{ "schema": "auturnal.deep-research-review.v2", "slug": "equivariance-under-symmetry", "title": "Equivariance Is a Contract About Transformations", "category": "Representation & Architecture", "dek": "A symmetry constraint is useful only when the input transformation, output transformation and target behavior agree.", "abstract": "Equivariant learning builds transformation rules into a model rather than hoping that data augmentation will teach them. This review uses group-equivariant convolutions, steerable feature types and E(n)-equivariant graph networks as primary anchors, then develops finite linear examples that expose the contract behind the construction. A group average is verified to commute with quarter-turn rotations; a deliberately anisotropic target quantifies the cost of imposing an incorrect symmetry; a signed-volume example distinguishes rotation invariance from reflection invariance. The analysis separates equivariance, invariance, empirical robustness and conservation laws, and explains how coordinate preprocessing and feature types can invalidate otherwise correct layers. All numeric examples are small algebraic constructions, not reproduced neural-network benchmarks. The central design requirement is to specify a transformation law for the task itself before choosing which symmetries the model should enforce.", "english": "Equivariant learning builds transformation rules into a model rather than hoping that data augmentation will teach them. This review uses group-equivariant convolutions, steerable feature types and E(n)-equivariant graph networks as primary anchors, then develops finite linear examples that expose the contract behind the construction. A group average is verified to commute with quarter-turn rotations; a deliberately anisotropic target quantifies the cost of imposing an incorrect symmetry; a signed-volume example distinguishes rotation invariance from reflection invariance. The analysis separates equivariance, invariance, empirical robustness and conservation laws, and explains how coordinate preprocessing and feature types can invalidate otherwise correct layers. All numeric examples are small algebraic constructions, not reproduced neural-network benchmarks. The central design requirement is to specify a transformation law for the task itself before choosing which symmetries the model should enforce.", "refs": [ "dl-09-1", "dl-09-2", "dl-09-3" ], "sections": [ [ "1. Start with the transformation of the answer", "A rotation of an input need not leave its answer unchanged. The energy assigned to a physical configuration might be a scalar invariant, while a predicted displacement should rotate with the coordinates. A map from an image to an orientation label has still another output transformation. Calling all three requirements rotational robustness obscures their mathematical differences.\n\nEquivariance specifies a relation between an input action and an output action. Invariance is the special case in which the output action is trivial. Before choosing an equivariant architecture, the task designer must decide which action is appropriate and whether the data acquisition process respects it. Rotating an object while holding illumination fixed is not always equivalent to rotating the coordinate frame. A physically meaningful claim should distinguish these interventions. The remainder of this article treats symmetry as a testable contract about the complete pipeline, rather than a decorative label attached to one layer.", [] ], [ "2. Group convolutions provide a structural route", "Cohen and Welling introduce group-equivariant convolutional networks that extend convolutional weight sharing to specified transformations. Their source demonstrates architectures for discrete groups involving translations, rotations and reflections and evaluates them on visual benchmarks.\n\nThe key contribution used here is structural: transformation consistency can be built into the allowable operations. The source does not imply that every possible task should identify rotated or reflected inputs, nor that exact algebra survives arbitrary preprocessing. The finite quarter-turn averaging construction below is this article's own verification example. It illustrates a commutation constraint without reimplementing a published convolutional network or reproducing the source's measured sample-efficiency results. This separation keeps the mathematical claim precise and the empirical scope honest.", [ "dl-09-1" ] ], [ "3. Features have transformation types", "The steerable-CNN framework of Cohen and Welling treats features as objects with specified transformation behavior and constrains maps between those feature types. Its theoretical perspective explains why equivariant construction is more than repeating an ordinary filter at several orientations.\n\nFor this review, the practical implication is a type discipline. A scalar, a planar vector and an orientation-indexed feature array do not transform in the same way. Operations that mix them must respect their respective laws. The primary source supplies the feature-type viewpoint; the simple vector and signed-volume tests below are independently worked examples. The bibliography distinguishes the accessed preprint, which retains an under-review header, from the ICLR venue identified by its primary metadata.", [ "dl-09-2" ] ], [ "4. Coordinates and invariant messages", "Satorras, Hoogeboom and Welling introduce EGNNs with equivariance to Euclidean transformations and permutations. The architecture uses invariant quantities such as squared distances in message construction and updates coordinates through relative vectors weighted by scalar functions.\n\nThe distinction between scalars and vectors is essential: squared distances survive orthogonal transformations, while relative vectors transform with them. This permits an algebraic equivariance argument when the scalar functions depend only on suitable invariant inputs. It does not imply reflection-sensitive targets can be represented by a reflection-invariant scalar readout. Nor does equivariance alone certify conservation of energy, numerical stability of a rollout, or predictive accuracy. Those are separate properties requiring separate assumptions and evidence; they are not claimed by the examples in this article.", [ "dl-09-3" ] ], [ "5. A finite group average that can be checked exactly", "Let R denote a ninety-degree planar rotation and let W be an arbitrary two-by-two linear map. Average the four conjugates obtained by rotating into each group frame, applying W, and rotating back. The resulting matrix commutes with every element of the quarter-turn group. A change of summation index proves the claim because multiplying by a group element merely permutes the four terms.\n\nFor the anisotropic matrix with diagonal entries two and one, the average is one and a half times the identity. The averaging removes the preferred horizontal direction. The executable companion checks the commutation identity for every group element and several deterministic matrices. These are finite assertions about a discrete group, not tests of continuous-rotation equivariance for an arbitrary image network. In two dimensions, some linear maps that commute with quarter turns also commute with all rotations; that special fact should not be generalized to arbitrary representations or sampled image operations.", [] ], [ "6. A visual measure of a broken contract", "Panel 1 measures the Frobenius norm of the commutator between the anisotropic diagonal map and a planar rotation. The norm is the square root of two times the absolute sine of the angle. It vanishes at zero and one hundred eighty degrees but is largest at ninety degrees. An isotropic scalar multiple of the identity has zero commutator at every angle.\n\nThe example shows why testing only a convenient subset of transformations can be misleading. A map may pass a half-turn test while violating a quarter-turn contract. Conversely, a numerical implementation may have tiny interpolation errors even when the continuous operator is equivariant. Testing should therefore identify the exact group elements, their discretized implementation, the norm and the numerical tolerance. A curve over angles is a useful diagnostic but not a proof for all possible inputs. The analytic expression and its finite numerical checks make that distinction explicit here.", [] ], [ "7. The wrong symmetry introduces unavoidable bias", "Suppose the desired vector-valued target really is the anisotropic map with gains two and one in fixed world axes. Average it toward the isotropic group projection using a mixing coefficient lambda. For inputs uniformly distributed on the unit circle, the mean squared target error becomes one quarter of lambda squared. Meanwhile, its mean squared quarter-turn equivariance defect becomes the square of one minus lambda.\n\nPanel 2 plots these two explicitly labeled quantities. At full symmetry, the equivariance defect is zero but the target error is one quarter. At no averaging, the target is exact and the defect is one. This is a deliberate task mismatch, not a criticism of correctly specified equivariance. It proves that satisfying a transformation constraint is insufficient if the desired answer does not obey that constraint. Task validation must precede architectural enforcement; otherwise a precise implementation can enforce the wrong scientific assumption.", [] ], [ "8. Reflection invariance can erase handedness", "Take three displacement vectors from a common origin and form their oriented parallelepiped volume through a determinant. Proper rotations preserve its sign, whereas a reflection reverses it. The pairwise distances among the four points remain identical after reflection. A representation containing only those distances therefore cannot distinguish the two handed configurations for a target defined by this signed triple product.\n\nThe executable example reflects a concrete tetrahedral configuration, verifies identical distances, and checks opposite determinants. It is a mathematical statement about labeled coordinates and a chosen pseudoscalar target. It does not claim that every molecular prediction needs handedness or that a particular EGNN application fails. Instead, it identifies the question an application must answer: is the target invariant under all orthogonal transformations, only proper rotations, or something else? When the target transforms with a sign under reflection, that output type must be represented explicitly. More training data cannot repair information deliberately identified by an invariant representation.", [] ], [ "9. Nonlinear operations must respect the feature type", "For a vector that rotates continuously, applying an arbitrary coordinate-wise nonlinearity generally changes its direction in a frame-dependent way. For example, clipping negative coordinates treats the axes as privileged. A nonlinear map based on a vector's invariant norm times the vector can respect rotation behavior under suitable definitions at zero. These choices implement different mathematical contracts.\n\nThe same scrutiny applies to normalization, pooling and feature concatenation. Summing scalars is straightforward; pooling vectors into a scalar by their norm discards directional information. Concatenating features is valid only if the resulting output action is defined for each component. A network may contain individually familiar operations but fail globally because one operation assumes the wrong feature type. A verification suite should test intermediate layers as well as the final prediction so that a defect can be localized. Passing a final scalar test alone can conceal cancellation or information loss in an intermediate representation.", [] ], [ "10. Boundaries and preprocessing are part of the model", "Exact symmetry of a continuous operator may become approximate after rasterization, padding, truncation, neighbor selection or coordinate normalization. Rotating an image on a finite square grid generally requires interpolation except for special angles. Selecting a fixed number of nearest neighbors can change discontinuously at a tie, and deterministic tie-breaking can introduce an ordering convention.\n\nThese details do not make equivariant design useless. They define the level at which the claim is true and the error sources that should be reported. A practical evaluation should test transformed raw inputs through the entire preprocessing pipeline, not merely rotate already prepared internal tensors. It should also distinguish errors from floating-point arithmetic and discretization from errors due to an invalid transformation rule. If an application tolerates an approximate contract, its tolerance should be tied to decision quality or physical units rather than selected because it makes the test pass.", [] ], [ "11. Equivariance does not imply dynamics are correct", "A coordinate update based on scalar-weighted relative vectors can transform correctly under rotations and translations while producing inaccurate accelerations or an unstable numerical trajectory. Translation and rotation consistency are constraints on form; conservation and stability impose additional relationships across coordinates, forces, time steps and learned parameters.\n\nA proposed dynamics evaluation should therefore separate instantaneous equivariance error, one-step predictive error, conservation residuals where appropriate, and long-horizon behavior under a declared integrator. No one measure substitutes for the others. For a static recognition task, the analogous separation is between equivariance, calibration and classification risk. The accompanying code verifies only finite group identities, a distance-preserving reflection and the declared bias calculation. It does not run a physical simulator or train an EGNN. Explicitly separating these evidence levels prevents a valid symmetry proof from being promoted into a much broader claim about the learned system.", [] ], [ "12. A transformation contract that can travel", "A reusable model description should specify the input group action, the output representation, the transformation of auxiliary features, and the preprocessing conventions. It should include examples that ought to remain equivalent and counterexamples that ought to remain distinguishable. These declarations make symmetry useful to researchers integrating the model into a larger system.\n\nThe finite examples establish why both parts matter. Group averaging can enforce an exact relation; enforcing that relation can be harmful for an anisotropic target; and reflection invariance can identify configurations that a handed target separates. A rigorous application does not choose between theory and experiment. It uses algebra to identify what is guaranteed, finite checks to catch implementation mistakes, and held-out task evaluations to assess usefulness. The resulting claim is narrower than universal robustness but substantially stronger: under a stated transformation and numerical convention, the model preserves the behavior that the task actually requires.", [] ] ], "section_kinds": [ "author_analysis_or_verification_proposal", "source_based_review", "source_based_review", "source_based_review", "author_analysis_or_verification_proposal", "author_analysis_or_verification_proposal", "author_analysis_or_verification_proposal", "author_analysis_or_verification_proposal", "author_analysis_or_verification_proposal", "author_analysis_or_verification_proposal", "author_analysis_or_verification_proposal", "author_analysis_or_verification_proposal" ], "equations": [ { "tex": "f(\\rho_{\\rm in}(g)x)=\\rho_{\\rm out}(g)f(x),\\qquad g\\in G", "explanation": "Equivariance requires specified group actions on both input and output; invariant outputs use the trivial action." }, { "tex": "\\overline W=\\frac1{|G|}\\sum_{g\\in G}\\rho(g)^{-1}W\\rho(g),\\qquad \\overline W\\rho(h)=\\rho(h)\\overline W", "explanation": "Finite-group projection for a linear endomorphism using the same representation on input and output." }, { "tex": "x_i^{\\rm new}=x_i+\\sum_{j\\ne i}(x_i-x_j)\\,\\phi\\!\\left(\\|x_i-x_j\\|^2,h_i,h_j\\right)", "explanation": "Illustrative equivariant coordinate update when h values are invariant scalars and phi produces an invariant scalar; constants can be absorbed into phi." }, { "tex": "V(a,b,c)=\\det[a\\ b\\ c],\\qquad V(Qa,Qb,Qc)=\\det(Q)V(a,b,c)", "explanation": "V is the oriented parallelepiped volume (signed triple product), not tetrahedron volume. Orthogonal Q preserves squared distances; V changes sign when det(Q)=-1." } ], "table": { "headers": [ "Property", "Contract", "Additional claim not implied" ], "rows": [ [ "Invariance", "Output unchanged under a specified action", "Preservation of orientation or handedness" ], [ "Equivariance", "Output follows its prescribed transformation", "Task accuracy or resistance to unrelated corruption" ], [ "Group averaging", "Finite linear map commutes with the group", "Correctness of the symmetry for the target" ], [ "Euclidean coordinate update", "Coordinates transform consistently", "Energy conservation or stable long-term dynamics" ] ] }, "protocol": [ "Specify input and output actions, including reflection behavior and auxiliary features.", "Enumerate all elements of the finite quarter-turn test group and verify the commutator.", "Construct matched and deliberately mismatched targets to measure symmetry-induced bias.", "Test mirror configurations with identical distances and a signed target.", "Run transformations through preprocessing and intermediate layers, recording numerical tolerances.", "Evaluate task risk and, where relevant, conservation and stability independently of equivariance." ], "limitations": [ "Finite quarter-turn checks do not certify arbitrary continuous image transformations.", "The mismatch curve uses a chosen anisotropic linear target and uniform-circle inputs.", "Signed volume is a specific reflection-sensitive target, not a claim about every scientific application.", "No published equivariant network is trained or independently benchmarked in this companion." ], "evidence": [ { "source": "dl-09-1", "result": "Group-equivariant convolutions extend weight sharing to explicitly chosen transformation groups.", "boundary": "Usefulness depends on the task symmetry; this guide does not replicate the network experiments.", "locator": "Primary publisher abstract." }, { "source": "dl-09-2", "result": "Steerable features have structured transformation types that constrain admissible maps.", "boundary": "The accessed preprint version is distinguished from venue metadata.", "locator": "Abstract and introduction on steerable representations." }, { "source": "dl-09-3", "result": "EGNNs combine invariant message inputs with equivariant relative-coordinate updates.", "boundary": "Transformation consistency does not by itself imply conservation, stability or accuracy.", "locator": "Architecture equations and introductory symmetry discussion." } ], "research_questions": [ "How should approximate symmetry be selected when acquisition conditions break an otherwise physical invariance?", "Which intermediate-layer tests best detect feature-type violations before they affect task accuracy?", "Can symmetry constraints be relaxed adaptively while retaining interpretable error bounds?" ], "kind": "ai_assisted_research_guide", "author": "Codex (AI)", "original_language": "en", "created_date": "2026-10-22", "published_date": null, "publication_status": "local_draft_pending_owner_upload", "version": 1, "previous_version": null, "analytic": "equivariance-under-symmetry", "peer_reviewed": false, "journal_publication": false, "disclosure": { "authorship": "Written and technically self-reviewed by Codex (AI). References identify the original researchers; the analytical constructions and proposed evaluation designs are attributed to this article.", "figures": "Covers are conceptual ImageGen illustrations. Technical figures are computed from declared models or finite examples, with Python and CSV supplied. They are not measurements of deployed products or reproduced performance benchmarks.", "review": "This local draft has no independent peer-review or journal-acceptance record. Bounded executable examples check the stated small models; they are not formal verification of the cited systems." }, "review_scope": "A focused analytical review of three selected foundational primary sources, with original worked examples. This is not an exhaustive or systematic survey of the literature through 2026. Source access and version notes are recorded with the bibliography.", "revision_note": "Initial local English manuscript; no server upload or publication action." }