Temporal Translation Invariance

Updated 5 July 2026

Temporal Translation Invariance is the property whereby signals, functions, or operators remain unchanged or shift consistently under time translations.
It is implemented as either invariance or equivariance, with each formulation affecting model architecture and robustness in domains like video recognition and neural operator design.
Leveraging temporal translation symmetry improves adversarial robustness, enhances signal representation stability, and facilitates efficient dynamical system predictions.

Searching arXiv for relevant papers on temporal translation invariance across machine learning, physics, and dynamical systems. Temporal translation invariance denotes invariance under shifts of the time coordinate, but its precise meaning depends on the mathematical object being transformed. For a temporal signal or field, the basic action is a shift operator such as $T_\tau x(t)=x(t-\tau)$ or, for spatio-temporal data, $(S_\tau f)(x,t)=f(x,t-\tau)$ . An invariant map satisfies $f(T_\tau x)=f(x)$ , whereas an equivariant map satisfies $f(T_\tau x)=T_\tau f(x)$ . In stationary stochastic processes and dynamical prediction, equivariance is typically the physically appropriate symmetry; in representation learning, classification, or adversarial robustness, approximate invariance of outputs or features may be the operative goal; and in non-equilibrium and high-energy physics, temporal translation invariance can be deformed, generalized, or spontaneously broken rather than simply imposed (Ashman et al., 2024, Pellegrini, 18 Mar 2026).

1. Core definitions and domain-specific meanings

Temporal translation invariance is the requirement that a quantity or map remain unchanged under a shift of the time argument. In video recognition, a model $f$ is temporally translation-invariant if its output is approximately unchanged by a temporal shift,

$f(T_\tau(x)) \approx f(x),$

for $\tau$ in a range of interest. In this setting, real video models are not perfectly invariant, and different architectures rely on different discriminative temporal patterns (Wei et al., 2021).

For dynamical operators, the stronger and usually more relevant property is temporal translation equivariance rather than invariance. In the FitzHugh–Nagumo neural-operator benchmark, the shift action is

$(S_\tau f)(x,t)=f(x,t-\tau),$

and a neural operator $O$ is temporally translation equivariant if

$O[S_\tau I](x,t)=S_\tau O[I](x,t),$

whereas temporal translation invariance would mean $(S_\tau f)(x,t)=f(x,t-\tau)$ 0, which is explicitly described as not the desired property for dynamical systems with delayed responses (Pellegrini, 18 Mar 2026).

In Neural Processes, the symmetry is formulated at the level of the posterior predictive map $(S_\tau f)(x,t)=f(x,t-\tau)$ 1. Translation equivariance is written as

$(S_\tau f)(x,t)=f(x,t-\tau)$ 2

which matches the structure of stationary processes (Ashman et al., 2024).

In physics, temporal translation invariance is often identified with stationarity. In equilibrium or stationary dynamics,

$(S_\tau f)(x,t)=f(x,t-\tau)$ 3

whereas breaking the symmetry produces explicit dependence on both times. In Euclidean quantum dynamics, the same condition becomes frequency conservation,

$(S_\tau f)(x,t)=f(x,t-\tau)$ 4

so non-diagonal frequency structure signals time-translation symmetry breaking (Achitouv et al., 2024, Henkel, 23 Apr 2025).

Domain	Symmetry statement	Operational consequence
Video recognition	$(S_\tau f)(x,t)=f(x,t-\tau)$ 5	reduced sensitivity to temporal alignment
Neural operators	$(S_\tau f)(x,t)=f(x,t-\tau)$ 6	delayed dynamics shift with stimuli
Neural Processes	$(S_\tau f)(x,t)=f(x,t-\tau)$ 7	posterior predictions respect stationarity
Scattering transforms	$(S_\tau f)(x,t)=f(x,t-\tau)$ 8 as output scale grows	time-shift stable representations
Relativistic phase space	time translations generated by $(S_\tau f)(x,t)=f(x,t-\tau)$ 9, not necessarily $f(T_\tau x)=f(x)$ 0	deformed conservation and dynamics
Ageing systems	ordinary TTI broken, generalized generator retained	constrained non-stationary scaling

A persistent misconception is that “temporal translation invariance” always means unchanged output under time shifts. The literature instead separates true invariance from equivariance, and in several dynamical settings the latter is the correct notion (Ashman et al., 2024, Pellegrini, 18 Mar 2026).

2. Stationarity, posterior prediction, and operator learning

For stationary stochastic processes, temporal translation symmetry is not an ad hoc architectural preference but a structural property of the underlying law. In the translation-equivariant Transformer Neural Process framework, the paper establishes the equivalence that the ground-truth stochastic process $f(T_\tau x)=f(x)$ 1 is stationary and the likelihood $f(T_\tau x)=f(x)$ 2 is translation invariant if and only if the prediction map $f(T_\tau x)=f(x)$ 3 is translation equivariant. For Gaussian-process regression with stationary kernels, the posterior mean and covariance inherit this symmetry because the kernel depends only on differences, so simultaneous shifts of context inputs and query inputs leave the posterior structure unchanged (Ashman et al., 2024).

This perspective determines the architecture. Standard Transformers with absolute positional encodings break translation symmetry, whereas TE-TNPs enforce translation equivariance through attention weights that depend on relative differences $f(T_\tau x)=f(x)$ 4:

$f(T_\tau x)=f(x)$ 5

Because the encoder depends only on relative positions and permutation-invariant aggregation, the resulting posterior predictive map is translation equivariant by construction (Ashman et al., 2024).

Empirically, the effect is strongest under distribution shifts. On synthetic 1D regression, non-equivariant TNPs degrade as the shift $f(T_\tau x)=f(x)$ 6 increases, whereas TE models and ConvCNP remain stable. The paper reports that TE-PT-TNP-M32 is approximately $f(T_\tau x)=f(x)$ 7 for all $f(T_\tau x)=f(x)$ 8, while TNP worsens to $f(T_\tau x)=f(x)$ 9 at $f(T_\tau x)=T_\tau f(x)$ 0 and CNP degrades to $f(T_\tau x)=T_\tau f(x)$ 1. On ERA5 environmental data, TE-PT-TNP achieves $f(T_\tau x)=T_\tau f(x)$ 2 in Central Europe, $f(T_\tau x)=T_\tau f(x)$ 3 in Western Europe, and $f(T_\tau x)=T_\tau f(x)$ 4 in Northern Europe, outperforming GP, CNP, RCNP, and PT-TNP across all regions (Ashman et al., 2024).

A closely related issue appears in neural operators for PDE dynamics. For the FitzHugh–Nagumo model, the physical operator is exactly equivariant with respect to time shifts of an applied current pulse: shifting the stimulus simply shifts the solution in time. The benchmark exploits this by training on stimuli with fixed start time $f(T_\tau x)=T_\tau f(x)$ 5 ms and varying $f(T_\tau x)=T_\tau f(x)$ 6 and $f(T_\tau x)=T_\tau f(x)$ 7, then testing out-of-distribution on translated dynamics with $f(T_\tau x)=T_\tau f(x)$ 8 ms (Pellegrini, 18 Mar 2026).

The benchmarked architectures exhibit sharply different symmetry behavior. CNOs are naturally suited to temporal translation invariance because they apply convolutions across both space and time, and the reported best configuration attains training L2 error $f(T_\tau x)=T_\tau f(x)$ 9, test L2 mean $f$ 0, and test L2 median $f$ 1. FNOs achieve the lowest training error, $f$ 2, but their translated-dynamics performance is substantially worse, with test L2 mean $f$ 3, and they also have the highest inference time. DONs and their variants are efficient but do not generalize well to the temporally shifted test set (Pellegrini, 18 Mar 2026).

The contrast between TE-TNPs and the FHN benchmark suggests a common principle: symmetry should be implemented at the representation level rather than hoped for from data coverage alone. In both cases, relative-position parameterization or spatiotemporal convolution controls generalization under time shifts more reliably than architectures that can encode absolute time.

3. Robustness, adversarial transfer, and stateless temporal consistency

In video adversarial machine learning, temporal translation invariance is enforced at the perturbation level. The temporal translation attack of “Boosting the Transferability of Video Adversarial Examples via Temporal Translation” is motivated by the observation that different video recognition models emphasize different discriminative temporal patterns, which makes standard transfer-based attacks weak across architectures. The paper defines a cyclic temporal translation operator

$f$ 4

and aggregates attack losses over the shift set $f$ 5, typically with $f$ 6 (Wei et al., 2021).

The attack objective replaces a single-clip loss with a translation-aggregated loss

$f$ 7

subject to $f$ 8, and the practical gradient is

$f$ 9

The method uses symmetric weights with three strategies—Uniform, Linear, and Gaussian—and adopts adjacent shifting with stride $f(T_\tau(x)) \approx f(x),$ 0 because random and “remote” shifting perform worse (Wei et al., 2021).

The empirical results are large on the stated benchmarks. For transfer-based attack against video recognition models, the method achieves a $f(T_\tau(x)) \approx f(x),$ 1 average attack success rate on the Kinetics-400 and $f(T_\tau(x)) \approx f(x),$ 2 on the UCF-101. Compatibility with other transfer methods is also reported: TI(1) gives $f(T_\tau(x)) \approx f(x),$ 3 on UCF and $f(T_\tau(x)) \approx f(x),$ 4 on Kinetics, whereas TI+TT(1) gives $f(T_\tau(x)) \approx f(x),$ 5 and $f(T_\tau(x)) \approx f(x),$ 6; MI(10) gives $f(T_\tau(x)) \approx f(x),$ 7 and $f(T_\tau(x)) \approx f(x),$ 8, whereas MI+TT(10) gives $f(T_\tau(x)) \approx f(x),$ 9 and $\tau$ 0 (Wei et al., 2021).

A distinct but related use of temporal invariance appears in sim2real image translation. TRITON defines temporal consistency as surface consistency: preserving the visual appearance of an object’s surface as it moves or is viewed from different angles across frames, including both camera motion and object motion or deformation. Operationally, outputs are consistent if the same 3D surface point maps to the same appearance in image space across time. The mechanism is stateless: each object label selects a learnable continuous texture function

$\tau$ 1

implemented as a Fourier feature MLP, and texture projection is performed by

$\tau$ 2

with $\tau$ 3 (Burgert et al., 2022).

The key regularizer is the unprojection consistency loss

$\tau$ 4

which penalizes variance across batch unprojections, together with the texture realism loss $\tau$ 5. Under the stated assumptions—fixed object identities, fixed UV parameterizations, and deterministic inference—the resulting pipeline is temporally consistent “over indefinite timescales” without temporal memory (Burgert et al., 2022).

The reported results connect the symmetry directly to downstream utility. On AlphabetCube-3, TRITON outperforms CycleGAN and CUT, with masked LPIPS $\tau$ 6 and masked L2 $\tau$ 7. On RobotFive, it yields the lowest mean localization error, $\tau$ 8 units, versus CUT $\tau$ 9, CycleGAN $(S_\tau f)(x,t)=f(x,t-\tau),$ 0, MUNIT $(S_\tau f)(x,t)=f(x,t-\tau),$ 1, Raster TRITON $(S_\tau f)(x,t)=f(x,t-\tau),$ 2, and TRITON without textures or consistency losses $(S_\tau f)(x,t)=f(x,t-\tau),$ 3 (Burgert et al., 2022).

Across these two lines of work, temporal translation invariance is not a single mechanism but a design pattern. In one case it is imposed by shift-aggregated adversarial optimization; in the other it is achieved by geometry-anchored appearance functions and variance minimization in UV space. This suggests that the most stable form of temporal invariance is often mediated by a latent object or perturbation representation rather than by framewise post hoc smoothing.

4. Signal representations and alignment-based formulations

In signal processing, temporal translation invariance is often sought at the level of a representation rather than a predictor. The general scattering-transform framework treats the temporal shift operator as

$(S_\tau f)(x,t)=f(x,t-\tau),$ 4

with outputs in $(S_\tau f)(x,t)=f(x,t-\tau),$ 5. For a general scattering transform of the form

$(S_\tau f)(x,t)=f(x,t-\tau),$ 6

the main quantitative result is a translation-contraction bound:

$(S_\tau f)(x,t)=f(x,t-\tau),$ 7

where $(S_\tau f)(x,t)=f(x,t-\tau),$ 8 is the Fourier support radius of the output-generating function. For coherent sequences with $(S_\tau f)(x,t)=f(x,t-\tau),$ 9, this yields

$O$ 0

thereby recovering and generalizing Mallat’s limiting translation invariance (Czaja et al., 2024).

The same argument applies to a modified Fourier scattering transform built from a uniform covering frame satisfying

$O$ 1

for all $O$ 2. The paper explicitly states the new upper bound

$O$ 3

which makes the time-shift stability constant depend only on the low-pass support radius and the signal energy (Czaja et al., 2024).

A different formulation appears in time-series alignment. “Time Series Alignment with Global Invariances” treats temporal variability through DTW or soft-DTW and global feature-space transformations. For discrete-time series $O$ 4, the temporal shift operator is

$O$ 5

and a distance is translation-invariant if

$O$ 6

for any admissible $O$ 7. Standard DTW, however, uses boundary constraints and therefore does not automatically yield pure global translation invariance. Within the framework, two equivalent realizations are natural: align $O$ 8 with a time-shifted $O$ 9 and take the minimum over $O[S_\tau I](x,t)=S_\tau O[I](x,t),$ 0, or enlarge the admissible path family to open-begin/open-end alignments (Vayer et al., 2020).

The paper’s base geometry is DTW with Global Invariances,

$O[S_\tau I](x,t)=S_\tau O[I](x,t),$ 1

together with its differentiable counterpart

$O[S_\tau I](x,t)=S_\tau O[I](x,t),$ 2

Translation invariance can then be incorporated explicitly by optimizing over $O[S_\tau I](x,t)=S_\tau O[I](x,t),$ 3 in addition to $O[S_\tau I](x,t)=S_\tau O[I](x,t),$ 4 and the alignment variables (Vayer et al., 2020).

These two literatures separate two classical routes to temporal invariance. Scattering achieves it by contraction under shifts after low-pass readout; alignment methods achieve it by quotienting over shifts or by admitting them as alignment parameters. The former yields a stable representation, while the latter yields a symmetry-aware distance.

5. Deformations and symmetry breaking in relativistic and quantum systems

In relativistic phase-space models, temporal translation invariance can be deformed without deforming Lorentz symmetry. In the Snyder realization of doubly special relativity, the translation generators are distinguished from the physical momenta. The physical momenta are nonlinear functions of auxiliary variables, and in physical variables $O[S_\tau I](x,t)=S_\tau O[I](x,t),$ 5 the generators are

$O[S_\tau I](x,t)=S_\tau O[I](x,t),$ 6

The generator of time translations is therefore

$O[S_\tau I](x,t)=S_\tau O[I](x,t),$ 7

so $O[S_\tau I](x,t)=S_\tau O[I](x,t),$ 8 except in the $O[S_\tau I](x,t)=S_\tau O[I](x,t),$ 9 limit. The action of translations on coordinates becomes momentum dependent:

$(S_\tau f)(x,t)=f(x,t-\tau)$ 00

and specifically

$(S_\tau f)(x,t)=f(x,t-\tau)$ 01

Time-translation invariance is therefore encoded in the conservation of $(S_\tau f)(x,t)=f(x,t-\tau)$ 02, not of the physical energy $(S_\tau f)(x,t)=f(x,t-\tau)$ 03 (Mignemi, 2010).

The Lorentz sector remains undeformed: “the action of the Lorentz group on the momentum variables is not affected.” The translation algebra itself also remains standard, with $(S_\tau f)(x,t)=f(x,t-\tau)$ 04, but conservation laws and momentum composition are changed because physical momenta add nonlinearly through the map $(S_\tau f)(x,t)=f(x,t-\tau)$ 05. For free dynamics, the Hamiltonian

$(S_\tau f)(x,t)=f(x,t-\tau)$ 06

produces

$(S_\tau f)(x,t)=f(x,t-\tau)$ 07

while the 3-velocity remains classical,

$(S_\tau f)(x,t)=f(x,t-\tau)$ 08

Thus the deformation acts through translation generators, conservation laws, and the massive dispersion relation rather than through a modified velocity law (Mignemi, 2010).

A very different phenomenon occurs in the large- $(S_\tau f)(x,t)=f(x,t-\tau)$ 09 disordered quantum “2+p” model. There, temporal translation invariance means stationarity of two-time quantities,

$(S_\tau f)(x,t)=f(x,t-\tau)$ 10

or, in Matsubara space,

$(S_\tau f)(x,t)=f(x,t-\tau)$ 11

The FRG analysis performs coarse-graining over the Wigner spectrum of the rank-2 disorder, which makes the beta functions explicitly scale dependent and prevents global fixed points. For sufficiently strong rank- $(S_\tau f)(x,t)=f(x,t-\tau)$ 12 disorder, finite-scale singularities appear in the flow, and the paper shows via the Luttinger–Ward formalism that these singularities hide a phase transition that breaks time-translation invariance (Achitouv et al., 2024).

The diagnostic is a frequency-nondiagonal quadratic vertex,

$(S_\tau f)(x,t)=f(x,t-\tau)$ 13

together with a self-energy ansatz

$(S_\tau f)(x,t)=f(x,t-\tau)$ 14

Any non-diagonal dependence in $(S_\tau f)(x,t)=f(x,t-\tau)$ 15 breaks frequency conservation and hence Euclidean TTI. The resulting Landau potentials, such as

$(S_\tau f)(x,t)=f(x,t-\tau)$ 16

admit nonzero minima along singular trajectories, which the paper interprets as TTSB phases. Replica-diagonal TTSB is continuous, while some replica off-diagonal channels are first order (Achitouv et al., 2024).

These two physics uses of temporal translation invariance are structurally opposite. In the Snyder model, the symmetry is preserved but realized through a deformed generator. In the disordered quantum model, the symmetry is genuinely broken and stationarity fails. The shared element is that the temporal generator, or the two-time kernel that replaces it, is the correct object for diagnosis.

6. Generalised time-translation invariance and ageing dynamics

Ageing systems provide a third possibility: ordinary time-translation invariance is broken, but a generalized representation of the time-translation generator still constrains observables. In equilibrium, stationarity is generated by

$(S_\tau f)(x,t)=f(x,t-\tau)$ 17

so two-time correlators and responses depend only on $(S_\tau f)(x,t)=f(x,t-\tau)$ 18. After a quench, this ordinary TTI is lost, and simple ageing takes the form

$(S_\tau f)(x,t)=f(x,t-\tau)$ 19

with $(S_\tau f)(x,t)=f(x,t-\tau)$ 20 (Henkel, 23 Apr 2025).

The central postulate of generalized time-translation-invariance is a representation change

$(S_\tau f)(x,t)=f(x,t-\tau)$ 21

which yields

$(S_\tau f)(x,t)=f(x,t-\tau)$ 22

Whenever $(S_\tau f)(x,t)=f(x,t-\tau)$ 23, standard TTI is softly broken: the algebra still contains a generator denoted $(S_\tau f)(x,t)=f(x,t-\tau)$ 24, but its action on correlators is no longer that of ordinary stationarity (Henkel, 23 Apr 2025).

For a two-point function $(S_\tau f)(x,t)=f(x,t-\tau)$ 25, covariance under generalized time translations and dilations gives the Ward identities

$(S_\tau f)(x,t)=f(x,t-\tau)$ 26

$(S_\tau f)(x,t)=f(x,t-\tau)$ 27

Their general solution for $(S_\tau f)(x,t)=f(x,t-\tau)$ 28 is

$(S_\tau f)(x,t)=f(x,t-\tau)$ 29

This formulation preserves an algebraic remnant of time translations while allowing non-stationary dependence on both $(S_\tau f)(x,t)=f(x,t-\tau)$ 30 and $(S_\tau f)(x,t)=f(x,t-\tau)$ 31 (Henkel, 23 Apr 2025).

One of the principal consequences is the equality of autocorrelation and autoresponse exponents. For the auto-correlator and auto-response, the paper derives

$(S_\tau f)(x,t)=f(x,t-\tau)$ 32

hence

$(S_\tau f)(x,t)=f(x,t-\tau)$ 33

It also derives the extended Janssen–Schaub–Schmittmann relation, valid for all $(S_\tau f)(x,t)=f(x,t-\tau)$ 34 for global observables,

$(S_\tau f)(x,t)=f(x,t-\tau)$ 35

For a fully finite system, generalized TTI reduces finite-size scaling to one-variable forms such as

$(S_\tau f)(x,t)=f(x,t-\tau)$ 36

and predicts plateau scalings including $(S_\tau f)(x,t)=f(x,t-\tau)$ 37 at fixed $(S_\tau f)(x,t)=f(x,t-\tau)$ 38 (Henkel, 23 Apr 2025).

The framework is not universal without qualification. Its stated assumptions include the scaling regime $(S_\tau f)(x,t)=f(x,t-\tau)$ 39, simple ageing with a single growing length $(S_\tau f)(x,t)=f(x,t-\tau)$ 40, and temperatures $(S_\tau f)(x,t)=f(x,t-\tau)$ 41. Conserved dynamics, multiscaling, logarithmic growth, or initial long-range correlations lie outside the basic hypotheses (Henkel, 23 Apr 2025).

Generalized time-translation-invariance therefore occupies an intermediate logical position between exact stationarity and explicit symmetry breaking. It does not restore equilibrium TTI, but it replaces it with a covariant generator whose Ward identities fix large parts of the ageing phenomenology.

7. Conceptual synthesis

Across the cited literatures, temporal translation invariance is not a single theorem but a family of symmetry statements indexed by the object being transformed. In stationary prediction problems, it is most naturally an equivariance of operators or posterior maps; in adversarial robustness and representation learning, it becomes an approximate invariance engineered through gradient aggregation, low-pass readout, or geometry-anchored textures; in alignment methods, it is induced by optimization over shifts or relaxed boundary conditions; in high-energy and disordered quantum physics, it is either deformed at the level of generators or broken through non-stationary two-time kernels; and in ageing, it reappears as a generalized symmetry represented by $(S_\tau f)(x,t)=f(x,t-\tau)$ 42 rather than by $(S_\tau f)(x,t)=f(x,t-\tau)$ 43 alone (Ashman et al., 2024, Wei et al., 2021, Mignemi, 2010, Henkel, 23 Apr 2025).

A second unifying distinction is between output invariance and latent or generator-level invariance. TE-TNPs use relative-position attention; CNOs exploit spatiotemporal convolutions; TRITON anchors appearance to fixed texture functions; TT attacks optimize perturbations over cyclically shifted clips; scattering transforms contract translations through the support of the output-generating filter; the Snyder model distinguishes translation generators from physical momenta; and the 2PI treatment of quantum disorder identifies TTI breaking through frequency-nondiagonal self-energy structure (Ashman et al., 2024, Burgert et al., 2022, Czaja et al., 2024, Achitouv et al., 2024).

A plausible implication is that temporal translation invariance is most robust when encoded in the symmetry action itself—through group-aware architectures, generator definitions, or quotient constructions—rather than inferred indirectly from finite training distributions. The comparative evidence in neural processes, neural operators, video attacks, sim2real consistency, and scattering representations is consistent with that interpretation (Pellegrini, 18 Mar 2026, Wei et al., 2021, Czaja et al., 2024).