Resolution Drift Fundamentals

• Resolution drift is the systematic change in spatial or predictive resolution caused by variable operational, environmental, or data conditions.
• In particle detectors, factors such as ionization statistics, electron diffusion, and electronics response contribute to spatial resolution variability that can be mathematically modeled and corrected.
• In federated learning and high-resolution imaging, resolution drift challenges performance fidelity, leading to solutions like multi-resolution modeling and advanced drift correction techniques.

Resolution drift refers to the systematic change or degradation of measurement or predictive resolution as a function of operational, environmental, or data heterogeneity factors. The term appears with significant technical meaning in at least two distinct domains: (i) particle physics and nuclear instrumentation—particularly drift chambers and time projection chambers (TPCs), where it denotes the spatial resolution variation with detector parameters such as drift distance, gas mixture, irradiation rate, and electronics; and (ii) distributed machine learning, where it formalizes the performance gap in federated learning (FL) caused by input-resolution heterogeneity across clients. Additionally, it has relevance in electron holography, where mechanical or sample drift must be corrected to preserve atomic-scale spatial resolution. A precise characterization of resolution drift and its mitigation is essential for sustaining fidelity in both measurement and inference under variable experimental or data conditions.

1. Definition and Fundamental Mechanisms

Resolution drift is, in all contexts, a manifestation of the dependence of spatial (or functional) resolution on variable system parameters. In drift chambers (e.g., ATLAS MDT, MEG II, COMET CDC), resolution drift quantifies the change in hit or track coordinate resolution $\sigma$ with varying drift time, distance from sense wire, gas properties, irradiation rate, or applied voltage. Mathematically, for drift chambers:

$$\sigma(r) = \sqrt{\sigma_{\rm ion}^2(r) + \sigma_{\rm diff}^2(r) + \sigma_{\rm ele}^2}$$

where:

• $\sigma_{\rm ion}(r)$: resolution component due to primary-ionization statistics,
• $\sigma_{\rm diff}(r)$: diffusion-induced component,
• $\sigma_{\rm ele}$: electronics time/walk contribution.

In federated learning, resolution drift is defined as the performance (e.g., accuracy) degradation when models trained on heterogeneous input resolutions are evaluated at a resolution different from any participating client:

$$\forall r' \notin \{r_k\}, \quad |\text{Perf}(w^*; r') - \text{Perf}(w^*; r_k)| \text{ is large}$$

Here, $w^*$ is the global model, $r_k$ are client-native resolutions, and Perf denotes a task-specific metric (e.g., PCKh in pose estimation) (Lim et al., 31 Jul 2025).

2. Resolution Drift in Drift and Tracking Chambers

In gaseous drift chambers, resolution drift is influenced by the interplay of stochastic ionization, diffusion, and technical factors:

• Primary-ionization cluster statistics: Low numbers of primary electrons per mm (e.g., $\sim$14–15 clusters/cm in He-based mixtures (Baldini et al., 2016, Wu et al., 2021)) lead to $\sigma_{\rm ion}$ dominating near the wire and for low-$$\sigma(r) = \sqrt{\sigma_{\rm ion}^2(r) + \sigma_{\rm diff}^2(r) + \sigma_{\rm ele}^2}$$0 gases.
• Transverse and longitudinal electron diffusion: $$\sigma(r) = \sqrt{\sigma_{\rm ion}^2(r) + \sigma_{\rm diff}^2(r) + \sigma_{\rm ele}^2}$$1 increases with drift distance; e.g., in the COMET CDC, $$\sigma(r) = \sqrt{\sigma_{\rm ion}^2(r) + \sigma_{\rm diff}^2(r) + \sigma_{\rm ele}^2}$$2m at $$\sigma(r) = \sqrt{\sigma_{\rm ion}^2(r) + \sigma_{\rm diff}^2(r) + \sigma_{\rm ele}^2}$$3 mm (Wu et al., 2021).
• Electronics jitter and time walk: Fast low-noise electronics and time-slewing corrections can significantly reduce $$\sigma(r) = \sqrt{\sigma_{\rm ion}^2(r) + \sigma_{\rm diff}^2(r) + \sigma_{\rm ele}^2}$$4 (e.g., in ATLAS, slewing reduced $$\sigma(r) = \sqrt{\sigma_{\rm ion}^2(r) + \sigma_{\rm diff}^2(r) + \sigma_{\rm ele}^2}$$5 from $$\sigma(r) = \sqrt{\sigma_{\rm ion}^2(r) + \sigma_{\rm diff}^2(r) + \sigma_{\rm ele}^2}$$6m to $$\sigma(r) = \sqrt{\sigma_{\rm ion}^2(r) + \sigma_{\rm diff}^2(r) + \sigma_{\rm ele}^2}$$7m without irradiation) (Deile et al., 2016).
• High irradiation rates or background (“rate effects”): Increased photon or $$\sigma(r) = \sqrt{\sigma_{\rm ion}^2(r) + \sigma_{\rm diff}^2(r) + \sigma_{\rm ele}^2}$$8 background rates lead to additional space-charge, pulse pile-up, and masking dead time, linearly degrading spatial resolution as $$\sigma(r) = \sqrt{\sigma_{\rm ion}^2(r) + \sigma_{\rm diff}^2(r) + \sigma_{\rm ele}^2}$$9 (Deile et al., 2016).

The following table summarizes observed dependencies:

System	Gas	Key Drift Parameter	$\sigma_{\rm ion}(r)$0 (low)	$\sigma_{\rm ion}(r)$1 (high)
ATLAS MDT	Ar:CO$\sigma_{\rm ion}(r)$2	Rate $\sigma_{\rm ion}(r)$3 (Hz/cm$\sigma_{\rm ion}(r)$4)	82 $\sigma_{\rm ion}(r)$5m	108 $\sigma_{\rm ion}(r)$6m (500 Hz)
COMET CDC	He:iC$\sigma_{\rm ion}(r)$7H$\sigma_{\rm ion}(r)$8	Drift distance $\sigma_{\rm ion}(r)$9 (mm)	100 $\sigma_{\rm diff}(r)$0m ($\sigma_{\rm diff}(r)$1→0)	200 $\sigma_{\rm diff}(r)$2m ($\sigma_{\rm diff}(r)$3~8)
MEG II prototypes	He:iC$\sigma_{\rm diff}(r)$4H$\sigma_{\rm diff}(r)$5	Gas mixture, $\sigma_{\rm diff}(r)$6 (mm)	97 $\sigma_{\rm diff}(r)$7m	138 $\sigma_{\rm diff}(r)$8m

Resolution drift in these systems is thus a direct function of physical operating conditions (Deile et al., 2016, Baldini et al., 2016, Wu et al., 2021).

3. Quantitative Modeling and Correction Methods

Mathematical modeling of resolution drift in gaseous detectors involves decomposing observed resolution into irreducible statistical components and those addressable by hardware or analysis improvements:

• Quadrature Decomposition:

$\sigma_{\rm diff}(r)$9

with $\sigma_{\rm ele}$0 growing linearly with irradiation rate and dominating at large radii (Deile et al., 2016).

• Time-slewing and pulse-height corrections: Application of $\sigma_{\rm ele}$1 exploits correlation between pulse height and discriminator time walk, yielding substantial $\sigma_{\rm ele}$2 gains in $\sigma_{\rm ele}$3 (Deile et al., 2016).
• Neural Network (NN) corrections: Recent work demonstrates replacement of analytic or simulation-based time-to-distance calibrations with data-driven dense NNs or CNNs that learn directly from waveform features. These approaches achieve a further $\sigma_{\rm ele}$4 reduction in $\sigma_{\rm ele}$5 across drift distances, with the greatest gains near the sense wire (Palo et al., 2023).

Empirical parameterizations allow predictive modeling, as in ATLAS MDT:

• Without slewing: $\sigma_{\rm ele}$6
• With slewing: $\sigma_{\rm ele}$7 (Deile et al., 2016).

4. Resolution Drift in Federated Learning and AI Workflows

In FL, resolution drift is a form of non-IID client heterogeneity that is orthogonal to label distribution shifts:

• Formalization: Performance at test time on unseen or low-resolution data diverges from in-sample accuracy due to the altered conditional distributions $\sigma_{\rm ele}$8 induced by resolution variability (Lim et al., 31 Jul 2025).
• Manifestation: In human pose estimation, for example, FedAvg trained with client images at (128×96, 192×144, 256×192) achieves only ~51–52% PCKh at 128×96, indicating severe drift (Lim et al., 31 Jul 2025).
• Mitigation: Resolution-Adaptive Federated Learning (RAF) augments the local client objective with a multi-resolution knowledge-distillation loss, using higher-resolution model outputs (“teachers”) to guide lower-resolution (“students”). This addresses performance drift by enforcing feature consistency across resolutions without penalizing global convergence rates.

A plausible implication is that tasks relying on spatial detail, such as regression or semantic segmentation, are especially susceptible to resolution drift, which cannot be mitigated solely by strategies developed for label non-IIDness.

5. Drift Correction in High-Resolution Imaging

In quantitative phase-imaging modalities such as phase-shifting off-axis electron holography, spatial resolution drift arises primarily from mechanical specimen or instrument drift:

• Problem: Even sub-Ångstrom scale sample displacement between exposures mixes amplitude and phase, destroying atomic resolution (Lindner et al., 2023).
• Correction Pipeline: A multi-stage registration—using phase-correlation on vacuum–specimen interfaces, drift vector correction, and pixel-wise least-squares fitting—permits reliable extraction of phase information up to the information limit (here, 1 Å).
• Outcome: The application of precise per-frame drift corrections enables the recovery of high-fidelity amplitude and phase signals up to 2π/452 at the 1 Å limit, matching multislice simulation benchmarks (Lindner et al., 2023).

This approach generalizes to any transmission electron microscope equipped with an electron biprism and underscores the criticality of robust correction schemes for atomic-resolution imaging.

6. Practical Implications, Limitations, and Design Strategies

Resolution drift impinges directly on the design, calibration, and operation of precision measurement and inference systems:

• Tracking detectors: Failure to control $\sigma_{\rm ele}$9 and $$\forall r' \notin \{r_k\}, \quad |\text{Perf}(w^*; r') - \text{Perf}(w^*; r_k)| \text{ is large}$$0 compromises pattern recognition, Kalman filtering, and physics reach in experiments; redundancy (multiple wire layers), optimized cell geometries, wider bandwidth electronics, and advanced analytics (e.g., NNs, cluster timing) are standard countermeasures (Deile et al., 2016, Baldini et al., 2016, Palo et al., 2023).
• Federated ML systems: Robustness to resolution drift requires explicit architectural and objective modifications (e.g., knowledge distillation), as opposed to generic techniques for non-IIDness (Lim et al., 31 Jul 2025).
• Imaging and microscopy: Experimental calibration pipelines must incorporate real-time or post-processing drift correction to extract meaningful high-fidelity information (Lindner et al., 2023).

A plausible implication is that resolution drift, in all forms, acts as a unifying constraint on the attainable granularity of both measurement and inference, demanding domain-specific engineering solutions.

7. Summary

Resolution drift is a rigorously defined phenomenon describing systematic changes in measurement or prediction granularity under parameter variability. In drift and tracking chambers, it is primarily governed by physical stochasticity (ionization, diffusion), operational parameters (irradiation rates, voltage), electronics response, and, increasingly, the sophistication of reconstruction algorithms. In federated learning, it embodies a distinct axis of client heterogeneity, unamenable to correction by label-level solutions. In high-resolution imaging, uncontrolled drift imposes strict limits on attainable information bandwidth. Across domains, resolution drift elucidates the boundary between fundamental limits and achievable performance, and ongoing innovation in hardware, calibration procedures, and algorithmics continues to refine its mitigation (Deile et al., 2016, Baldini et al., 2016, Wu et al., 2021, Palo et al., 2023, Lim et al., 31 Jul 2025, Lindner et al., 2023).