When modeling real-world phenomena, researchers frequently encounter a fundamental dilemma. Parametric models (such as linear regression) are simple and statistically efficient, but they risk severe model misspecification if the true relationship is non-linear. Conversely, fully nonparametric models can capture arbitrary shapes, but they are often mathematically intractable because they operate in an infinite-dimensional space of possible functions.
Sieve estimation resolves this tension. The core intuition is straightforward: instead of searching over an infinite-dimensional space of all possible functions, we restrict our search to a sequence of simpler, finite-dimensional approximating spaces—called a "sieve." As more data becomes available, we allow the sieve to grow finer and more complex, enabling the model to capture increasingly intricate details without ever losing tractability.
The Mechanics of Basis Function Expansion
To build a sieve in practice, an unknown, complex function $f(x)$ is expressed as a linear combination of known, flexible basis functions $\psi_k(x)$:
$$f(x) \approx \sum_{k=1}^{K} \beta_k \psi_k(x)$$
This transformation turns a complex nonparametric problem into a parametric-like problem where only the coefficients $\beta_k$ need to be estimated. As highlighted by Gao et al. (2020), representing unknown functions with polynomial splines maps a nonparametric additive model into a high-dimensional space of spline coefficients, making it highly amenable to standard regularization techniques.
Depending on the structure of the data, different types of basis functions are chosen:
- Polynomial Splines and B-splines: These are excellent for localized smoothness and capturing non-linear relationships step-by-step. They are widely used to handle smooth relationships in high-dimensional additive models (Fan and Feng (2011)) and to model baseline event rates in panel count data (Zhao et al. (2017)).
- Wavelets and Fourier Series: These are ideal for periodic or highly localized spatial structures. For example, Bigot et al. (2010) project infinite-dimensional stochastic covariance processes onto these bases to simplify estimation without assuming a rigid parametric shape.
Beyond standard regression, this framework adapts well to complex structural data. For instance, Chen et al. (2024) incorporate auxiliary side-information into low-rank tensor decompositions by using sieve approximations to flexibly map covariates to latent factors.
The Dynamic Balance: Why the Sieve Must Grow
The power of sieve estimation lies in how the number of basis functions $K$ scales with the sample size $n$. If $K$ remains fixed, the model behaves like a rigid parametric model and suffers from approximation bias. If $K$ grows too quickly relative to $n$, the model overfits, causing the estimation variance to explode.
As Ding et al. (2022) explain, sieve estimators explicitly manage this trade-off between approximation errors and estimation variance within infinite-dimensional function spaces. By carefully calibrating the growth of the basis functions as more data is acquired, the estimator adaptively captures the underlying smoothness of the target function.
This careful scaling allows sieve estimators to achieve minimax optimal convergence rates—the theoretical speed limit for learning from data. For instance, Kwon and Chae (2024) demonstrated that restricting generative density estimation to an expanding sequence of function classes achieves these optimal rates. Similarly, in Bayesian frameworks, Yoo et al. (2023) proved that a data-dependent sieve prior using B-splines achieves minimax optimal convergence for complex hazard models.
Robust Causal Inference and Treatment Effects
In causal inference, researchers must estimate "nuisance functions"—such as propensity scores or baseline outcomes—to isolate the true impact of a treatment. If these nuisance models are misspecified, the resulting treatment effect estimates become biased.
Sieve estimation provides a robust shield against this misspecification. Fan et al. (2021) utilize nonparametric sieve estimation to achieve optimal covariate balancing without assuming rigid functional forms for propensity scores. This flexibility is vital when evaluating complex causal pathways, such as estimating mediation effects (Huang et al. (2024)) or continuous causal functions where unobserved confounding is present (Kawakami et al. (2024)).
When generalizing clinical trial results to broader target populations, Lee et al. (2022) employ sieves to approximate complex survival outcomes flexibly. Furthermore, when integrating randomized experiments with external controls or real-world data, the sieve framework allows for the modeling of complex, non-linear bias functions. Wu et al. (2023) and Mao et al. (2025) demonstrate that combining sieve expansions with roughness penalties allows models to safely "borrow" information from external datasets while protecting against model misspecification and overfitting.
Policy Evaluation in Reinforcement Learning
Estimating the value of a policy from offline data (Off-Policy Evaluation, or OPE) requires learning continuous value functions (Q-functions). Sieve methods have become a cornerstone for establishing the statistical guarantees of these algorithms.
Wang et al. (2024) showed that linear sieve approximations allow Fitted Q-evaluation (FQE) to achieve optimal convergence rates by adaptively capturing value function smoothness over time horizons. By framing Q-function estimation as a nonparametric instrumental variables problem, Chen and Qi (2022) used sieve bases to establish minimax optimal rates even for continuous states and actions.
Crucially, this finite-dimensional projection provides the mathematical tractability needed for formal statistical inference. Luo et al. (2025) leveraged this property to construct simultaneous confidence bands, moving beyond pointwise estimates to quantify uncertainty across an entire continuous state space at once.
Enhancing Modern Neural Architectures
The principles of sieve estimation are not confined to classical statistics; they are actively reshaping deep learning architectures. A prime example is the emergence of Kolmogorov–Arnold Networks (KANs) as alternatives to Multi-Layer Perceptrons (MLPs).
Instead of using fixed activation functions on nodes, KANs place learnable univariate functions on the edges. As analyzed by Zhang and Zhou (2025), these edge functions are implemented exactly as linear combinations of fixed basis functions, such as B-splines, wavelets, or Fourier series. This structural connection allows researchers to apply the theoretical machinery of sieve estimation directly to deep networks, proving that generalization bounds depend on the smoothness and coefficients of these bases rather than merely the count of hidden nodes.
Key Foundational Work
Time-dependent coefficients in a Cox-type regression model (1991) — contributed 2 concept elements to this direction.
Instrumental variable estimation of nonseparable models (2006) — contributed 1 concept elements to this direction.