On the Learnability of Test-Time Adaptation: A Recovery Complexity Perspective¶
Conference: ICML 2026
arXiv: 2605.28057
Code: None
Area: Learning Theory / Test-Time Adaptation / Non-stationary Online Learning
Keywords: TTA Learnability, Recovery Complexity, Wasserstein Quantization, ϕ-mixing, Minimax Lower Bounds
TL;DR¶
This paper establishes the first theoretical framework for Test-Time Adaptation (TTA) learnability by introducing \((\epsilon, \delta)\)-Recovery Complexity to measure the time required to reduce excess risk to \(\epsilon\) after a distribution shift. By extending local recovery to non-stationary test streams via \((\epsilon, \rho)\)-TTA Learnability and deriving matching minimax upper/lower bounds, the work reveals the intrinsic "adaptation speed vs. information constraint" tradeoff in TTA.
Background & Motivation¶
Background: TTA has achieved empirical success across vision, tabular, and NLP tasks (e.g., Tent, CoTTA, NOTE, ODS) by performing online model updates using only unlabeled test data guided by a proxy loss \(\psi\). However, these methods often collapse under complex distribution shifts, prompting the community to reconsider the conditions under which TTA is reliable.
Limitations of Prior Work: Existing online learning theories (regret frameworks) characterize cumulative performance rather than instantaneous reliability questions like "how long does it take to recover after a shift?" Meanwhile, prior theoretical analyses of TTA (e.g., AdaNPC or ATTA) impose restrictive assumptions, such as memory-based architectures or the availability of active labeling, which do not hold for general unlabeled streams.
Key Challenge: The fundamental goal of TTA is to maintain acceptable instantaneous risk at every time step. Existing frameworks fail to characterize either post-shift recovery or the information constraints caused by the mismatch between the proxy loss and the true loss on unlabeled streams.
Goal: Construct a unified framework that simultaneously expresses (i) continuous and abrupt distribution shifts, (ii) temporal correlations, (iii) proxy-task mismatch, and (iv) post-shift recovery speed, while deriving matching minimax bounds under a stochastic proxy-gradient oracle.
Key Insight: Discretize the test stream into "piecewise stationary" approximations using Wasserstein-1 distance to analyze each segment as an independent recovery problem. Utilize ϕ-mixing coefficients to characterize temporal dependencies within batches and apply Le Cam’s two-point method for information-theoretic lower bounds.
Core Idea: Transform instantaneous reliability into an analyzable complexity measure called "recovery complexity" \(\tau(\epsilon, \delta)\), and aggregate multi-segment recovery behaviors into a global \((\epsilon, \rho)\)-TTA Learnability. This reduces the problem of TTA learnability to sample/batch complexity analysis in classical stochastic optimization.
Method¶
Overall Architecture¶
The theoretical framework consists of four interconnected components forming a "reduction pipeline": (1) Test stream formalization using Wasserstein-1 quantized distribution shifts and ϕ-mixing temporal dependencies; (2) Definition of a proxy-optimal competitor \(\theta_t^\star \in \arg\min_{\theta\in\mathcal{N}_r(\theta_1)} \psi_t(\theta)\) and its true risk \(R_t:=\ell_t(\theta_t^\star)\) as a baseline; (3) Local analysis of minimax bounds for \((\epsilon, \delta)\)-recovery complexity after a single shift; (4) Global aggregation of recovery behaviors into \((\epsilon, \rho)\)-TTA Learnability compared against dynamic regret. The input is a non-stationary stream \(\mathcal{S}=\{D_t\}_{t=1}^{T}\), and the output is a provable bound regarding the parameter regimes (alignment \(\alpha\), mismatch bias \(\zeta\), batch size \(B\), mixing constant \(C_\phi\), shift magnitude \(\Delta_W\)) under which TTA is learnable.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Input: Non-stationary test stream<br/>S = {D_t}, unlabeled + proxy loss ψ"]
A --> B["Wasserstein-1 Quantized Distribution Shift<br/>Segmented into stationary approximations, shift count K_S ≤ ⌈2V_T/Δ_W⌉"]
A --> C["ϕ-mixing Temporal Dependency<br/>Correlated batch B → Effective batch B/C_φ"]
B --> D["(ε,δ)-Recovery Complexity Local Analysis<br/>Relative to proxy-optimal competitor θ*, Le Cam lower bound + PL upper bound match"]
C --> D
D --> E["(ε,ρ)-TTA Learnability Global Aggregation<br/>ρ ≤ δ + (K_S+1)·τ/T, piecewise recovery → long-term reliability"]
E --> F["Output: Provable bounds for TTA Learnability<br/>Comparison with dynamic regret, revealing adaptation-information tradeoff"]
Key Designs¶
1. Wasserstein-1 Quantization of Distribution Shifts: Approximating Non-stationary Trajectories To analyze continuous shifts and abrupt changes, the authors discretize the trajectory \(\{\mathcal{P}_t\}\) into piecewise stationary approximations \(\{\tilde{\mathcal{P}}_t\}\) using Wasserstein-1 distance, ensuring approximation error \(\le \Delta_W/2\). A greedy algorithm maintains an anchor; if the \(W_1\) distance of \(\mathcal{P}_t\) from the anchor is \(\le \Delta_W/2\), the anchor is kept; otherwise, a shift is declared and the anchor is reset. The total number of shifts is bounded by \(\tilde{K}_S(T) \le \lceil 2V_T/\Delta_W\rceil\) (where \(V_T\) is total variation). \(W_1\) is chosen to unify covariate and label shifts on the joint space \(\mathcal{X}\times\mathcal{Y}\) and support information-theoretic lower bounds.
2. ϕ-mixing Dependency and Effective Batch Size: Compressing Temporal Correlation into \(C_\phi\) TTA test streams often contain non-i.i.d. samples (e.g., video frames). Authors use the ϕ-mixing coefficient to characterize this: assuming geometric decay \(\phi(i)\le \varrho^i\), the variance of a correlated batch of size \(B\) is equivalent to an i.i.d. batch of size \(B_{\text{eff}} = B/C_\phi\), where $\(C_\phi = 1 + \frac{4\varrho^{1/2}}{1-\varrho^{1/2}}.\)$ This compresses complex stochastic process properties into a single scalar, allowing batch size and independence to be compared directly in the complexity bound.
3. Minimax Bounds for \((\epsilon, \delta)\)-Recovery Complexity: Characterizing Difficulty The authors define recovery complexity \(\tau(\epsilon, \delta) := \inf\{t: \sup_{u\ge t}\mathbb{P}(\mathcal{E}_u > \epsilon) \le \delta\}\), representing how fast the excess risk \(\mathcal{E}_t := \ell_t(\theta_t) - R_t\) drops below \(\epsilon\) with probability \(\ge 1-\delta\). Using Le Cam’s method, they prove any stochastic proxy-gradient oracle algorithm faces \(\tau \ge \Omega\big(\frac{C_\phi}{B}\cdot\frac{1}{\alpha(\sqrt{\zeta+2\alpha\epsilon}+\sqrt{\zeta})^2}\big)\). A simplified TTA baseline achieves a matching upper bound under \(L\)-smooth and PL conditions. This reveals that mismatch \(\zeta\) creates an "error floor," and \(\tau\) scales with \(1/\alpha^2\).
4. \((\epsilon, \rho)\)-TTA Learnability: Aggregating Local Recovery for Long-term Reliability \((\epsilon, \rho)\)-TTA Learnability is defined as \(\frac{1}{T}\sum_{t=1}^{T}\mathbb{P}\big(\ell_t(\theta_t)-R_t>\epsilon\big)\le\rho\), meaning the proportion of time steps where excess risk exceeds \(\epsilon\) is at most \(\rho\). Theorem 4.3 provides the local-to-global transition: if an algorithm has \((\epsilon', \delta)\)-recovery complexity \(\tau\) on each stationary segment, the stream is \((\epsilon, \rho)\)-learnable with $\(\rho\le\delta+\frac{(\tilde{K}_S(T)+1)\,\tau(\epsilon',\delta)}{T}.\)$ This demonstrates that designing learnable TTA reduces to improving local recovery complexity.
Loss & Training¶
The paper is theoretical and analyzes a TTA baseline using stochastic proxy-gradient descent on \(\psi\) with step size \(\eta\) within a local neighborhood \(\mathcal{N}_r(\theta_1)\). Key assumptions include \((\alpha, \zeta)\)-Alignment (Assumption 2.1), \(L\)-smoothness + PL condition + bounded gradient variance (Assumption 2.2), and Wasserstein quantization + ϕ-mixing (Assumptions 2.4/2.7).
Key Experimental Results¶
Main Results¶
The primary "experiment" is the theoretical comparison of the framework against existing tools regarding their expressiveness for TTA.
| Framework | Instantaneous Reliability | Unlabeled Proxy Mismatch | Non-stationary Stream Shape | Temporally Correlated Batch | Minimax Bounds |
|---|---|---|---|---|---|
| Static Online Regret | No | No | Fixed comparator only | Weak | No |
| Dynamic Regret | Partial | No | Arbitrary path variation | Weak | No |
| AdaNPC | No | Memory-based only | Covariate only | No | No |
| ATTA | Partial | Requires active labels | General | No | No |
| Ours | Yes | \(\alpha, \zeta\) explicit | \(W_1\) Quantization | ϕ-mixing → \(C_\phi\) | Matching |
Ablation Study¶
The impact of shutting down various factors in the recovery complexity lower bound (from Remark 3.4–3.7):
| Configuration | Lower Bound Magnitude | Description |
|---|---|---|
| Full bound | \(\Omega\big(\frac{C_\phi}{B\alpha}(\sqrt{\zeta+2\alpha\epsilon}+\sqrt{\zeta})^{-2}\big)\) | Full matching order lower bound |
| \(\zeta=0\) | \(\Omega(C_\phi/(B\alpha^2\epsilon))\) | Matches classical stochastic optimization orders |
| \(\varrho=0\) (\(C_\phi=1\)) | \(\Omega(1/(B\alpha^2\epsilon))\) | Temporal correlation term disappears |
| \(\alpha \downarrow\) | \(\tau \uparrow\) by \(1/\alpha^2\) | Weaker proxy-task alignment slows recovery |
| \(\Delta_W\) change | Fixed | Shift magnitude triggers but does not determine recovery difficulty |
Key Findings¶
- Proxy mismatch \(\zeta\) creates an error floor; recovery complexity does not reach zero even as \(\epsilon \to 0\), explaining the unavoidable minimum risk in unlabeled TTA.
- Batch size \(B\) and ϕ-mixing \(C_\phi\) jointly dominate difficulty; increasing batch size cannot compensate for high temporal correlation if the effective batch \(B/C_\phi\) remains small.
- Alignment strength \(\alpha\) has a quadratic impact, implying high returns for designing proxy losses better aligned with task gradients.
- TTA Learnability is distinct from dynamic regret: the former requires recovery to a threshold \(\epsilon\) per segment, while the latter measures cumulative performance.
Highlights & Insights¶
- The use of Wasserstein-1 quantization to discretize non-stationary paths into piecewise stationary approximations is a powerful reduction technique applicable to other streaming learning scenarios like continual learning or online RL.
- Compressing temporal dependencies into a single scalar \(C_\phi\) provides a quantifiable measure that can be directly compared with batch size and alignment.
- The \(\zeta\)-induced error floor provides a formal explanation for why self-training or pseudo-labeling TTA methods fail under high-noise or biased scenarios.
Limitations & Future Work¶
- The analysis is restricted to a local neighborhood \(\mathcal{N}_r(\theta_1)\), assuming PL and smoothness properties hold locally, which may not be true under extreme shifts.
- The upper bounds rely on a simplified stochastic gradient descent baseline; actual TTA methods using BN-only updates or teacher-student mechanisms still have a theoretic gap.
- Since this is a theoretical work, there is a lack of empirical validation on standard TTA benchmarks (e.g., CIFAR-C) to quantify these complexity factors.
Related Work & Insights¶
- Comparison with Dynamic Regret (Zhao et al., 2024): While dynamic regret looks at cumulative gap, this work focuses on per-segment recovery time, providing the conditions to convert learnability to regret.
- Comparison with AdaNPC (Zhang et al., 2023): Unlike AdaNPC which is restricted to memory-based non-parametric classifiers, this work abstracts analysis to the algorithm-class level using energy oracles.
- Comparison with ATTA (Gui et al., 2024): ATTA uses active labeling to simplify the problem to supervised online learning; this work focuses on the harder unlabeled setting where proxy mismatch \(\zeta\) is the primary hurdle.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ First theoretical framework for TTA learnability.
- Experimental Thoroughness: ⭐⭐ Purely theoretical; lacks empirical validation on TTA benchmarks.
- Writing Quality: ⭐⭐⭐⭐ Clear structure and flow, though notation-heavy.
- Value: ⭐⭐⭐⭐ Provides a formal language and targets for future proxy loss and reset strategy designs.
Related Papers¶
- [ICML 2026] MMD-Balls as Credal Sets: A PAC-Bayesian Framework for Epistemic Uncertainty in Test-Time Adaptation
- [ICLR 2026] Expressive Power of Implicit Models: Rich Equilibria and Test-Time Scaling
- [ICML 2026] Semi-Supervised Noise Adaptation: Transferring Knowledge from Noise Domain
- [ICML 2026] Catastrophic Forgetting is Low-Rank: A Function-Space Theory for Continual Adaptation
- [ICLR 2026] A Statistical Learning Perspective on Semi-dual Adversarial Neural Optimal Transport Solvers