---
title: Ordered Pivotal Sampling
url: https://www.emergentmind.com/topics/ordered-pivotal-sampling
type: topic
---

# Ordered Pivotal Sampling

Searching arXiv for the cited papers to ground the article.
Ordered pivotal sampling is a fixed-size, without-replacement unequal-probability sampling design for a finite population \(U\) with prescribed first-order inclusion probabilities \(\pi_U=(\pi_1,\dots,\pi_N)^\top\), \(0<\pi_k\le 1\), and \(\sum_{k\in U}\pi_k=n\). It processes units in a fixed order, resolves a sequence of local pivots or “battles,” and produces a \(0\)–\(1\) decision for each unit while matching exactly the target inclusion probabilities. The design is due to Deville and Tillé (1998) and was independently rediscovered by Srinivasan (2001). Its defining feature is that the order is not neutral: by constraining which units interact, it induces spreading along the chosen order and is therefore especially useful in longitudinal and spatial sampling, where spatial or temporal balance is desirable [1510.08895].

## 1. Origins and position within sampling theory

Ordered pivotal sampling sits within the broader theory of unequal-probability sampling without replacement. In the language of balanced sampling, pivotal sampling is the special case of the cube method in which the only balancing constraint is fixed sample size. With prescribed inclusion probabilities \(\pi_k\), it yields a fixed-size sample of size \(n=\sum_{k\in U}\pi_k\), exactly preserves the first-order marginals, and is strictly without replacement [1211.5442].

What distinguishes the ordered variant from standard pivotal sampling is the imposition of a fixed ordering of the units. Standard pivotal sampling allows an arbitrary pairing structure; ordered pivotal sampling determines the pairing structure from the given order, often spatial or temporal. This produces implicit spreading: sampled units tend to be well dispersed in that order. In contrast to simple random sampling without replacement, ordered pivotal sampling accommodates unequal \(\pi_k\); in contrast to Poisson sampling, it has fixed sample size and no independent inclusion mechanism; and in contrast to with-replacement unequal-probability designs, it induces negative dependence and avoids repeated selection of the same unit [1510.08895].

A central structural result is that ordered pivotal sampling and Deville’s systematic sampling induce the same sampling design. Algorithmically the procedures differ—ordered pivotal sampling uses sequential fights, winners, and jumpers, whereas Deville’s systematic sampling uses uniform variables and threshold crossings—but design-wise they coincide. This characterization makes it possible to derive second-order inclusion probabilities and clarifies why an appropriate ordering can reduce variance while a poor ordering causes only limited efficiency loss relative to simple random sampling [1211.5442].

## 2. Finite-population formulation and microstrata

Let
\[
V_k=\sum_{l=1}^k \pi_l,\qquad k=1,\dots,N,\qquad V_0=0.
\]
A unit \(k\in U\) is cross-border if, for some integer \(i\in\{1,\dots,n\}\),
\[
V_{k-1}< i \le V_k.
\]
These units, denoted \(k_i\) for \(i=1,\dots,n-1\), are precisely those whose cumulative inclusion probabilities straddle an integer boundary. Introducing phantom units \(k_0\) and \(k_n\) with \(\pi_{k_0}=\pi_{k_n}=0\), define
\[
a_i=i-V_{k_i-1},\qquad b_i=V_{k_i}-i,\qquad i=1,\dots,n-1,
\]
so that \(\pi_{k_i}=a_i+b_i\). The quantity \(a_i\) belongs to microstratum \(i\), whereas \(b_i\) is carried into microstratum \(i+1\) [1510.08895].

The ordered structure is encoded through microstrata
\[
U_i=\{k\in U: k_{i-1}\le k\le k_i\},\qquad i=1,\dots,n.
\]
Within each \(U_i\), define local weights
\[
\alpha_{ik}=
\begin{cases}
b_{i-1}, & k=k_{i-1},\\[4pt]
\pi_k, & k_{i-1}<k<k_i,\\[4pt]
a_i, & k=k_i,
\end{cases}
\]
which satisfy
\[
\sum_{k\in U_i}\alpha_{ik}=1.
\]
Hence each microstratum carries a local set of pseudo-probabilities summing to one, and each cross-border unit is split across two adjacent microstrata through the pair \((a_i,b_i)\). This representation is the key to both the constructive algorithm and the later martingale analysis [1510.08895].

The same structure can be reformulated on a clustered population \(U^c\) of size \(2n-1\). In that representation, cross-border units become singleton clusters and non-cross-border units between two consecutive cross-border units are grouped together. The clustered view is particularly useful for comparing ordered pivotal sampling with multinomial sampling and for proving efficiency dominance over with-replacement unequal-probability designs [1609.02688].

## 3. Sequential algorithm and design characterization

The algorithm proceeds microstratum by microstratum while propagating a single residual unit, called the jumper. Initialize with the phantom unit \(L_0=k_0\). At step \(i\), microstratum \(U_i\) receives the jumper \(L_{i-1}\) with residual probability \(b_{i-1}\). A temporary contender \(S_i\) is then selected among
\[
\{L_{i-1},k_{i-1}+1,\dots,k_i-1\}
\]
with probabilities proportional to \((b_{i-1},\pi_{k_{i-1}+1},\dots,\pi_{k_i-1})\). The contender \(S_i\) then confronts the cross-border unit \(k_i\), and the pair is updated as
\[
(S_i,k_i)\mapsto
\begin{cases}
(F_i,L_i)=(S_i,k_i) & \text{with probability } \dfrac{1-a_i}{1-b_i},\\[6pt]
(F_i,L_i)=(k_i,S_i) & \text{with probability } \dfrac{a_i}{1-b_i}.
\end{cases}
\]
The fixed unit \(F_i\) is definitively included in the sample; the loser \(L_i\) becomes the new jumper and carries residual probability \(b_i\) into the next microstratum. After the final step \(i=n\), the design returns exactly \(n\) fixed units \(F_1,\dots,F_n\), and Deville and Tillé’s result guarantees \(\Pr(k\in S)=\pi_k\) for every unit \(k\) [1510.08895].

This sequential mechanism admits a two-stage interpretation on the clustered population \(U^c\): first select \(n\) clusters by ordered pivotal sampling on cluster-level inclusion probabilities, then select one unit within each chosen cluster with probability proportional to its original \(\pi_k\). The same two-stage representation holds for multinomial sampling, which is why the comparison between the two designs reduces to a cluster-level variance comparison [1609.02688].

Chauvet’s characterization theorem shows that ordered pivotal sampling on the original population and Deville’s systematic sampling with the same parameter \(\pi\) induce the same design. The result is not merely heuristic equivalence: all sample probabilities and all joint inclusion probabilities coincide. This equivalence explains why ordered pivotal sampling inherits the implicit stratification behavior associated with systematic-type procedures while retaining the constructive pivotal representation [1211.5442].

## 4. Inclusion structure, dependence, and efficiency

The first-order inclusion probabilities are exact by construction, but the second-order inclusion probabilities \(\pi_{kl}\) are structurally more intricate. They can be computed explicitly for ordered pivotal sampling, yet many of them are zero, especially when microstrata contain several non-cross-border units. In particular, two non-cross-border units in the same microstratum cannot be selected together. This sparsity is a manifestation of the design’s negative dependence and spreading behavior, but it also complicates classical variance estimation based on pairwise inclusion probabilities [1211.5442].

The dependence structure is central to the design’s efficiency. Ordered pivotal sampling avoids with-replacement duplication, spreads selections along the ordered frame, and tends to avoid neighboring units when the order is spatial or temporal. On this basis, it is more accurate than multinomial sampling with the same parameter \(\pi\): for any variable \(y\),
\[
V(t_y^{\mathrm{ops}})\le V(t_y^{\mathrm{ms}}).
\]
The proof proceeds through the clustered population \(U^c\): both designs share the same second-stage within-cluster variance, so the comparison reduces to the first-stage cluster selection, where the variance under ordered pivotal sampling is no larger than the variance under multinomial sampling [1609.02688].

In the equal-probability case \(\pi_k=n/N\) with \(N=np\), ordered pivotal sampling reduces to stratified simple random sampling with one unit per stratum. Then
\[
V_{\mathrm{ops}}(\hat t_{y\pi})=N^2\frac{1-f}{n}\frac{1}{n}\sum_{i=1}^n S_{yi}^2,
\]
where \(f=n/N\) and \(S_{yi}^2\) is the within-stratum variance in microstratum \(U_i\). By contrast, for simple random sampling without replacement,
\[
V_{\mathrm{srs}}(\hat t_{y\pi})=N^2\frac{1-f}{n}S_y^2,
\]
and for ordered systematic sampling,
\[
V_{\mathrm{sys}}(\hat t_{y\pi})=N^2\frac{1-f}{n}\frac{1}{n}S_Y^2.
\]
In worst-case terms, the maximum design effect satisfies
\[
DMAX(\mathrm{ops})=\frac{N-1}{N-n},\qquad
DMAX(\mathrm{sys})=n\,\frac{N-1}{N-n}.
\]
Thus ordered pivotal sampling can exploit informative orderings, yet its worst-case degradation relative to simple random sampling remains much more limited than for systematic sampling [1211.5442].

A recurrent misconception is that the equivalence with Deville’s systematic sampling makes ordered pivotal sampling identical to ordinary systematic sampling. That is not the case. The equivalence is design-theoretic: the induced distribution on samples is the same as Deville’s systematic design, not the same as every systematic scheme, and the algorithmic constructions remain different [1211.5442].

## 5. Horvitz–Thompson estimation, asymptotics, and variance estimation

For a deterministic survey variable \(y=(y_1,\dots,y_N)^\top\), with total
\[
t_y=\sum_{k\in U} y_k,
\]
the Horvitz–Thompson estimator under ordered pivotal sampling is
\[
\hat t_{y\pi}=\sum_{k\in S}\frac{y_k}{\pi_k}
=\sum_{k\in S}\check y_k,\qquad \check y_k=\frac{y_k}{\pi_k}.
\]
The central asymptotic result of the martingale analysis is a central-limit theorem for \(\hat t_{y\pi}\) under the randomization induced by the design, and also under a model-assisted approach allowing correlations between values, which is especially relevant in spatial sampling [1510.08895].

The key representation is a decomposition of the estimation error into microstratum-wise increments,
\[
\hat t_{y\pi}-t_y=\sum_{i=1}^n \xi_i,
\]
where
\[
\xi_i=\check y_{F_i}+b_i\check y_{L_i}
-\sum_{k\in U_i'}\alpha_{ik}\check y_k
-b_i\check y_{k_i}.
\]
These increments satisfy
\[
E(\xi_i\mid y_U,L_{i-1})=0,
\]
and are uncorrelated. With the filtration
\[
\mathcal F_0=\sigma(y_U),\qquad
\mathcal F_i=\sigma(y_U,S_1,F_1,L_1,\dots,S_i,F_i,L_i),
\]
the normalized increments \(\eta_i=\xi_i/\sqrt{V_U(\hat t_{y\pi})}\) form a martingale difference sequence. This creates exactly the probabilistic framework required for a martingale central-limit theorem [1510.08895].

The design-based CLT is obtained under three assumptions. \(H1\) requires that no inclusion probability is too large: there exists \(f<1\) such that \(\pi_k\le f\) for all \(k\). \(H2\) is a fourth-moment bound:
\[
\sum_{k\in U}\pi_k\left(\check y_k-\frac{t_y}{n}\right)^4 \le C\frac{N^3}{n^3}.
\]
\(H3\) is a non-degeneracy condition within microstrata:
\[
\sum_{i=1}^n\sum_{k,l\in U_i}\alpha_{ik}\alpha_{il}
(\check y_k-\check y_l)^2 \ge C\frac{N^2}{n^2}.
\]
Under these assumptions,
\[
\frac{\hat t_{y\pi}-t_y}{\sqrt{V_U(\hat t_{y\pi})}}
\xrightarrow{d} N(0,1).
\]
The proof combines a lower bound on \(V_U(\hat t_{y\pi})\), a fourth-moment bound for the increments \(\xi_i\), and a bound on the variance of the conditional variances, thereby verifying the martingale CLT conditions [1510.08895].

A separate asymptotic result establishes weak consistency of the Horvitz–Thompson estimator under mild moment conditions, using the fact that ordered pivotal sampling is more efficient than multinomial sampling. If a finite-population moment condition analogous to
\[
\sum_{k\in U}\pi_k\bigl(\check y_k-t_y/n\bigr)^2 \le A_1\,\frac{N}{n^{1/2}}
\]
holds, then
\[
E\Big[N^{-1}\bigl(t_y^{\mathrm{ops}}-t_y\bigr)^2\Bigr]=O(n^{-1}),
\]
so the estimator is weakly consistent [1609.02688].

Variance estimation is more delicate. Because many \(\pi_{kl}\) are zero, standard Horvitz–Thompson variance estimators based on second-order inclusion probabilities can be downward biased or undefined. A conservative alternative is the Hansen–Hurvitz variance estimator applied to the ordered pivotal sample. Its expectation satisfies
\[
E\!\left[v_{HH}(t_y^{ps})-V(t_y^{\mathrm{ops}})\right]
=\frac{1}{n-1}\Bigl[V(t_y^{\mathrm{ms}})-V(t_y^{\mathrm{ops}})\Bigr],
\]
which is nonnegative because ordered pivotal sampling is at least as efficient as multinomial sampling. Hence the Hansen–Hurvitz estimator is conservative for ordered pivotal sampling [1609.02688].

## 6. Applications, generalizations, and later developments

The most established applications are spatial sampling and ordered survey frames. Because sampled units tend to be well spread in the imposed order, ordered pivotal sampling is particularly suited to spatially balanced and temporally balanced designs. When the survey variable is spatially or temporally correlated, the order-induced spreading can translate directly into variance reduction. This is the operational setting emphasized in longitudinal surveys and spatial sampling [1510.08895].

Within the broader pivotal family, the local pivotal method generalizes the ordered case by replacing one-dimensional adjacency with nearest-neighbor pairing in an arbitrary metric space. In that framework, ordered pivotal sampling is a special case obtained when the metric is induced by a fixed one-dimensional order. The shared mechanism is the same pivotal update rule that preserves first-order inclusion probabilities while pushing nearby units apart, creating “automatic stratification” and a “thin” well-spread sample [2305.02446].

Recent machine-learning work has adapted pivotal sampling to leverage-score-based active learning. There, the pivotal algorithm is implemented on a binary tree whose structure reflects geometry, so that siblings compete early and nearby design points become negatively correlated in inclusion. The method keeps the same leverage-score marginals as independent sampling, but enforces spatial coverage. In the reported experiments, it reduces the number of samples needed to reach a given target accuracy by up to \(50\%\). The general agnostic regression guarantee is \(O(d\log d)\) samples under a one-sided \(\ell_\infty\) independence condition that includes pivotal sampling, and for polynomial regression the bound improves to \(O(d)\) samples [2310.04966].

The pivotal mechanism has also appeared in online algorithms. In online stochastic bipartite matching, a polynomial-time approximation algorithm invokes pivotal sampling on offline nodes ordered by decreasing edge weight at each time step. The prefix property of pivotal sampling then turns matching probabilities into expressions of the form \(\mathbb E[\min(1,X)]\) for weighted sums of negatively correlated Bernoulli variables, and the analysis relies on the negative cylinder dependence delivered by the pivotal step [2407.15285].

These developments suggest a broader interpretation of ordered pivotal sampling. In survey sampling, the fixed order is usually a spatial, temporal, or auxiliary-variable order. In later extensions, the same principle is encoded through trees, geometric partitions, or online priority orderings. A plausible implication is that the enduring value of ordered pivotal sampling lies not only in exact unequal-probability fixed-size selection, but also in the way it converts an ordering into a controlled dependence structure that can be exploited for dispersion, robustness, and variance reduction across distinct domains.

Source: https://www.emergentmind.com/topics/ordered-pivotal-sampling