frame consistency under perturbations. For symmetric objects, consider returning an equivalence class of poses rather than one arbitrary pose, or maintain equivariant features that do not require choosing a unique frame.

D. Representations turn group elements into linear transformations on features

In representation theory, a representation ρ(g) assigns a linear transformation to each group element while respecting composition: ρ(g₁g₂)=ρ(g₁)ρ(g₂). Scalars use the trivial representation under rotation. Planar vectors use the usual 2×2 rotation matrices. Higher-order tensors use larger structured representations.

This gives feature channels a semantic type. A block of channels need not be interpreted one coordinate at a time; the whole block may form one irreducible or reducible representation. Under a rotation, coordinates mix according to ρ(g). The mixing is not entanglement in the undesirable sense. It is the correct transformation law.

This is an important correction to naïve interpretability. If a 3D vector feature is represented by three numbers, no one coordinate should be expected to mean “the vector” independently. The meaning lives in the transformation of the block.

E. Irreducible representations are the elementary transformation types

An irreducible representation cannot be decomposed into smaller invariant subspaces under the group action. General feature spaces can often be decomposed into sums of irreducible types. For rotations in three dimensions, spherical-harmonic degrees provide a familiar hierarchy: scalar-like degree zero, vector-like degree one, and higher angular structures beyond that.

Why does this matter to neural networks? Linear equivariant maps become block-structured by representation type. Some feature types can communicate freely; others require tensor products or learned couplings that respect the group’s algebra. The architecture stops treating every channel as interchangeable.

A diagnostic model card should list which representation types are present at each layer. “128 channels” is incomplete if 32 are scalars, 16 are vectors and the rest form higher-order features. Capacity depends on type as well as count.

F. Schur-style reasoning explains why equivariant linear maps are constrained

The earlier commutation example WR=RW is the small case of a general principle. Intertwining operators between irreducible representation spaces are highly constrained. For many groups and fields, Schur’s lemma states that an intertwiner between inequivalent irreducible representations must vanish, while an intertwiner from an irreducible representation to itself has a restricted form.

Neural architectures exploit this by parameterising only admissible maps. The parameter reduction is not arbitrary pruning. It follows from symmetry. A model can therefore use fewer free parameters while covering all functions compatible with the chosen representation structure.

This also creates a test. Construct a random linear layer, measure its equivariance residual, then project or constrain it into the commutant allowed by the group. The residual should drop to numerical tolerance. If it does not, the implementation of the action or projection is wrong.

G. The regular representation shows why group convolution carries orientation channels

For a finite group, the regular representation acts on functions defined over group elements by shifting their argument. A feature map indexed by position and group element can therefore carry not only “where” a pattern occurs but “under which transformed orientation” it is detected. Group convolution shares filters across those orientations according to the group law.

For a square-image rotation group, one can imagine four orientation channels corresponding to quarter turns. Rotate the input by ninety degrees and the orientation-channel index cycles. The feature tensor changes, but in a predictable permutation. Pooling over the orientation index later can create rotation invariance if the receiver wants it.

This gives a concrete interpretation of equivariance as organised bookkeeping. The model does not forget orientation; it moves orientation information through a known channel structure.

H. Haar measure is the continuous version of averaging over a group

The finite-group average |G|^{-1}Σ_g generalises to integration over a continuous compact group using an invariant Haar measure. An invariant summary can be written schematically as F(x)=∫_G h(T_gx)dμ(g). The measure is chosen so that shifting the integration variable by a group element does not change the integral.

In computation, the integral is approximated: discrete orientation samples, quadrature, harmonic coefficients or Monte Carlo estimates. The approximation introduces a second symmetry error beyond the model’s mathematical design. A paper can be exactly equivariant in continuous theory and only approximately equivariant after discretisation.

Report group-sampling resolution and test transformations between sampled orientations. If the model works only at the exact angles used in quadrature, its effective symmetry is smaller than the continuous claim suggests.

I. Fourier analysis turns translation into phase and convolution into multiplication

Translation symmetry is deeply connected to Fourier analysis. A spatial translation multiplies a Fourier component by a phase factor. Convolution becomes multiplication in frequency space. This is why complex exponentials are natural irreducible components for translations on periodic domains.

The perspective explains both power and failure. Frequency truncation can reduce computation while discarding fine spatial information. Discrete sampling can alias high frequencies into lower ones, breaking the intended transformation relation. Padding changes the domain and therefore the symmetry.

When a model claims translation equivariance, inspect the spectral consequences of downsampling. Anti-alias filtering before stride changes can preserve approximate shift behaviour better than naïve subsampling. The representation pipeline includes sampling theory.

J. Spherical harmonics provide rotation-aware basis functions

Functions on the sphere can be decomposed into spherical harmonics. Under 3D rotation, coefficients of a fixed angular degree mix through Wigner D matrices but do not mix arbitrarily across degrees. This structured transformation makes spherical harmonics a natural basis for 3D equivariant networks.

Degree zero behaves like a scalar. Degree one transforms like a vector representation. Higher degrees encode more complex angular patterns. Tensor Field Networks, SE(3)-Transformers and many molecular architectures use this machinery directly or indirectly.

Truncating at a maximum degree limits angular resolution and model expressivity. Increasing degree expands capacity and computational cost. “Maximum angular momentum” is therefore a representational budget analogous to channel count, with a clear geometric meaning.

K. Tensor products let equivariant features interact without losing their types

Linear equivariant layers alone cannot express every useful nonlinear interaction. Tensor products combine representation types: two vector-like features can produce scalar-like dot-product information, vector or pseudovector structure, and higher-order components depending on the group and decomposition.

Clebsch–Gordan coefficients provide the algebra for decomposing tensor products of rotation representations into irreducible components. In neural networks, learned weights can modulate admissible couplings while the fixed coefficients ensure the output transforms correctly.

This gives symmetry-aware networks a principled nonlinear language. Instead of applying arbitrary componentwise activations to vector coordinates, they form invariant scalars, gates and typed tensor interactions whose transformation laws are controlled.

L. Parity separates rotations from reflections

The rotation group SO(3) contains orientation-preserving rotations. O(3) also includes reflections. Features can have parity: some remain unchanged under inversion, others change sign. A polar vector such as displacement and an axial vector such as angular momentum respond differently under reflection.

If the receiver needs handedness, this distinction is essential. A model that uses only O(3)-invariant distances cannot distinguish mirror configurations. A model that includes orientation-sensitive triple products or parity-aware feature types can.

Audit the physical quantity before choosing SO(3), SE(3), E(3) or O(3). The larger group gives stronger parameter sharing and stronger blindness. Symmetry power and information loss rise together.

M. Semidirect products appear when transformations interact rather than simply add

Rigid motions combine rotations and translations. The order matters because rotating a translation vector changes its direction. This structure is captured by a semidirect product such as SE(3), not by treating rotation and translation as two independent symmetry lists.

This matters when implementing coordinate transforms. A pose can be represented by a rotation and a translation, but composition must follow the group law. Adding translation vectors while ignoring the current frame is a common robotics error. The representation should know which frame each vector belongs to.

Frame-labelled quantities make equivariance auditable. World-frame position, camera-frame direction and robot-base action are not interchangeable vectors even though each is three numbers.

N. Homogeneous spaces explain features that have a position without a full pose

Sometimes the state is not a group element itself but a coset or homogeneous space. A direction in 3D can be represented as a point on the sphere: rotating the direction matters, but rotation around the direction may be irrelevant. An unoriented axis identifies a direction with its negative. These spaces have their own stabilisers and transformation structure.

Trying to encode such states as full rotations introduces redundant degrees of freedom and can create artificial discontinuities. The correct state space can simplify learning and make identifiability clearer.

The design rule is to represent the quotient implied by the physical equivalence, not a larger coordinate system merely because it is convenient.

O. Gauge choices are local coordinate conveniences, not physical information

On a curved surface or bundle, local feature components require a choice of basis. Different bases describe the same physical object. A gauge transformation changes the local coordinate representation while leaving the underlying field unchanged. A gauge-equivariant network ensures predictions transform consistently under those local changes.

This is stronger than global rotation equivariance. Different locations may use different local frame rotations. Convolution-like operations must transport or compare features in a way that accounts for those frame differences.

A practical audit deliberately changes local frames without changing the underlying geometry. If predictions change beyond the expected gauge transformation, the model has learned a coordinate convention rather than the intended geometric object.

P. Approximate symmetry should be measured as a spectrum, not declared yes or no

Real datasets rarely obey perfect symmetry at every scale. Gravity breaks full 3D rotational symmetry for many embodied tasks. Sensor housing can privilege directions. Manufacturing tolerances produce near-symmetries. Biological systems can exhibit approximate bilateral symmetry with meaningful exceptions.

Measure symmetry defect as a function of transformation. For each g, evaluate both input-label consistency and model equivariance residual. Some transformations may be nearly exact; others systematically break the relation. Plotting this defect creates a symmetry profile.

A model can then use a hybrid prior: hard-code the reliable subgroup, learn residual corrections for the broken directions, or condition the symmetry on context variables. This is more faithful than choosing between “fully equivariant” and “not equivariant”.

Q. Symmetry breaking can be signal rather than noise

When a law is symmetric but a particular state is not, the broken symmetry can identify meaningful structure. A ferromagnet can choose one magnetisation direction even though the governing equations are rotationally symmetric. A crystal chooses a lattice orientation. A bilateral organism develops left-right asymmetries.

Machine learning systems may need to represent both the symmetric law and the state-specific order parameter that breaks it. Forcing the state to remain invariant would erase the very variable of interest.

Recent subgroup and local-symmetry research is important for exactly this reason: it gives models ways to exploit broad symmetry while preserving evidence about which symmetry is broken in a particular context.

R. Symmetry discovery is a model-selection problem with dangerous shortcuts

If the symmetry group is not known, the model can attempt to infer transformations under which the task behaves consistently. But many accidental dataset regularities can mimic symmetry. A training set containing only front-facing objects may make some rotations untestable rather than symmetric.

Discovery therefore requires interventions or held-out transformations. Propose a candidate symmetry from training data, then test it on deliberately transformed examples or new environments. Penalise complexity so the model does not invent unnecessarily large groups that erase useful distinctions.

Symmetry discovery should output uncertainty: which transformations are strongly supported, weakly supported or contradicted. A group label without an evidence profile can be more confident than the data justify.

S. Permutation equivariance is the symmetry of entity order, not entity identity

A set has no meaningful input order. If we permute the rows of a point cloud or the nodes in a graph representation, a per-entity output should permute in the same way and a global summary should remain invariant. This is permutation equivariance and invariance.

But reordering entities is not the same as swapping their physical identities in the world. If node features include identity or state, permuting the data structure carries those features with the node. The group action changes representation order, not the underlying assignment of attributes to entities.

This distinction prevents a common conceptual error: symmetry under arbitrary storage order does not imply physical exchangeability when entities have different roles.

T. Equivariant universal approximation claims always have a domain and representation class

There are universal-approximation results showing that sufficiently expressive symmetry-constrained architectures can approximate broad classes of continuous equivariant functions. The phrase “universal” is easy to overread. The theorem assumes a domain, group, topology, feature type, continuity class and architecture family.

A universal approximator can still be statistically inefficient, computationally impractical or badly optimised. It may require feature orders or widths far beyond a real system. Approximation existence is different from sample complexity and learnability.

When reading a universality theorem, write down the exact function class and the limiting resource that grows. The theorem establishes representational possibility, not a deployment guarantee.

U. Generalisation benefit depends on alignment between true symmetry and imposed symmetry

Equivariant models can reduce effective hypothesis-space size and therefore improve generalisation when the target function really obeys the symmetry. This is the intuition behind formal results such as the strict generalisation-benefit analysis in the literature.

If the symmetry is wrong, the same restriction becomes bias. A model that is invariant to a feature that predicts the target for legitimate reasons cannot recover that information. There is no free generalisation benefit independent of the task.

Empirically, compare matched-capacity models under data scarcity and abundant-data regimes. If the symmetry prior is useful, the advantage should often be largest when examples of transformed configurations are scarce. Then test systematic symmetry breaking to measure the cost of the prior when its assumptions fail.

V. Equivariance residuals should be decomposed by layer

End-to-end residual tells us whether the complete model meets the transformation contract, but it does not localise a defect. Measure equivariance after preprocessing, after each major block, at the latent representation, and at the output head. A sudden increase identifies the layer that violates the action.

Also decompose by feature type. Scalar residuals, vector residuals and higher-order residuals need different transformations before comparison. Normalise residuals carefully so low-magnitude features do not create unstable ratios.

Store worst-case examples. A mean residual of 10^-5 can coexist with an edge case that is orders of magnitude larger due to padding or a canonicalisation switch. Safety-sensitive receivers care about the tail.

W. Composition tests are stronger than single-transform tests

A model might pass isolated rotation tests but fail the group law numerically. Test composition: compare applying g₁ then g₂ with applying g₂g₁ once. Check inverses: transform and transform back. Check identity. These tests diagnose whether the implemented action is internally coherent.

For SE(3), vary rotation and translation jointly rather than in separate test suites. For permutations, compose random permutations. For gauge changes, apply different local frame transformations and their inverses.

Group-law testing is particularly useful for numerical libraries and coordinate conversions. A model can appear approximately equivariant under individual transformations while accumulating drift under repeated composition.

X. Conservation laws and equivariance are related but not interchangeable

Physics readers often connect symmetry with conservation through Noether’s theorem. The connection is profound and also easy to misuse. Noether relates continuous symmetries of an action or Lagrangian to conserved quantities under specific variational assumptions. A neural network being rotation equivariant does not by itself prove conservation of angular momentum in its learned dynamics.

A force field derived from an invariant potential has useful structural properties, but conservation also depends on time integration, the form of the dynamics and whether the learned potential matches the physical system. A direct force predictor can be equivariant and still be non-conservative if it is not the gradient of a scalar potential.

Test conservation separately. Compute loop integrals or compare mixed derivatives where appropriate. Run trajectories and monitor conserved quantities. Symmetry is evidence about structure, not a substitute for every physical law.

Y. A full benchmark should include symmetry-preserving and symmetry-breaking splits

One benchmark split should test generalisation along the assumed group: train on a subset of orientations or poses and evaluate held-out transformations. A second should test ordinary in-distribution performance. A third should introduce controlled symmetry breaking: gravity cues, boundaries, chiral labels, asymmetric sensors or context-dependent rules.

Compare an exact equivariant model, an augmented ordinary model, a hybrid residual-symmetry model and a capacity-matched baseline. Report sample efficiency, final task error, equivariance residual, compute, memory and robustness to broken symmetry.

This design reveals the full trade-off. Exact symmetry may dominate sparse-data group generalisation, while the hybrid model can recover when systematic exceptions matter. Augmentation may approach the same performance with enough data but at greater sample cost.

Z. A symmetry evidence card

Object: what is being transformed?
Group: which transformations are claimed?
Action: how do transformations act on raw input?
Feature types: scalars, vectors, tensors, parity or local-frame fields.
Output action: what should change and what should remain invariant?
Exactness: theorem, discrete approximation or learned empirical relation.
Boundary conditions: padding, domain limits, coordinate conventions.
Stabilisers: which states have ambiguous canonical poses?
Broken symmetry: known exceptions and context variables.
Residual tests: mean, quantiles and worst case.
Composition tests: identity, inverse and group-law checks.
Task utility: sample efficiency and downstream improvement.
Cost: parameter, memory and compute overhead.
Scope: transformations, populations and domains actually tested.

AA. Worked laboratory: square symmetry as a finite group

Use the dihedral group D4 of a square: four rotations and four reflections. Represent an image patch on a square grid and explicitly implement each transformation. Build an orientation-indexed feature vector whose channels correspond to group elements. Verify that transforming the input permutes channels according to left multiplication.

Then average over the group and prove numerically that the pooled scalar is invariant. Remove the reflection elements and observe how the group changes from D4 to the cyclic rotation subgroup. Construct a chiral pattern whose mirror should remain distinct and show why D4 pooling erases it while rotation-only pooling does not.

This finite example makes group choice, representation action, invariance and information loss inspectable without a deep network.

AB. Worked laboratory: project a matrix onto an equivariant subspace

Take a random matrix W acting on feature vectors and a finite group representation ρ(g). Construct the Reynolds-style average

W_eq = (1/|G|) Σ_g ρ_out(g)^(-1) W ρ_in(g).

Under the appropriate representation conventions, this averaging projects the arbitrary map toward the subspace of intertwiners. Verify numerically that W_eq ρ_in(k) ≈ ρ_out(k) W_eq for every group element k while the original random W does not.

The experiment turns abstract representation theory into a direct engineering test: symmetry can be imposed by averaging a function over transformed copies, but projection also removes components incompatible with the symmetry. Those removed components are useful only if the task legitimately breaks the symmetry.

AC. Worked laboratory: chirality from oriented volume

Choose four non-coplanar points. Pairwise distances are unchanged by reflection. Compute an oriented volume using a scalar triple product such as det[b-a, c-a, d-a]. Under a proper rotation, the sign is preserved. Under reflection, the sign flips.

This one scalar demonstrates information unavailable to a distance-only O(3)-invariant representation. If the receiver needs chirality, parity-sensitive structure must enter somewhere. If chirality is nuisance, the sign can be deliberately discarded.

The lesson is precise: invariants are lossy summaries chosen relative to a group. Every invariant should be accompanied by the question, “which distinctions did this quotient remove?”

AD. Worked laboratory: equivariant message passing and permutation order

Build a small point graph. Define messages from invariant squared distances and scalar node features. Sum messages over neighbours. Because summation is permutation invariant, reordering neighbours does not change the aggregated message. Update coordinates using weighted relative vectors. Rotate and translate the entire point cloud, then verify that scalar outputs remain invariant and coordinate outputs transform accordingly.

Next deliberately break one ingredient. Replace relative coordinates by absolute x-coordinates inside the scalar message. Rotate the cloud. The residual rises. Replace neighbour summation with an order-dependent concatenation. Permute nodes. The residual rises again.

This failure-by-construction method is excellent for teaching because every broken line of code has a clear semantic consequence.

AE. Worked laboratory: canonicalisation under near symmetry

Create rectangles whose aspect ratio approaches one. Define the canonical orientation by the principal axis of the point cloud. For a clearly elongated rectangle, the leading eigenvector is stable. As the shape approaches a square, the top eigenvalues become equal and the principal direction becomes ill-conditioned. Tiny noise can rotate the estimated frame dramatically.

Plot frame angle error against eigenvalue gap. The experiment shows that canonicalisation uncertainty can be predicted from symmetry: as the stabiliser grows, frame identifiability shrinks. A robust system should report pose ambiguity instead of forcing one confident orientation.

AF. Equivariance and causal representation solve different invariance questions

Equivariance says how a representation should transform under a specified action. Causal representation asks which variables and mechanisms retain meaning under interventions or environment changes. A rotation is often a known symmetry action; a causal intervention changes the generating mechanism or variable value.

The two can reinforce one another. A causal variable representing physical position should transform equivariantly under a coordinate rotation. A mechanism such as Newtonian force may obey rotational symmetry across environments. But a model can be perfectly rotation equivariant while relying on a confounded causal variable, and a causal representation can be correct without using a hard-coded group architecture.

Keep the contracts separate: symmetry tests transformation consistency; causal tests intervention stability.

AG. Equivariance and object-centric representation meet at persistent entity state

An object slot can carry identity as an invariant attribute and pose as an equivariant attribute. If a scene is globally rotated, the object identities remain, positions and orientations transform, and relations such as distance can remain invariant while directional relations rotate.

This typed slot representation is more powerful than one undifferentiated vector because it tells a world model which parts of state should move predictably. Dynamics can then be learned on object states without relearning the coordinate-system law for every object.

Evaluation should track slot identity through transformations, compare object-wise transformation residuals, and distinguish global scene actions from local object actions.

AH. Equivariance should survive model versioning if downstream interfaces depend on it

Downstream systems may rely on the transformation semantics of a representation. A robotics controller may assume a vector block rotates under a specified matrix. If a new model version changes coordinate convention or feature typing, the controller can silently receive incompatible state.

Version the representation contract. Store group, convention, handedness, axis order, units and reference frame. Run regression tests across model updates. A representation with mathematical semantics is an interface, not merely an embedding.

This is the final engineering consequence of equivariance: when change has a known meaning, that meaning must travel with the data.

AI. Final audit: when does an equivariance claim deserve release?

The physical or semantic transformation is defined.
The mathematical group and action are stated.
Feature types and their representations are documented.
Input, latent and output transformation laws are tested.
Identity, inverse and composition tests pass within declared tolerance.
Boundary and discretisation effects are measured.
Stabiliser-induced ambiguities are surfaced.
Reflection and parity assumptions match the receiver.
Known symmetry breaking is represented rather than hidden.
Task information does not collapse while residuals look good.
Matched baselines establish whether the prior improves sample efficiency or robustness.
Worst-case residuals and not only averages are inspected.
Preprocessing and postprocessing are inside the end-to-end test.
Versioned coordinate conventions are available to downstream users.
The public claim does not exceed the transformations and domains actually tested.

A strong equivariant model does not merely score well on rotated examples. It carries a contract about change. The contract tells us which states are equivalent, which features rotate or permute, which asymmetries matter and which coordinate choices are arbitrary. That contract can reduce sample complexity and improve transfer precisely because the model is forbidden to waste capacity relearning a law we already know. The same restriction becomes harmful when the law is wrong. Equivariance is therefore best understood as disciplined prior knowledge: powerful because it is explicit, and trustworthy only when its boundaries are tested as carefully as its successes.

Sources and research boundaries

Continue in the Representation & Cognitive Tools library

Return to the World Representation & Cognitive Tools canonical owner. Related advanced reading: Disentangled Representation Learning · Object-Centric Representation · 3D Tokenisation · Causal Representation Learning.

Equivariant Representation Learning | How Models Preserve Structure Through Rotation, Translation, Reflection and Symmetry