Temporal Discrepancy Tomography

• Temporal Discrepancy Tomography is a framework that models time-localized events to reconstruct dynamic system changes with sub-frame resolution.
• It utilizes methods such as event-parameterization, basis decomposition, and wavelet transforms to quantify temporal discrepancies effectively.
• TDT reduces data complexity and enhances efficiency across imaging, signal processing, and quantum optics by compressing high-dimensional data.

Temporal Discrepancy Tomography (TDT) encompasses a set of methodologies for reconstructing the evolution of physical, informational, or quantum systems with explicit modeling of time-localized changes and non-stationary structure. Across materials science, signal processing, and quantum optics, TDT refers to approaches that exploit temporal discrepancies—shifts, transitions, or anomalies occurring in time—to achieve high-resolution dynamical insight. Core to these techniques is the replacement of traditional frame-based paradigms or scalar summaries with models or transforms that preserve, extract, and localize events or statistical deviations in time, enabling sub-frame or multi-scale detection and reconstruction in both experimental and algorithmic domains.

1. Principles and Conceptual Foundation

TDT methods depart from classical static or frame-averaging approaches by leveraging explicit temporal modeling. Central concepts include:

• Event-parameterization: Representing system evolution as discrete or continuous transitions parameterized by local times or statistical discrepancies, rather than as a stack of global time frames.
• Temporal localization: Identifying and reconstructing the timing and extent of dynamic events—such as phase transitions in imaging, structural changes in quantum states, or statistical anomalies in generated texts—at resolutions finer than acquisition intervals.
• Discrepancy quantification: Defining and extracting “temporal discrepancies” via basis function decompositions, transition-time mappings, or signal processing transforms, such as wavelets, to capture both the location and the magnitude of temporal variation.
• Dimension reduction and efficiency: Compressing data representations from high-dimensional stacks (frame-based) to compact summaries in terms of key temporal parameters, substantially reducing memory and computational loads.

These techniques are driven by the need to overcome the limitations of observing only a single viewpoint at a time (imaging), intrinsic non-stationarity (text), or limited mode distinguishability (quantum optics).

2. Mathematical Models and Representations

TDT implementations are characterized by mathematical formalisms that encapsulate temporal evolution:

A) Event-based Piecewise Constant Models (Goethals et al., 2024)

Each spatial voxel $x$ is modeled as switching once from an initial value $\mu_0(x)$ to a final value $\mu_1(x)$ at unknown time $\tau(x)$:

$$A(x, t) = \mu_0(x)\; H(\tau(x) - t) + \mu_1(x)\; H(t-\tau(x))$$

with $H(\cdot)$ the Heaviside function. This representation reduces the continuous temporal axis to per-voxel transitions, enabling memory- and data-efficient reconstruction of dynamic material processes.

B) Temporal Basis Function Decomposition (Nikitin et al., 2018)

The time-variant state $x(x,y,z,t)$ is decomposed as:

$$x(x, y, z, t) \approx \sum_{k=0}^{M-1} x_k(x, y, z)\, \phi_k(t)$$

where $\phi_k(t)$ are chosen basis functions (e.g., Fourier modes, wavelets, steps) and $x_k(\cdot)$ the corresponding spatial volumes. This reduces $N_x N_y N_z N_t$ unknowns to $M N_x N_y N_z$ coefficients, capturing complex dynamics with adjustable temporal resolution.

C) Time-Frequency Signal Decomposition (West et al., 3 Aug 2025)

For a sequence of discrepancy scores $\{z_i\}$ on e.g. tokenized text, the signal

$$\widetilde Z(x, t) = \sum_{i=1}^n z_i\, \frac{1}{\sqrt{2\pi}h} \exp\left(-\frac{(t-i)^2}{2h^2}\right)$$

is subjected to Continuous Wavelet Transform (CWT):

$$W(a, b) = \frac{1}{\sqrt{a}} \int_{-\infty}^\infty \widetilde Z(x, t)\, \psi^*\left(\frac{t-b}{a}\right)\, dt$$

yielding a two-dimensional time-scale “scalogram” for anomaly localization at multiple linguistic scales.

3. Computational Algorithms and Optimization

Event-Based Iterative SIRT (Goethals et al., 2024)

TDT in CT employs an alternating-minimization, ordered-subsets SIRT-like algorithm. At each iteration:

• Updates are performed for transition times $\tau(x)$ by evaluating time-aware covariance balances of projection residuals before and after the current $\tau(x)$.
• Initial and final values $\mu_0(x), \mu_1(x)$ are updated by least squares over rays segregated by the current estimate of $\tau(x)$.
• Memory requirements are reduced to three 3D volumes.
• Complexity scales as $$O(N_{\text{rays}}\, N_{\text{samples per ray}}) + O(N_{\text{voxels}}\, N_{\text{proj per voxel}})$$.

Primal-Dual TV-Regularized Reconstruction (Nikitin et al., 2018)

The reconstruction is formalized as minimization:

$$\min_x \left\{ \frac{1}{2} \|A x - d\|_2^2 + \lambda_1 \|\, |\nabla_{\lambda_2} x|\, \|_1 \right\}$$

with joint spatio-temporal TV regularization and solved using the Chambolle–Pock primal-dual method. Efficient GPU implementations exploit repeated Radon transforms and memory tiling by $z$-slabs for large datasets.

Continuous Wavelet Classification and SVM (West et al., 3 Aug 2025)

• Discrepancy signals are wavelet-transformed, and the wavelet energy is computed for empirically-determined linguistic bands (morphological, syntactic, discourse).
• A three-dimensional energy vector is classified by an RBF-kernel SVM.
• The approach delivers robust anomaly detection in highly non-stationary and adversarial scenarios.

Quantum Overlap Tomography (Tiedau et al., 2017)

• Uses controlled mode overlap (by delay) between a signal and a probe.
• Measurement statistics (on/off click probability as a function of delay) yield linear constraints on photon-number populations, allowing for inversion (via regularized least squares or maximum-likelihood EM).
• Modal overlap functions are reconstructed, linking time-delay statistics to spectral properties of quantum states.

4. Performance Metrics, Experimental Validation, and Practical Impact

Time-Resolved Imaging and Event Recovery (Goethals et al., 2024, Nikitin et al., 2018)

• Synthetic and experimental validation demonstrate sub-frame temporal resolution, with mean absolute errors of $\sim$0.088 rotations, or less than one-tenth the time frame of conventional CT.
• TDT reliably identifies local events, such as rupture times in metallic foams, to within a single projection, outperforming frame-based SART reconstuction which distributes events over broader time windows.
• Robustness is retained across changes in angular acquisition and flow direction, with errors consistently well below frame duration.

LLM Anomaly Detection (West et al., 3 Aug 2025)

• TDT achieves AUROC of 0.855 on the RAID benchmark (6.9% gain), and excels under adversarial paraphrasing (0.812, a 14.1% gain over alternatives).
• Generalizes to multiple languages and models; e.g., Spanish HART L2 scores 0.699 AUROC (+25.5%).
• Only 13% computational overhead relative to baseline (e.g., 58 vs. 51 ms for 512 tokens), with fundamentally linear scaling.

Quantum State and Mode Characterization (Tiedau et al., 2017)

• Experimental demonstration with heralded single- and two-photon states shows that TDT reconstructs diagonal density matrices with fidelities up to 98%, with correct identification of photon-number populations.
• Modal structure can be extracted directly, provided probe temporal profile is known, through Fourier inversion of recovered overlap functions.

5. Limitations, Variants, and Future Directions

• Single-event per voxel restriction: Some TDT forms (Goethals et al., 2024) are limited to one transition per voxel. Systems with repeated or non-monotonic changes require model extension to multi-step temporal parameterizations.
• Sensitivity to temporal basis selection: Accuracy and computational load depend on the number and nature of temporal basis functions chosen (Nikitin et al., 2018). Adaptive or wavelet-like bases may better capture local dynamics with fewer components.
• Regularization and priors: Lack of spatial prior on transition times may introduce noise in reconstructions (Goethals et al., 2024). Total variation or localized priors are recommended for stabilizing solutions.
• Data requirements: Event-based updates typically require at least three rotations (for SIRT convergence and covariance estimation). Sparse or partial data settings may necessitate reformulation.
• Scalability and real-time potential: GPU-based implementations scale efficiently with hardware improvements (Nikitin et al., 2018). Near-real-time analysis is plausible with further optimization of memory hierarchies and data transfer pipelines.

A plausible implication is that, as acquisition rates in imaging and complexity in generative models increase, TDT frameworks—which directly target the temporal structure of discrepancies—will become central to high-fidelity, efficient, and robust dynamical reconstruction across scientific domains.

6. Applications Across Domains

Domain	TDT Formulation	Key Capabilities
Materials tomography	Event-based, piecewise-constant models (Goethals et al., 2024)	Sub-frame localization of transitions
Medical imaging	Temporal basis decomposition, TV-regularization (Nikitin et al., 2018)	Denoised, motion-robust 4D reconstructions
AI text analysis	CWT on discrepancy signals (West et al., 3 Aug 2025)	Non-stationary anomaly localization
Quantum optics	Overlap-based delay tomography (Tiedau et al., 2017)	Simultaneous state and mode recovery

These applications highlight the transdisciplinary nature of Temporal Discrepancy Tomography, with the common objective of revealing structure “in time” that is otherwise obscured by aggregation or frame-based analysis.

7. Historical Context and Theoretical Significance

The conceptual origin of TDT reflects persistent challenges in time-resolved observation: limitations of sequential data acquisition, non-stationary statistics, and the need for interpretable, temporally-resolved models. Early developments in quantum optics (Tiedau et al., 2017) exploited time-domain overlaps to reconstruct modal structure; advances in imaging (Goethals et al., 2024, Nikitin et al., 2018) broke the static-sample paradigm by directly inferring event times from projection data; in computational linguistics (West et al., 3 Aug 2025), TDT revealed fundamental non-stationarity in machine-generated text, prompting more robust detection algorithms.

This suggests that Temporal Discrepancy Tomography, by emphasizing the detection and explicit modeling of temporal events and anomalies, has established itself as an essential methodological framework for the dynamical sciences.