TimeMoDE: A Multi-Domain Temporal Framework

• TimeMoDE is a multifaceted term that covers diverse applications from quantum temporal modes and spatiotemporal forecasting to generative diffusion models and process algebra.
• It employs structured time representations—such as temporalmode quantum pulse gates, dynamic mode decomposition, and sparse expert diffusion—to enhance data encoding and model fidelity.
• The framework has practical implications across high-dimensional quantum state tomography, improved forecasting performance, efficient multi-mode spatial data modeling, and foundational theories of time.

TimeMoDE is a non-standard label that appears in several distinct arXiv research contexts rather than denoting a single canonical framework. In the supplied literature it refers, explicitly or descriptively, to: a temporalmode-based quantum-information platform built around temporal modes and quantum pulse gates; a DMD-based time representation for spatiotemporal forecasting; an alias applied to SimMST, a simple framework for multi-mode spatial-temporal modeling; a diffusion-based Mixture-of-Experts generator for scarce time series; and, more conceptually, a modular algebraic theory of time grounded in process, durons, and bi-algebraic structure (Ansari et al., 2017, Ashby et al., 2020, Kong et al., 1 Jun 2025, Liu et al., 2023, Yao et al., 13 Jun 2026, Hiley, 2013).

1. Terminological scope and principal usages

A common misconception is that TimeMoDE denotes one standardized architecture. In the available literature, the label is applied to technically unrelated constructions spanning quantum optics, spatiotemporal forecasting, generative modeling, and process algebra. In some cases it is the formal model name; in others it is a descriptive shorthand for a temporalmode-based or dynamic-mode-based program.

Usage of TimeMoDE	Domain	Technical core
Temporalmode-based platform	Quantum information	TM encoding, QPG tomography, TM unitary control
“time via dynamic modes”	Spatiotemporal forecasting	TDMD, SPDMD, DMD temporal embeddings
SimMST alias	Multi-mode spatial-temporal forecasting	TDL, CSRL, CCL
Time Mixture-of-Domain-Experts	Time-series generation	DiT backbone, MoE, Domain Prompts, prototypes
Time-theoretic program	Process algebra and foundations	durons, commutator/anti-commutator dynamics, bi-algebra

This plurality is substantive rather than merely terminological. The quantum-optical usage is centered on temporal modes of single photons; the forecasting usage treats dynamic modes as data-driven time coordinates; the generative usage deploys domain-aware sparse expert routing inside diffusion transformers; and the algebraic usage treats temporal extension itself as primitive. A plausible implication is that the label has functioned less as a fixed acronym than as a reusable marker for research programs in which time is represented through structured modes rather than as a bare scalar index.

2. Temporal-mode quantum information and temporalmode devices

In quantum optics, the temporalmode-oriented usage of TimeMoDE is built on temporal modes (TMs), defined as orthogonal broadband wavepacket shapes of a single photon in time or frequency. A TM basis is written through mode operators

$$\hat{A}_i = \int f_i(\omega)\,\hat{a}(\omega)\,\mathrm{d}\omega,$$

with orthonormal spectral mode functions $f_i(\omega)$. In the QPG tomography experiment, the working spaces are $d=5$ and $d=7$, and the basis functions are Hermite–Gaussian temporal or spectral modes (Ansari et al., 2017).

The quantum pulse gate (QPG) is the central device in this usage. It is a dispersion-engineered sum-frequency generation process in a titanium-indiffused, periodically poled LiNbO$_3$ waveguide, using a 1550 nm weak coherent signal, an 873 nm shaped pump, and a green SFG output. The governing frequency-conversion Hamiltonian is

$$\hat{H}_\mathrm{FC} = \theta \iint f^\alpha(\omega_\mathrm{in},\omega_\mathrm{out})\, \hat{a}(\omega_\mathrm{in}) \hat{b}^\dagger(\omega_\mathrm{out})\, \mathrm{d}\omega_\mathrm{in}\mathrm{d}\omega_\mathrm{out} + \mathrm{h.c.},$$

with a Schmidt decomposition that yields independent TM conversion channels and efficiencies $\eta_k^\alpha = \sin^2(|\theta \lambda_k^\alpha|)$. In the ideal single-mode limit, only one Schmidt coefficient is nonzero, so the QPG acts only on one selected temporal mode while leaving orthogonal modes unconverted (Ansari et al., 2017).

The 2017 measurement-tomography work reconstructs the QPG as a family of POVM elements

$$\hat{M}^\alpha = \sum_k \eta_k^\alpha |k^\alpha\rangle\langle k^\alpha|,$$

using weak coherent states in a tomographically complete set of mutually unbiased bases. The reconstruction is performed via weighted least squares with Hermiticity and positivity constraints. With spectral filtering on the SFG output, the reported average measurement-operator purity and fidelity are $\mathcal{P}_\text{measured}=0.920\pm0.024$ and $\mathcal{F}_\text{measured}=0.912\pm0.046$ for $f_i(\omega)$0; for $f_i(\omega)$1, the corresponding values are $f_i(\omega)$2 and $f_i(\omega)$3. Once these calibrated POVMs are used in state reconstruction, TM state tomography reaches $f_i(\omega)$4 in $f_i(\omega)$5 and $f_i(\omega)$6 in $f_i(\omega)$7, showing that an imperfect QPG can nonetheless serve as a high-dimensional TM analyzer after calibration (Ansari et al., 2017).

The complementary 2020 work addresses the manipulation component of the same temporalmode program. It proves, at the level of finite-dimensional discretized TM spaces, that any unitary can be implemented through alternating time-domain and frequency-domain phase operations,

$f_i(\omega)$8

realized physically by electro-optic phase modulators and large group-delay dispersion from chirped fiber Bragg gratings. Numerical simulations report greater than $f_i(\omega)$9 fidelity for several experimentally relevant transformations, including $d=5$0 single-mode shaping with $d=5$1 at $d=5$2, a 2-mode Hadamard with $d=5$3 and $d=5$4, a 4-mode Hadamard with $d=5$5 and $d=5$6, and a 3-mode TM demultiplexer with average fidelity $d=5$7 (Ashby et al., 2020).

Taken together, these works specify the detection-and-control stack required for a temporalmode platform: preparation and encoding of photons in chosen TMs, mode-selective conversion and measurement via calibrated QPG POVMs, and general TM unitaries via sequential time and frequency phase modulation. This suggests a concrete hardware interpretation of TimeMoDE as a temporalmode information-processing architecture rather than a single component.

3. Dynamic modes as a time representation for spatiotemporal forecasting

A very different usage treats TimeMoDE as “time via dynamic modes”, a DMD-based time embedding method for spatiotemporal forecasting. The starting point is a spatial network $d=5$8 with node signals $d=5$9, where long-range seasonal dependencies are poorly captured by short observation windows and conventional encodings such as time-of-day, day-of-week, fixed sinusoidal embeddings, or even learned sinusoidal encodings like Time2Vec. The proposed alternative is to derive time representations directly from the observed data through Dynamic Mode Decomposition and a Koopman-operator perspective (Kong et al., 1 Jun 2025).

The DMD approximation is written as

$d=7$0

with $d=7$1 a Vandermonde matrix defined by eigenvalues $d=7$2, so that

$d=7$3

To make the decomposition robust in the forecasting regime $d=7$4, the method uses a circulant Hankel embedding, Total DMD with the method of snapshots, and sparsity-promoting DMD. The resulting temporal modes are the rows of the pruned Vandermonde matrix $d=7$5, and the real-valued embedding at time $d=7$6 is

$d=7$7

No calendar features or absolute timestamps are required; the time coordinates are inferred from the multivariate dynamics themselves (Kong et al., 1 Jun 2025).

The embedding is model-agnostic. For an input tensor $d=7$8, it is broadcast across nodes and concatenated as

$d=7$9

and can therefore be inserted into FC-LSTM, DCRNN, Graph WaveNet, AGCRN, or analogous architectures. The reported datasets are GZ-METRO, PEMS04, and Daymet, with forecasting horizon $_3$0. Representative 12-step RMSE improvements include Graph WaveNet on GZ-METRO from $_3$1 to $_3$2, DCRNN on GZ-METRO from $_3$3 to $_3$4, DCRNN on PEMS04 from $_3$5 to $_3$6, and Graph WaveNet on Daymet from $_3$7 to $_3$8. Against explicit time encodings on Graph WaveNet, TimeMoDE achieves the best 12-step scores on all listed datasets, including MAE/RMSE $_3$9 on GZ-METRO, $$\hat{H}_\mathrm{FC} = \theta \iint f^\alpha(\omega_\mathrm{in},\omega_\mathrm{out})\, \hat{a}(\omega_\mathrm{in}) \hat{b}^\dagger(\omega_\mathrm{out})\, \mathrm{d}\omega_\mathrm{in}\mathrm{d}\omega_\mathrm{out} + \mathrm{h.c.},$$0 on PEMS04, and $$\hat{H}_\mathrm{FC} = \theta \iint f^\alpha(\omega_\mathrm{in},\omega_\mathrm{out})\, \hat{a}(\omega_\mathrm{in}) \hat{b}^\dagger(\omega_\mathrm{out})\, \mathrm{d}\omega_\mathrm{in}\mathrm{d}\omega_\mathrm{out} + \mathrm{h.c.},$$1 on Daymet (Kong et al., 1 Jun 2025).

The method also reduces systematic periodic structure left in the residuals. For GZ-METRO with Graph WaveNet, the mean absolute residual correlation decreases from $$\hat{H}_\mathrm{FC} = \theta \iint f^\alpha(\omega_\mathrm{in},\omega_\mathrm{out})\, \hat{a}(\omega_\mathrm{in}) \hat{b}^\dagger(\omega_\mathrm{out})\, \mathrm{d}\omega_\mathrm{in}\mathrm{d}\omega_\mathrm{out} + \mathrm{h.c.},$$2 to $$\hat{H}_\mathrm{FC} = \theta \iint f^\alpha(\omega_\mathrm{in},\omega_\mathrm{out})\, \hat{a}(\omega_\mathrm{in}) \hat{b}^\dagger(\omega_\mathrm{out})\, \mathrm{d}\omega_\mathrm{in}\mathrm{d}\omega_\mathrm{out} + \mathrm{h.c.},$$3 at lag $$\hat{H}_\mathrm{FC} = \theta \iint f^\alpha(\omega_\mathrm{in},\omega_\mathrm{out})\, \hat{a}(\omega_\mathrm{in}) \hat{b}^\dagger(\omega_\mathrm{out})\, \mathrm{d}\omega_\mathrm{in}\mathrm{d}\omega_\mathrm{out} + \mathrm{h.c.},$$4, from $$\hat{H}_\mathrm{FC} = \theta \iint f^\alpha(\omega_\mathrm{in},\omega_\mathrm{out})\, \hat{a}(\omega_\mathrm{in}) \hat{b}^\dagger(\omega_\mathrm{out})\, \mathrm{d}\omega_\mathrm{in}\mathrm{d}\omega_\mathrm{out} + \mathrm{h.c.},$$5 to $$\hat{H}_\mathrm{FC} = \theta \iint f^\alpha(\omega_\mathrm{in},\omega_\mathrm{out})\, \hat{a}(\omega_\mathrm{in}) \hat{b}^\dagger(\omega_\mathrm{out})\, \mathrm{d}\omega_\mathrm{in}\mathrm{d}\omega_\mathrm{out} + \mathrm{h.c.},$$6 at $$\hat{H}_\mathrm{FC} = \theta \iint f^\alpha(\omega_\mathrm{in},\omega_\mathrm{out})\, \hat{a}(\omega_\mathrm{in}) \hat{b}^\dagger(\omega_\mathrm{out})\, \mathrm{d}\omega_\mathrm{in}\mathrm{d}\omega_\mathrm{out} + \mathrm{h.c.},$$7, and from $$\hat{H}_\mathrm{FC} = \theta \iint f^\alpha(\omega_\mathrm{in},\omega_\mathrm{out})\, \hat{a}(\omega_\mathrm{in}) \hat{b}^\dagger(\omega_\mathrm{out})\, \mathrm{d}\omega_\mathrm{in}\mathrm{d}\omega_\mathrm{out} + \mathrm{h.c.},$$8 to $$\hat{H}_\mathrm{FC} = \theta \iint f^\alpha(\omega_\mathrm{in},\omega_\mathrm{out})\, \hat{a}(\omega_\mathrm{in}) \hat{b}^\dagger(\omega_\mathrm{out})\, \mathrm{d}\omega_\mathrm{in}\mathrm{d}\omega_\mathrm{out} + \mathrm{h.c.},$$9 at $\eta_k^\alpha = \sin^2(|\theta \lambda_k^\alpha|)$0. Residual ACF peaks at lags $\eta_k^\alpha = \sin^2(|\theta \lambda_k^\alpha|)$1 and $\eta_k^\alpha = \sin^2(|\theta \lambda_k^\alpha|)$2 are likewise substantially reduced. In this usage, TimeMoDE is therefore a spectral time-coordinate system extracted from multivariate dynamics, not a neural architecture in itself (Kong et al., 1 Jun 2025).

4. SimMST as a TimeMoDE-style framework for multi-mode spatial-temporal data

A separate technical exposition maps the TimeMoDE label onto SimMST, “A Simple framework for Multi-mode Spatial-Temporal data modeling.” Here the underlying object is a tensor

$\eta_k^\alpha = \sin^2(|\theta \lambda_k^\alpha|)$3

where $\eta_k^\alpha = \sin^2(|\theta \lambda_k^\alpha|)$4 is the number of modes, $\eta_k^\alpha = \sin^2(|\theta \lambda_k^\alpha|)$5 the number of spatial objects, $\eta_k^\alpha = \sin^2(|\theta \lambda_k^\alpha|)$6 the input history, and $\eta_k^\alpha = \sin^2(|\theta \lambda_k^\alpha|)$7 the channel count. The forecast target is

$\eta_k^\alpha = \sin^2(|\theta \lambda_k^\alpha|)$8

A “mode” is explicitly a type of coexisting signal over the same spatial objects, such as bike and taxi in NYC or railway and bus in Beijing (Liu et al., 2023).

SimMST is organized into three modules repeated across $\eta_k^\alpha = \sin^2(|\theta \lambda_k^\alpha|)$9 layers: Temporal Dependencies Learning (TDL), Cross-mode Spatial Relationships Learning (CSRL), and Channel Correlations Learning (CCL). TDL is an MLP-based temporal mixer with residual and LayerNorm,

$$\hat{M}^\alpha = \sum_k \eta_k^\alpha |k^\alpha\rangle\langle k^\alpha|,$$0

while CCL applies an analogous MLP across channels,

$$\hat{M}^\alpha = \sum_k \eta_k^\alpha |k^\alpha\rangle\langle k^\alpha|,$$1

The distinctive component is CSRL, which learns directed mode-pair adjacency matrices $$\hat{M}^\alpha = \sum_k \eta_k^\alpha |k^\alpha\rangle\langle k^\alpha|,$$2 from trainable in-affect and out-affect embeddings: $$\hat{M}^\alpha = \sum_k \eta_k^\alpha |k^\alpha\rangle\langle k^\alpha|,$$3 followed by self-loops and row normalization. Spatial propagation then uses both $$\hat{M}^\alpha = \sum_k \eta_k^\alpha |k^\alpha\rangle\langle k^\alpha|,$$4 and its transpose,

$$\hat{M}^\alpha = \sum_k \eta_k^\alpha |k^\alpha\rangle\langle k^\alpha|,$$5

and aggregates contributions across source modes with learnable scalar weights (Liu et al., 2023).

The framework is intentionally simple relative to GNN- and attention-heavy baselines. The reported overall time and space complexities are

$$\hat{M}^\alpha = \sum_k \eta_k^\alpha |k^\alpha\rangle\langle k^\alpha|,$$6

Empirically, SimMST is reported as best in most settings across NYC, Chicago, and Beijing. At $$\hat{M}^\alpha = \sum_k \eta_k^\alpha |k^\alpha\rangle\langle k^\alpha|,$$7, examples include NYC-Bike MAE/RMSE/CORR $$\hat{M}^\alpha = \sum_k \eta_k^\alpha |k^\alpha\rangle\langle k^\alpha|,$$8, NYC-Taxi $$\hat{M}^\alpha = \sum_k \eta_k^\alpha |k^\alpha\rangle\langle k^\alpha|,$$9, Chicago-Bike $\mathcal{P}_\text{measured}=0.920\pm0.024$0, Chicago-Taxi $\mathcal{P}_\text{measured}=0.920\pm0.024$1, Beijing-Railway $\mathcal{P}_\text{measured}=0.920\pm0.024$2, and Beijing-Bus $\mathcal{P}_\text{measured}=0.920\pm0.024$3. The paper also reports parameter-size and inference-time reductions relative to the strongest baselines: on NYC, SimMST uses $\mathcal{P}_\text{measured}=0.920\pm0.024$4k parameters versus up to $\mathcal{P}_\text{measured}=0.920\pm0.024$5k for MTGNN+, and $\mathcal{P}_\text{measured}=0.920\pm0.024$6 s/epoch inference versus $\mathcal{P}_\text{measured}=0.920\pm0.024$7 for GWNet+ (Liu et al., 2023).

In this sense, TimeMoDE does not refer to dynamic modes or diffusion, but to a simple multi-mode spatiotemporal modeling stack in which temporal, cross-mode spatial, and channel dependencies are separated into succinct MLP- and matrix-multiplication-based blocks.

5. TimeMoDE as a diffusion-based Mixture-of-Domain-Experts for scarce time series

The 2026 paper makes TimeMoDE an explicit model name and defines it as a unified generative framework for scarce time series. Its central claim is that standard deep generative models assume abundant in-domain data and therefore underperform in low-data settings. TimeMoDE addresses this through a Diffusion Transformer backbone, a Mixture-of-Domain-Experts (MoDE) module, Domain Prompts and prototypes for domain-aware routing, and diffusion-stage awareness inside the experts (Yao et al., 13 Jun 2026).

The model adopts the DDPM formulation

$\mathcal{P}_\text{measured}=0.920\pm0.024$8

with a learned reverse process

$\mathcal{P}_\text{measured}=0.920\pm0.024$9

and trains a noise predictor $\mathcal{F}_\text{measured}=0.912\pm0.046$0 with

$\mathcal{F}_\text{measured}=0.912\pm0.046$1

Noised sequences are tokenized into non-overlapping patches, linearly embedded, and processed by a DiT stack in which the standard MLP is replaced by MoDE. Each expert is a GLU block acting on an AdaLN-conditioned representation $\mathcal{F}_\text{measured}=0.912\pm0.046$2, where $\mathcal{F}_\text{measured}=0.912\pm0.046$3 embeds the diffusion timestep and any available class label (Yao et al., 13 Jun 2026).

The routing problem is nontrivial because heavily noised tokens become visually indistinguishable across domains. TimeMoDE therefore introduces Domain Prompts (DPs), constructed from representative real time series by convolution, linear projection, positional encoding, and average pooling,

$\mathcal{F}_\text{measured}=0.912\pm0.046$4

Each expert is associated with a learnable prototype subspace $\mathcal{F}_\text{measured}=0.912\pm0.046$5. Prototype learning is regularized by an intra-prototype orthogonality term and an inter-prototype separability term,

$\mathcal{F}_\text{measured}=0.912\pm0.046$6

with

$\mathcal{F}_\text{measured}=0.912\pm0.046$7

Routing scores combine DP–prototype similarity with token-dependent terms and are sparsified by top-$\mathcal{F}_\text{measured}=0.912\pm0.046$8 selection followed by softmax. A shared expert $\mathcal{F}_\text{measured}=0.912\pm0.046$9 always runs, while domain experts are activated sparsely (Yao et al., 13 Jun 2026).

Pre-training is conducted on a large multi-domain corpus covering 10 domains—including healthcare, finance, energy, traffic, cloud, human activity, machine sensors, physics, space, and nature—and 5 tasks: forecasting, simulation, anomaly detection, classification, and generation. The total pre-training loss is

$f_i(\omega)$00

where $f_i(\omega)$01 is a standard MoE load-balancing objective. During few-shot fine-tuning, prototypes are frozen, and both $f_i(\omega)$02 and $f_i(\omega)$03 are removed; only $f_i(\omega)$04 remains. This design is meant to preserve learned domain subspaces while adapting the DiT and expert parameters to scarce target data (Yao et al., 13 Jun 2026).

The reported evaluation covers percentage-based regimes $f_i(\omega)$05, $f_i(\omega)$06, $f_i(\omega)$07, fixed-sample regimes $f_i(\omega)$08, $f_i(\omega)$09, $f_i(\omega)$10, and full-shot training, using c-FID, discriminative score, predictive score, and additional feature/population metrics. Under few-shot percentage-based evaluation, TimeMoDE achieves average c-FID $f_i(\omega)$11 at $f_i(\omega)$12 data versus $f_i(\omega)$13 for ImagenFew, and the discriminative score improves by about $f_i(\omega)$14 to $f_i(\omega)$15. With only $f_i(\omega)$16 samples, c-FID is about $f_i(\omega)$17 versus $f_i(\omega)$18 for ImagenFew and above $f_i(\omega)$19 for the weaker baselines. In full-shot evaluation, TimeMoDE reports c-FID $f_i(\omega)$20, discriminative score $f_i(\omega)$21, and predictive score $f_i(\omega)$22, compared with ImagenFew at $f_i(\omega)$23 a slightly larger predictive error. The paper explicitly notes that TimeMoDE fine-tuned with only $f_i(\omega)$24 data surpasses ImagenTime trained on the full dataset (Yao et al., 13 Jun 2026).

Ablations identify the main contributors. Replacing MoDE with dense layers degrades average c-FID at $f_i(\omega)$25 data from $f_i(\omega)$26 to $f_i(\omega)$27; removing the DP/prototype router increases c-FID from $f_i(\omega)$28 to $f_i(\omega)$29; removing timestep conditioning also worsens performance; and allowing prototypes to update during fine-tuning is slightly inferior to freezing them. In the explicit 2026 usage, TimeMoDE is therefore a sparse, domain-aware, stage-aware diffusion foundation model for few-shot time-series generation (Yao et al., 13 Jun 2026).

6. TimeMoDE-like process algebra, durons, and the theory of time

The most abstract usage is not a forecasting or generative architecture at all, but a conceptual program in which a TimeMoDE-like theory of time is assembled from durons, process algebra, and bi-algebraic structure. The primitive temporal object is the duron, written as $f_i(\omega)$30, or in quantum form as the bi-local operator

$f_i(\omega)$31

A duron is an extended, ambiguous moment rather than an instantaneous time point; the interval $f_i(\omega)$32 is treated as basic, and the sharp instant emerges only in the $f_i(\omega)$33 limit (Hiley, 2013).

The associated algebra of process uses ordered pairs $f_i(\omega)$34 with defining relations including scaling, involution, addition, and especially the groupoid-like composition law

$f_i(\omega)$35

undefined when the adjacent endpoints do not match. The involution is

$f_i(\omega)$36

and coexistence is encoded by

$f_i(\omega)$37

The paper relates this structure to, but distinguishes it from, the monoidal-category language of Abramsky and Coecke: the similarity lies in compositionality, while the difference lies in the explicitly non-commutative *-algebraic setting with commutators, anti-commutators, idempotents, ideals, and later bi-algebra co-products (Hiley, 2013).

When the duration of the duron approaches the infinitesimal, the bi-local dynamics yields two coupled real equations. The mean-time equation becomes the quantum Liouville equation

$f_i(\omega)$38

while the time-difference equation becomes

$f_i(\omega)$39

The paper states that these two equations are equivalent to the Schrödinger equation and its dual. This splits the usual complex quantum evolution into a commutator dynamics and an anti-commutator energy equation (Hiley, 2013).

The theory is then doubled into a bi-algebra with co-products

$f_i(\omega)$40

which supports a direct connection to Umezawa’s thermal quantum field theory and Bogoliubov transformations between inequivalent vacua. On this doubled structure the paper identifies two types of time: Schrödinger time, associated with reversible unitary evolution within one representation or vacuum, and irreversible time, associated with transitions between inequivalent vacuum states and therefore with thermodynamic behavior (Hiley, 2013).

In this foundational usage, TimeMoDE is best understood as a modular theory of temporal extension and emergence. The term does not designate a computational model, but a program in which time is generated from process, durons, and algebraic doubling. This suggests a broader interpretive continuity across the otherwise disparate usages of the label: in each case, time is not treated as a primitive scalar alone, but is represented through structured modes, relations, or spectra.