Monotone Curve Estimation Framework

• The paper demonstrates how the framework recovers unknown monotone functions using isotonic least-squares estimation without needing derivative estimates.
• It details efficient computation via PAVA and robust inference under both short- and long-range dependence using universal limit laws.
• The approach yields asymptotically valid confidence intervals while adapting to various noise structures for practical data analysis.

A monotone curve estimation framework provides a principled approach for recovering an unknown function under a monotonicity constraint, typically non-decreasing or non-increasing, from observed data subject to stochastic noise or dependence. Such problems occur across time series, regression, Bayesian nonparametrics, and shape-constrained inference. Recent advances address both estimation and inference under dependence, optimize computational efficiency, and circumvent the need to estimate nuisance parameters like derivatives, thus broadening the applicability of monotone curve estimation in the analysis of dependent and independent observations (Bagchi et al., 2014).

1. Problem Definition and Model Setup

The core monotone curve estimation problem considers data $Y_i = m(t_i) + \varepsilon_i$, where $m:[0,1] \to \mathbb{R}$ is an unknown non-decreasing trend function, and $\varepsilon_i$ is a stationary mean-zero error sequence. Two major dependence structures are addressed:

• Short-range dependence (SRD): $$\sum_k |\text{Cov}(\varepsilon_k, \varepsilon_0)| < \infty$$; noise variance grows linearly, $\sigma_n^2 \sim n \tau^2$.
• Long-range dependence (LRD): $\sigma_n^2$ grows at rate $n^{2-d}$ with Hurst index $H=1-d/2$ ($H \in (1/2,1)$); partial sums converge to fractional Brownian motion (fBm).

A central goal is to estimate $m(\cdot)$ and construct valid local confidence intervals for its value at a fixed point $x_0 \in (0,1)$, accounting for serial dependence in errors.

2. Isotonic Least Squares Estimation and Discrepancy Statistics

Estimation proceeds via the isotonic (monotone) least-squares estimator:

$$(\hat{m}_n(t_1), \dots, \hat{m}_n(t_n)) = \underset{m_1 \leq \cdots \leq m_n}{\arg\min} \sum_{i=1}^n (Y_i - m_i)^2$$

which is extended to $[0,1]$ as a left-continuous step function.

For inference at $x_0$ about $m(x_0)$, a constrained isotonic fit $\hat{m}_n^{\,0,\theta}$ is constructed under an additional pointwise constraint at $x_0$, and discrepancy statistics quantify evidence against the null $H_0\!: m(x_0)\le\theta$:

$$T_n(x_0,\theta) = \frac{n}{\sigma_n^2} \sum_{i=1}^n\Bigl[ \hat{m}_n(t_i) - \hat{m}_n^{\,0,\theta}(t_i) \Bigr]^2$$

Both the fit and the constrained fit are computed efficiently by two runs of the Pool-Adjacent-Violators Algorithm (PAVA), yielding $O(n)$ computational complexity.

3. Confidence Interval Construction Under Dependence

Inference is based on inverting the discrepancy statistic. The limiting distribution of $T_n(x_0, \theta)$ depends crucially on the dependence structure:

• SRD Case: Under $m(x_0) = \theta$ and $\sigma_n^2/n \to \tau^2$, $T_n(x_0, \theta)/\tau^2 \to_d \mathbb{T}$ where $\mathbb{T}$ is the universal distribution (functional) of Brownian motion plus quadratic drift; the quantiles $q_{1-\alpha}$ are tabulated.

The $100(1 - \alpha)\%$ confidence interval is

$$\{\theta : T_n(x_0, \theta) \le \widehat{\tau}^2 q_{1-\alpha} \}$$

where $\widehat{\tau}^2$ is a consistent estimator, such as a Bartlett-window estimate of the long-run variance. No derivative (nuisance) estimation is needed.

• LRD Case: Under LRD, a normalized statistic converges to a functional of fBm plus quadratic drift:

$$\mathbb{T}^{(H)} = \int_{\mathbb{R}} [S(z) - S^0(z)]^2 dz$$

where $S$ and $S^0$ are left-derivatives of the greatest convex minorants of the fBm-plus-drift process, with and without clamping at $z = 0$.

The critical quantile for $H \in (0.6, 0.9)$ grows slowly in $H$; for robustness, use $H=0.95$ for conservative intervals. The resulting confidence interval is

$$\{\theta: T_n(x_0, \theta) \le \widehat{\sigma}_n^2 d_n^3 q_{1-\alpha}(H)\}$$

where $d_n \sim n^{-d/(2+d)}$ is the bandwidth in LRD normalization.

4. Universal Limit Laws and Practical Recommendations

A significant innovation is the isolation and tabulation of universal limit laws for monotone inference in dependent error regimes. The limit distribution $\mathbb{T}$ for SRD and $\mathbb{T}^{(H)}$ for LRD are independent of nuisance parameters such as the local slope $m'(x_0)$, obviating the challenging task of derivative estimation under dependence.

Implementation workflow:

1. Compute the unconstrained isotonic estimator (PAVA).
2. For a grid of $\theta$ near $\hat{m}_n(x_0)$, compute the constrained estimator and $T_n(x_0,\theta)$.
3. Estimate the long-run variance (SRD) or long-range scale (LRD), as appropriate.
4. Look up or interpolate the critical quantile for the relevant universal law.
5. Invert: the interval is all $\theta$ where $T_n(x_0,\theta)$ does not exceed the critical level.

The framework yields asymptotically honest confidence intervals (i.e., coverage approaches the nominal level uniformly across the monotone function class). Adaptive grid expansion and careful centering at $\hat{m}_n(x_0)$ are recommended for numerical stability, especially in the search for endpoints of the confidence interval.

5. Connections and Advantages

This methodology generalizes previous approaches reliant on independence or local smoothing techniques. Its principal advantages are:

• Avoidance of Nuisance Parameter Estimation: Previously, many methods required estimation of $m'(x_0)$, which is ill-posed under dependence and can severely distort inference. The new universal laws for $T_n(x_0, \theta)$ inherently factor out such quantities.
• Adaptability to Error Structure: Whether the noise is i.i.d., short-range, or exhibits strong serial correlation (LRD), the framework remains valid with only adjustments to scaling and the limiting distribution quantile.
• Efficiency: Computational cost is $O(n)$ per interval endpoint owing to the use of PAVA.

For full theoretical details, limit theorems, further quantile tables, and algorithmic implementation, see Bagchi, Banerjee & Stoev (2017) (Bagchi et al., 2014).

6. Extensions, Limitations, and Software

The framework is applicable for general monotone trends under both short- and long-range dependence, and can be extended to stepwise, piecewise, or even irregular monotone functions provided the core model is retained. The published R package “ISOTREND” implements the approach for both dependence regimes. Finer points, such as edge corrections near boundaries and finer resolution of the bandwidth constant $d_n$ under LRD, are discussed in supplementary material to (Bagchi et al., 2014).

A limitation is the requirement for knowledge or consistent estimation of the long-run variance or dependence parameters (e.g., Hurst index). While the universal limits are robust within ranges of $H$, extreme LRD ($H$ very near $1$) or misspecification of variance scaling can affect accuracy of coverage.

---

For the contemporary theory and practical methodology of monotone curve estimation under dependence, the framework described above constitutes the central reference model (Bagchi et al., 2014).