Primary-Auxiliary Filtering (PAF) Overview

• Primary-Auxiliary Filtering (PAF) is a design paradigm that decomposes processing into a primary component modeling core tasks and an auxiliary component that validates or refines results via explicit filtering rules.
• In applications like traffic imputation and open-set test-time adaptation, PAF distinguishes internal dynamics from contextual cues, enhancing robustness and performance under data uncertainties.
• The framework is applied across diverse fields—from state-space inference and Nyquist filtering to person re-identification—demonstrating optimized inference, error reduction, and improved model stability.

Primary-Auxiliary Filtering (PAF) denotes a class of designs in which a “primary” process carries the main inferential burden while an “auxiliary” process supplies contextual guidance, validation, proposal refinement, or compensation, and the two are linked by an explicit filtering rule rather than by simple aggregation. In the current literature, the term is used explicitly for open-set test-time adaptation, while closely related mechanisms appear in traffic time series imputation, sequential Monte Carlo, finite-length Nyquist filtering, and domain-generalizable person re-identification, sometimes without the label itself being used in the paper (Lee et al., 26 Aug 2025, Hu et al., 17 Nov 2025, Guarniero et al., 2015, Zhou et al., 2022, Li et al., 2023).

1. Conceptual scope and recurring structure

Across domains, PAF is best understood as a structural pattern rather than a single algorithm. The recurring decomposition is asymmetric: the primary component models the core latent structure or task objective, whereas the auxiliary component contributes information that is informative but not identical in role. The filtering mechanism then determines when the auxiliary signal should amplify, validate, reweight, project, or compensate the primary signal.

Context	Primary component	Auxiliary component
Traffic imputation	Internal spatial-temporal correlations	Timestamps, node attributes, exogenous features
OSTTA	Adapting model entropy filter	EMA-based validator and soft reweighting
State-space inference	Particle system / baseline filter	Auxiliary indices, twisting functions, mixture weights
Nyquist ISI removal	Paired Nyquist FIR filters	Auxiliary factors added to the transmit block
Person ReID	ReID identity objective	Weakly labeled saliency objective

This recurring form has several concrete realizations. In traffic imputation, the primary stream models internal graph-structured dependencies while the auxiliary stream embeds exogenous context. In open-set test-time adaptation, the primary filter is the adapting model and the auxiliary filter is an EMA model used as a validator. In sequential Monte Carlo, the auxiliary object can be an ancestor index, a twisting function $\boldsymbol{\psi}$, or an optimized mixture weight vector. In Nyquist filtering, the auxiliary mechanism is a blockwise pre-compensation vector. In person ReID, the auxiliary signal is not an input but a task gradient that is projected to avoid harming the primary objective (Hu et al., 17 Nov 2025, Lee et al., 26 Aug 2025, Branchini et al., 2020, Zhou et al., 2022, Li et al., 2023).

A common misconception is to treat PAF as mere side-information concatenation. The literature instead emphasizes explicit control: cross-gating, hard validation, soft reweighting, proposal twisting, or gradient projection. This suggests that the defining feature of PAF is not the presence of auxiliary information per se, but the presence of a rule that regulates how auxiliary information is permitted to affect the primary process.

2. Traffic time series imputation: primary and auxiliary patterns in PAST

In traffic time series imputation, PAF is instantiated by PAST, which decomposes the signal used for imputation into primary and auxiliary patterns. The traffic measurements are represented as a multivariate time series

$\mathbf{X} \in \mathbb{R}^{N \times T},$

where $N$ is the number of sensors or nodes and $T$ is the number of time steps. Primary patterns are defined as patterns originating from internal relationships between data points, namely spatial-temporal correlations among sensors and across time, encoded by the traffic graph and the time series structure itself. Auxiliary patterns are defined as patterns influenced by external factors, such as timestamps, calendar features, sensor attributes, and exogenous covariates (Hu et al., 17 Nov 2025).

The architecture separates these roles into a Graph-Integrated Module (GIM) and a Cross-Gated Module (CGM). GIM models primary patterns through dynamic graphs, interval-aware dropout, and multi-order graph convolution. For each time step $t$, a dynamic adjacency matrix $\mathbf{A}^{(t)}$ is constructed from hidden states, and multi-order propagation is written as

$$\mathbf{Z}_{t} = \sum_{k=0}^{K} \alpha_k \left(\tilde{\mathbf{A}}^{(t)}\right)^k \tilde{\mathbf{H}}_t \mathbf{W}_k.$$

Interval-aware dropout reduces reliance on unreliable internal correlations when missing intervals become long, thereby shifting emphasis toward more robust signals. CGM embeds external features into a latent auxiliary representation and performs bidirectional gating: $$\mathbf{g}^{\text{aux}\to\text{prim}}_{i,t} = \sigma\left(\mathbf{W}_{g1} \mathbf{E}_{i,t}^{\text{aux}} + \mathbf{U}_{g1} \mathbf{H}_{i,t}^{\text{prim}} + \mathbf{b}_{g1}\right),$$

$$\mathbf{g}^{\text{prim}\to\text{aux}}_{i,t} = \sigma\left(\mathbf{W}_{g2} \mathbf{H}_{i,t}^{\text{prim}} + \mathbf{U}_{g2} \mathbf{E}_{i,t}^{\text{aux}} + \mathbf{b}_{g2}\right),$$

so that each stream is filtered by the other before fusion.

The decomposition is specifically designed for three missing types: random missing, fiber missing, and block missing. Random missing still preserves many local spatio-temporal correlations, so the primary stream remains strong. Fiber missing weakens node-specific temporal continuity, making auxiliary context comparatively more important. Block missing can remove both local spatial and temporal evidence, so multi-order graph propagation and auxiliary priors become critical. The model is trained with an ensemble self-supervised framework that synthesizes masks across different missing patterns and combines reconstruction with self-supervised loss. On MeTr-LA, PeMS-Bay, and LargeST-SD, under 27 missing data conditions, PAST outperforms seven state-of-the-art baselines by up to 26.2% in RMSE and 31.6% in MAE, with especially large gains in fiber and block missing scenarios (Hu et al., 17 Nov 2025).

In this setting, “filtering” means selective trust allocation between intrinsic traffic dynamics and exogenous context. The paper explicitly states that the term PAF is not used there, but the mechanism matches the concept: the primary and auxiliary streams are filtered through each other, and interval-aware dropout further modulates dependence on the primary stream when missing intervals become long.

3. Open-set test-time adaptation: primary filter, auxiliary validator, and KIP

In open-set test-time adaptation (OSTTA), PAF is used explicitly as a two-stage entropy-based filtering mechanism. The setting begins with a labeled source-domain dataset

$$\mathcal{D}_s = \{(x_i, y_i)\}_{i=1}^{N_s}, \qquad \mathcal{Y}_s = \{1, \dots, \mathcal{C}_s\},$$

and an unlabeled target stream $\mathbf{X} \in \mathbb{R}^{N \times T},$0. Closed-set TTA assumes $\mathbf{X} \in \mathbb{R}^{N \times T},$1, whereas OSTTA assumes $\mathbf{X} \in \mathbb{R}^{N \times T},$2, so the test stream contains unknown classes. In this regime, blind entropy minimization is harmful because low-entropy adaptation on unknown samples causes confident misclassification and error accumulation (Lee et al., 26 Aug 2025).

The proposed PAF maintains three models at test time: a fixed source model $\mathbf{X} \in \mathbb{R}^{N \times T},$3, an adapting model $\mathbf{X} \in \mathbb{R}^{N \times T},$4, and an EMA model $\mathbf{X} \in \mathbb{R}^{N \times T},$5. The primary filter is

$\mathbf{X} \in \mathbb{R}^{N \times T},$6

and the auxiliary filter is

$\mathbf{X} \in \mathbb{R}^{N \times T},$7

For samples classified as closed-set by the primary filter, PAF uses a soft EMA-derived weight

$\mathbf{X} \in \mathbb{R}^{N \times T},$8

For risky entropy maximization, it applies hard filtering: only samples with both $\mathbf{X} \in \mathbb{R}^{N \times T},$9 and $N$0 are treated as open-set candidates. If $N$1 but $N$2, the sample is excluded from training. The batch objective is

$N$3

The division of labor is explicit. The adapting model is the “fast learner,” because it incorporates current target-domain statistics; the EMA model is the “cautious moderator,” because temporal averaging suppresses rapid error propagation. PAF therefore uses the adapting model for domain awareness and the EMA model for stability. Knowledge-Integrated Prediction (KIP) is a separate inference-stage mechanism that combines the source, adapting, and EMA models by confidence-based weighting,

$N$4

followed by

$N$5

Empirically, the method improves both closed-set accuracy and open-set discrimination. On CIFAR10-C + Textures-C, UniEnt+ achieves ACC 81.21, AUR 88.36, and H-S 84.63, while PAF + KIP achieves ACC 87.80, AUR 97.71, and H-S 92.49. On CIFAR100-C + Textures-C, the method reaches ACC 62.81, AUR 94.54, and H-S 75.48. An ablation on CIFAR100-C + SVHN-C reports ACC 26.27, AUR 82.98, H-S 39.91 for only the primary filter; ACC 55.83, AUR 97.11, H-S 70.90 for only the auxiliary filter; ACC 58.27, AUR 97.61, H-S 72.98 for both filters without KIP; and ACC 62.59, AUR 97.61, H-S 76.27 for the full system (Lee et al., 26 Aug 2025).

This formulation makes the asymmetry of PAF particularly clear. The auxiliary signal does not replace the primary signal; it validates or attenuates it where the cost of a wrong decision is high, especially for entropy maximization on suspected open-set samples.

4. Sequential Monte Carlo and state-space inference: auxiliary variables, twisting, and optimized proposals

In state-space models, the language of auxiliary particle filtering predates recent uses of the term PAF, but the same primary-auxiliary architecture is already present. The primary objects are the particles approximating the filtering distribution; the auxiliary objects are the variables or functions used to anticipate the next observation, bias ancestor selection, or twist the proposal and weight structure. In a general state-space model,

$N$6

and the core inferential challenge is that the likelihood $N$7 is typically intractable (Pitt et al., 2010).

The auxiliary sampling importance resampling (ASIR) formulation introduces predictive weights

$N$8

to bias ancestor selection toward particles likely to explain the next observation, followed by adapted propagation and importance correction. The simulated likelihood obtained by the auxiliary particle filter is unbiased, and this property supports pseudo-marginal MCMC. The same work emphasizes that using the particle filter within MCMC sampling is $N$9, and that partially or fully adapted particle filters can be much more efficient than the standard particle filter, especially when the signal to noise ratio is high (Pitt et al., 2010).

A more formal primary-auxiliary interpretation appears in the iterated auxiliary particle filter. There, a class of twisted models is indexed by positive functions $T$0, and each $T$1 defines a $T$2-auxiliary particle filter with an unbiased estimator of the marginal likelihood $T$3. The optimal sequence

$T$4

yields a zero-variance estimator, but is generally intractable. The iterated scheme uses a primary run of a $T$5-APF to generate particles and an auxiliary backward approximation step to refine $T$6 toward $T$7, thereby improving likelihood estimation and particle efficiency (Guarniero et al., 2015).

The optimized auxiliary particle filter extends this line by interpreting APF auxiliary variables as mixture weights in an importance sampling proposal,

$T$8

and choosing $T$9 by solving a convex non-negative least squares problem. The resulting importance weights are

$t$0

This framework proves unbiasedness and consistency of the estimators and shows that the conditional variance of the OAPF partial normalizing constant estimator is always less than or equal to that of the corresponding generalized APF estimator based on local-component denominators (Branchini et al., 2020).

In this literature, “auxiliary” therefore has a distinct meaning from exogenous covariates or multi-task supervision. It refers to proposal control, ancestor selection, twisting functions, or optimized mixture weights. A plausible implication is that PAF in state-space inference is best viewed as a proposal-design doctrine: use an auxiliary object to anticipate future likelihood information while preserving unbiasedness.

5. Finite-length Nyquist filtering: auxiliary factors as exact ISI compensation

In digital communication, PAF appears as a primary Nyquist filter plus an auxiliary correction mechanism. The primary stage consists of the conventional paired Nyquist filters $t$1 at the transmitter and $t$2 at the receiver. Because ideal Nyquist filters have infinite impulse responses, practical FIR truncation destroys exact zero-ISI behavior and leaves residual intersymbol interference. The auxiliary factor method introduces a vector of auxiliary factors $t$3 added to the transmit block $t$4,

$t$5

with the design objective that the received samples satisfy

$t$6

after passage through the finite-length transmit and receive filters (Zhou et al., 2022).

The zero-ISI condition becomes a linear system

$t$7

which yields the closed-form solution

$t$8

The paper further derives an efficient Toeplitz-based form,

$t$9

where boundary effects are handled by partitioned submatrices. Provided $\mathbf{A}^{(t)}$0 is invertible, this construction guarantees

$\mathbf{A}^{(t)}$1

so the auxiliary factors exactly cancel the ISI induced by finite truncation, regardless of the window length $\mathbf{A}^{(t)}$2.

The method is blockwise rather than streaming. It therefore introduces a transmission delay of at least one block’s transmission time, because the whole block $\mathbf{A}^{(t)}$3 must be known before computing $\mathbf{A}^{(t)}$4. The dominant per-block complexity is

$\mathbf{A}^{(t)}$5

with smaller $\mathbf{A}^{(t)}$6 terms for Toeplitz submatrix inversions. The paper reports zero relative RMS error $\mathbf{A}^{(t)}$7 at zero jitter and significantly lower BER under timing jitter for AF-SRRC and AF-BTRC than for conventional SRRC and BTRC. It also reports that the PAPR impact is small, with Table I indicating less than 6% for BPSK and less than 14% for 16-QAM (Zhou et al., 2022).

This use of PAF differs sharply from its machine-learning uses. The auxiliary component is not a side feature, a second model, or a training objective. It is a pre-compensation vector that mathematically inverts the ISI produced by the primary filters on a known block.

6. Domain generalization in person ReID: primary task, auxiliary task, and gradient filtering

In person re-identification, the primary-auxiliary decomposition is defined at the level of objectives. The primary objective is instance-level ReID, implemented as identity classification plus triplet loss,

$\mathbf{A}^{(t)}$8

The auxiliary objective is weakly supervised pedestrian saliency detection. Weak saliency labels are generated by a pre-trained saliency detector $\mathbf{A}^{(t)}$9, and the auxiliary head predicts a saliency map trained with an $$\mathbf{Z}_{t} = \sum_{k=0}^{K} \alpha_k \left(\tilde{\mathbf{A}}^{(t)}\right)^k \tilde{\mathbf{H}}_t \mathbf{W}_k.$$0 loss. The motivation is that pedestrians are treated as domain-invariant structural entities, whereas background clutter, scale, viewpoint, and other contextual patterns are domain-specific (Li et al., 2023).

The key mechanism is Primary-Auxiliary Objectives Association (PAOA), which calibrates the auxiliary gradient relative to the primary gradient on the shared backbone. Let

$$\mathbf{Z}_{t} = \sum_{k=0}^{K} \alpha_k \left(\tilde{\mathbf{A}}^{(t)}\right)^k \tilde{\mathbf{H}}_t \mathbf{W}_k.$$1

If $$\mathbf{Z}_{t} = \sum_{k=0}^{K} \alpha_k \left(\tilde{\mathbf{A}}^{(t)}\right)^k \tilde{\mathbf{H}}_t \mathbf{W}_k.$$2, the auxiliary gradient is left unchanged. If $$\mathbf{Z}_{t} = \sum_{k=0}^{K} \alpha_k \left(\tilde{\mathbf{A}}^{(t)}\right)^k \tilde{\mathbf{H}}_t \mathbf{W}_k.$$3, the method projects the auxiliary gradient onto the normal plane of the primary gradient: $$\mathbf{Z}_{t} = \sum_{k=0}^{K} \alpha_k \left(\tilde{\mathbf{A}}^{(t)}\right)^k \tilde{\mathbf{H}}_t \mathbf{W}_k.$$4 The effective backbone update is then proportional to

$$\mathbf{Z}_{t} = \sum_{k=0}^{K} \alpha_k \left(\tilde{\mathbf{A}}^{(t)}\right)^k \tilde{\mathbf{H}}_t \mathbf{W}_k.$$5

rather than to the naive sum of gradients. This is an explicit filtering rule: the auxiliary task is allowed to contribute only in directions that do not conflict with the primary task.

The model uses a ResNet-50 backbone, a primary identity head, and a lightweight CNN saliency head. In experiments, the joint multitask baseline without calibration yields average mAP 46.4, whereas adding gradient calibration yields average mAP 50.7; adding deployment optimization gives PAOA+ with average mAP 51.4. The corresponding baseline without auxiliary supervision is 45.5. An additional calibration-design ablation reports 46.4 for no calibration, 46.8 when calibrating the primary gradient using the auxiliary gradient as reference, 48.5 for mutual referencing, and 50.7 for the primary-referenced calibration used in PAOA (Li et al., 2023).

PAOA+ extends the method to test time by optimizing the auxiliary saliency loss on unlabeled target-domain data using weak saliency labels generated by the same detector. The reported best trade-off is one update step per test batch. This supports a more general view of PAF in multitask learning: the auxiliary signal can be useful precisely because it is constrained, not because it is symmetric with the primary task.

7. Comparative interpretation, misconceptions, and limitations

Taken together, these formulations show that PAF is not a single universal algorithm. The literature instead presents multiple domain-specific realizations linked by a common asymmetry. In traffic imputation, the asymmetry is between intrinsic spatial-temporal structure and exogenous context. In OSTTA, it is between a domain-aware but unstable adapting model and a stable but slower EMA model. In particle filtering, it is between the particle approximation itself and auxiliary mechanisms that encode look-ahead information. In Nyquist filtering, it is between the main FIR chain and a blockwise compensation vector. In person ReID, it is between a primary supervision signal and an auxiliary supervision signal whose gradient must be filtered before reaching shared parameters (Hu et al., 17 Nov 2025, Lee et al., 26 Aug 2025, Guarniero et al., 2015, Zhou et al., 2022, Li et al., 2023).

A second misconception is that auxiliary information is always weaker or optional. In several formulations it becomes decisive precisely when the primary pathway degrades. PAST explicitly shifts toward auxiliary patterns under long missing intervals. OSTTA uses the auxiliary EMA model to veto dangerous entropy maximization. Auxiliary particle filtering improves likelihood estimation when the observation is informative and the bootstrap proposal degenerates. Auxiliary factors remove the ISI that the primary FIR approximation cannot avoid. PAOA allows the auxiliary task to help only after its harmful component has been removed. This suggests that “auxiliary” refers to role, not importance.

The main limitations are likewise domain-specific. OSTTA remains threshold-based, depends on the EMA decay, and updates only batch-norm layers in the reported experiments (Lee et al., 26 Aug 2025). Person ReID depends on saliency quality, introduces extra gradient-processing overhead, and may overfit if too many deployment-time updates are applied (Li et al., 2023). The Nyquist formulation requires block buffering and additional computation, hence non-negligible delay (Zhou et al., 2022). In state-space inference, the optimal twist $$\mathbf{Z}_{t} = \sum_{k=0}^{K} \alpha_k \left(\tilde{\mathbf{A}}^{(t)}\right)^k \tilde{\mathbf{H}}_t \mathbf{W}_k.$$6 is intractable, and particle-MCMC remains computationally expensive because likelihood evaluation is costly (Guarniero et al., 2015, Pitt et al., 2010). Traffic imputation benefits most when timestamps, node attributes, or other auxiliary features remain informative under extensive missingness (Hu et al., 17 Nov 2025).

The broad research significance of PAF lies in this repeated design lesson: auxiliary information is most effective when it is not simply merged with the primary process, but is admitted through a selective mechanism that preserves the objective or statistical guarantees of the primary process.