WorksMathematics
Sampling Design and Confidence Tubes for Integral Curves on Riemannian Manifolds
Abstract
We estimate the integral curve of an unknown vector field from noisy tangent vector field on a known compact Riemannian manifold. We consider two sampling settings: independent observations at random locations that have already been collected, and a fixed response budget that allows subsequent locations to be chosen using pilot information. Under suitable regularity assumptions, the passive kernel estimator has curve error of order in probability, with sample size and manifold dimension . Its functional central limit theorem has a limit characterized by a linear stochastic differential equation. For active sampling, an independent pilot directs one new batch of measurements to neighborhoods of the target curve. Under localization, smoothness and design conditions, this concentration yields expected maximum curve error of order , including the pilot in the total response budget . Both methods provide simultaneous geodesic confidence tubes with asymptotically exact coverage under mild conditions. Simulations show narrower active tubes in most reported settings, with improved small-sample coverage in some three-dimensional examples. Ocean-drifter observations illustrate mean-flow reconstruction on the sphere, and a canopy-flow LDA application uses retrospective active sampling to quantify lateral and vertical departures of local mean streamlines.
Introduction
Riemannian data arise when the objects being measured have geometric constraints: directions lie on spheres, orientations are rotations, and many shape representations form nonlinear spaces. Respecting this geometry is essential for comparing observations and describing their variation [23]. When the observations describe local velocities, integral curves connect these measurements to pathways and transport. Examples include direction-field trajectories in diffusion imaging and mean circulation inferred from ocean drifters [16, 20]. The statistical task is to recover a trajectory together with its uncertainty from noisy information.
Wall shear stress on a curved body provides a concrete example of this problem. The shear stress is tangent to the surface, and its integral curves, called skin-friction lines, describe the organization of near wall flow [27]. Recovering a finite segment of such a curve from sparse, noisy vector measurements requires accounting for both surface geometry and measurement uncertainty. Accurate shear measurements are often available only at a limited number of sensor locations (Rohlfs and Weiss, 2025). Since the target line is unknown in advance, a sparse pilot survey can first provide an approximate route. When further measurement locations can be chosen, the remaining budget can be concentrated along and across this estimated route to refine the curve while allowing it to depart from the pilot. Adaptive acquisition is also used in laser Doppler anemometry (LDA), Doppler lidar and underwater sampling [30, 34, 14], while streamlines inform motion planning in ocean currents (To et al., 2019). Together with existing archives such as ocean-drifter data, these settings motivate a distinction between using available observations and directing new measurements toward a curve of interest.
We formulate these questions as a statistical problem of estimating an integral curve from noisy vector measurements. At a recorded position , the tangent vector response follows
where is a known Riemannian manifold, is an unknown vector field and the errors are conditionally centered. The target is its parameterized integral curve from a given starting point over a fixed time interval. This leads to two questions.
Q1: With independent random-design observations already available, how can we estimate the entire curve and construct a simultaneous confidence tube?
Q2: If a total response budget allows us to choose new locations, how can a preliminary sample guide those measurements to improve accuracy while allowing valid simultaneous inference?
Recovering a mean wind trajectory from local velocity measurements identifies an average atmospheric transport route over both land and ocean. Figure 1 uses simulated velocities to illustrate passive estimation and pilot-guided sampling and reconstruction on Mercator maps (–), with Natural Earth land boundaries.
Figure 1. (a) Passive sampling — 160 observations. (b) Active sampling — 40 pilot + 120 new observations. Target curve; Estimated curve; Pilot curve; Pilot locations; Passive locations; New locations. Simulated atmospheric-flow reconstruction under passive and active sampling.
The first question has substantial Euclidean precedents. Koltchinskii et al. [16] study noisy vector observations at uniform random locations and derive a curve functional central limit theorem with linear stochastic error propagation. Koltchinskii and Sakhanenko [15] study curve functionals, and Sakhanenko [25] establishes global minimax rates in that regression setting. Later work studies how to accelerate integral curve estimation (Sakhanenko, [26]). Tensor based models instead infer a direction field from noisy tensor measurements; their curve estimation risks are studied by Banerjee et al. [3, 4]. These observation models must be distinguished from direct tangent vector regression. Likewise, Arias-Castro and Qiao [1] construct confidence regions for density-gradient lines from observations of positions alone. The derivative of a density estimator, rather than an observed velocity, drives those curves. Related sampling strategies arise in randomized initial value solvers, which select where to evaluate the field or its derivatives [13], and in scalar inverse regression, where a two-stage design yields a parametric rate and bootstrap inference [29].
On manifolds, Pelletier [21, 22] provides volume corrected kernel estimators for density and scalar regression. Lin and Yao [18] map observed trajectories to tangent vectors along their Fréchet mean curve to identify principal modes of variation. Furthermore, Lin and Yao [19] use local linear regression on a manifold to predict a scalar response from a noisy functional predictor. In our model, we observe noisy tangent velocities at sampled locations and recover an unobserved trajectory by integrating the estimated field. Parallel transport lets us average velocities in different tangent spaces, but field errors must then be propagated through the dynamics. The general evolution equation theory of Wei and Panaretos [33] explains this transfer and provides bootstrap results when the scaled field estimation error has a weak limit. For shrinking-bandwidth tangent regression, we additionally establish the integrated error law and feasible covariance along the random fitted curve.
To answer Q1, we integrate an intrinsic kernel estimate of and prove a functional central limit theorem for the estimation error of the entire integral curve. After rescaling and parallel transport to a common tangent space, the error process converges to a Gaussian process that solves a linear stochastic differential equation. Its drift describes how the field propagates earlier errors and incorporates the smoothing bias, while its diffusion describes how velocity noise accumulates along the curve. The curve rate agrees with the corresponding second-order Euclidean rate. Under undersmoothing, we use a Gaussian approximation at the fitted bandwidth and consistently estimate its covariance across times to construct simultaneous geodesic confidence tubes. In flat coordinates, transport is the identity and these constructions become normalized Euclidean kernel regression and its propagated error.
To answer Q2, we use an independent kernel pilot to locate the curve and collect one batch of location-response pairs in neighborhoods along the pilot. Sampling across each neighborhood identifies transverse changes in the field, allowing the reconstructed curve to depart from the pilot. Conditional on successful localization, the final responses determine the leading estimation error. Under the stated smoothness and design conditions, the expected maximum curve error is , with pilot responses included in . A kernel pilot and a fixed pilot fraction suffice in dimensions two and three at the stated smoothness.
For this active experiment, we establish simultaneous confidence tubes whose radii have order and whose coverage tends to the nominal level under undersmoothing and nondegenerate response noise. Thus concentrating observations near the curve improves estimation accuracy while retaining uncertainty quantification for the entire trajectory. The inference remains valid across changing local sampling designs, without requiring their covariances to approach a single limit. The simulations compare the two procedures at common response budgets: active tubes are narrower in most settings and improve small sample coverage in some three dimensional examples. The applications examine how observational support limits a regional mean flow curve on the sphere and use pilot-guided LDA reconstruction to quantify the stronger lateral and downward departure of local mean flow within a model canopy than above it.
The rest of the paper is organized as follows. § 2 gives the model and geometric assumptions. § 3 and § 4 develop the passive and active methods with theoretical guarantees and algorithms to answer Q1 and Q2. § 5 and § 6 report simulations and the applications. Implementation details, additional experiments and supporting application analyses appear in Appendices § 9–§ 11. Proofs are provided in the supplementary material.
Model and geometry
Let be a fixed, connected, compact Riemannian manifold without boundary, of dimension , with distance , Levi-Civita connection , exponential map , and Riemannian volume . If and lie in a common normal neighborhood, denotes parallel transport along their unique minimizing geodesic. Transport along a specified curve is written separately below.
Under passive sampling, the pairs in (1) are independent, with , and . Here is the design density and is the conditional response covariance. The pilot sample in the active design consists of independent copies of from the same regression model, with . Throughout, conditionally centered errors are sub-Gaussian in unit tangent directions:
The same finite applies to all laws in the model class and need not be known. For final active responses, the same bound is conditional on the pilot and all recorded positions. Fix and a starting point . The target is
Its parameterization is determined by : normalizing the field would generally change the target. Let denote parallel transport along from time to time zero. Its inverse transports vectors back to . We represent accumulated curve error in .
Smoothness norms use a fixed finite atlas with compact coordinate domains; for noninteger orders they include the corresponding Hölder seminorm. The known geometry determines fixed bounds for coordinate maps and their derivatives.
Assumption 2.1. The density and field are , and is . The density is bounded away from zero, and is uniformly positive definite.
Assumption 2.2. The nonnegative radial function is , supported in the open unit ball, and satisfies , , .
Assumption 2.3. The solution is an embedding and . There is a constant , common to all laws in the model class, such that
Assumption 2.4. For some , the admissible laws have fixed known bounds , , . Conditional on the pilot and all recorded positions, the final response errors are mutually independent.
The density, smoothness and kernel assumptions are familiar from nonparametric regression on manifolds. A design density bounded away from zero gives each neighborhood positive sampling probability, and a smooth regression function controls the bias of local averages [22]. Cheng and Wu [9] likewise use a compactly supported radial kernel and allow the conditional response variance to vary with location. Here radial symmetry cancels the first-order bias, compact support keeps parallel transport within normal neighborhoods, and the sub-Gaussian bound controls fluctuations of the local averages.
Recovering an integral curve also requires controlling how velocity errors accumulate along it. For passive inference, the additional smoothness in Assumption 2.1 controls derivatives of the fitted field, and the covariance lower bound ensures nondegenerate uncertainty. Assumption 2.3 limits the time the curve spends in any kernel neighborhood, as in the Euclidean integral curve model of Koltchinskii et al. (2007, Theorem 1). It excludes stationary points and returns over the observation interval. For active sampling, the known bounds in Assumption 2.4 determine the sampling radii and fit thresholds. Active inference in Section 4.3 additionally requires a lower bound on the response covariance.
Notation. Let denote the common intrinsic bound on and the diameter of . For curves and , write .
Passive sampling for Q1
We first consider observations collected at independent random locations. Intrinsic kernel regression estimates the vector field, and integrating the fitted field gives the curve estimate. We derive its limiting error process and use that law to construct simultaneous confidence tubes.
Curve estimation
The estimator averages nearby velocity measurements after parallel transport to a common tangent space, then integrates the fitted field from . The weights correct for Riemannian volume and stabilize small denominators. In normal coordinates at , write . For a radial kernel as in Assumption 2.2 and smaller than the injectivity radius, fix and a smooth function satisfying for and , and define
The replacement of by is a smooth stabilization: it prevents division by a density estimate close to zero, while leaving the ratio unchanged wherever . Uniform consistency of and make this replacement inactive with probability tending to one. Volume correction makes the change of variables exact and removes curvature from the second-order scalar integration formula; it is standard in intrinsic kernel estimation [21, 22].
The plug-in estimate of (2) solves
For the passive estimator we use Assumptions 2.1–2.3 and the following bandwidth condition.
Assumption 3.1. As ,
The first two conditions ensure that both and its spatial derivative converge uniformly to and , respectively. The derivative describes how velocity changes with position, and hence how an error in the curve affects its subsequent motion. The last condition controls the accumulated smoothing bias; confidence tubes use . The same balance appears in Koltchinskii et al. [16], together with . Our logarithmic strengthening supports uniform control of the fitted field and its derivative. The exponent five in (7) suffices to control the derivatives of the kernel numerator and denominator, the interpolation error within coordinate charts, and the remainder along the trajectory.
Curve error and simultaneous inference
We first identify the distribution of the error accumulated along the estimated curve. After scaling and parallel transport to a common tangent space, this error converges to a Gaussian process satisfying a linear stochastic differential equation. This functional limit provides the basis for the simultaneous inference developed below.
Put and . Let solve with ; it propagates an error made at time to time . When the estimated curve stays in a common normal tube around , define . The logarithm represents the geodesic displacement from the true curve, and puts these errors in one tangent space so that they can be compared across time. Outside this tube the short logarithm need not be uniquely defined, so we set to define the process for every sample. This exceptional event has probability tending to zero and does not affect the limiting distribution.
Define the connection-Laplacian bias field
The sign convention is , where is the trace of the second covariant derivative with respect to the Riemannian metric .
For , put
The integral is invariant under orthonormal changes of frame and therefore also defines for nonzero tangent vectors.
Theorem 3.2 (Integral-curve functional central limit theorem). Under Assumptions 2.1–2.3 and 3.1, in , where, for standard Brownian motion in , and
Equivalently, with the transition matrix defined above,
Its mean and covariance are the unique solutions of
For , .
The drift propagates earlier errors through the dynamics, while the diffusion term adds the local contribution of response noise. In a flat chart, , and is the ordinary derivative. Unlike the known-density numerator estimator in Koltchinskii et al. [16], the normalized ratio here removes the leading design fluctuation proportional to ; its diffusion coefficient therefore involves the response covariance . The theorem also includes the leading smoothing bias when . For confidence tubes we undersmooth so that and the limiting process is centered. Since on the normal tube, a quantile of , divided by , describes the radius needed to cover the entire curve. This connects the curve error to the radius of a simultaneous tube.
We now choose the bandwidth for the second-order estimator of § 3. Its curve error combines a bias of order and a random term of order . The next result gives the balanced estimation rate. All probability statements in this section are for a fixed model satisfying the passive assumptions.
Corollary 3.3. Under Assumptions 2.1–2.3, let . If , then
Balancing and gives . This minimizes the upper bound on curve error; the error can be smaller when the leading bias vanishes. For centered inference, choose a smaller bandwidth satisfying both and . The latter condition prevents undersmoothing so far that derivative noise invalidates the curve linearization. The logarithmic power is sufficient for our uniform bounds.
The curve rate is faster than the pointwise field rate for second-order kernel regression [28]. Thus estimating the trajectory can be more accurate than estimating the field at an individual location. The curve rate agrees with the Euclidean result of Sakhanenko [25], which establishes minimax bounds under different regression assumptions. Corollary 3.3 gives a probability bound for each fixed manifold model.
The limiting SDE explains which error paths must be reproduced, but its coefficients depend on the unknown field, density and response covariance. We construct a computable Gaussian approximation by propagating observation-level perturbations through the fitted dynamics. At the chosen bandwidth, one observation can affect several nearby points on the curve. Using the same perturbation throughout the trajectory preserves this dependence. Under undersmoothing, the resulting scaled process approximates the curve error and has the same limit as the SDE in Theorem 3.2.
We first define the Gaussian law whose supremum quantile will determine the tube radius. Its covariance includes each observation’s contribution at bandwidth . Put and . For a possible observation define its propagated contribution
It has mean zero. Let be the centered continuous Gaussian process on with covariance
Both the random positions and the conditional response noise contribute to this covariance. The kernel overlap at two different times is kept inside the expectation. The bounded-Lipschitz distance below uses the supremum norm on .
Theorem 3.4. Under Assumptions 2.1–2.3 and 3.1 and , the following uniform
expansion and Gaussian approximation hold:
Moreover, .
Theorem 3.4 approximates the distribution of the entire path and its maximum. The Gaussian law depends on , and the approximation is asymptotic as grows. Its limiting covariance is that of the centered SDE in (10).
To approximate the curve error, we perturb the observation weights and responses by independent Gaussian variables and propagate these perturbations through the fitted dynamics. Each observation uses the same perturbation throughout time, preserving the dependence between overlapping kernel neighborhoods. This reproduces entire error paths, so their maximum displacement can determine a simultaneous tube radius.
We estimate the auxiliary mean and response covariance without using an observation in its own fit. Choose a fixed number of outer folds , independently of the observations, and partition each complement into inner validation sets . All training and validation sizes are at least for a fixed . Let denote (3.3) fitted to sample , with bandwidth and normalization . For and , put
Pool the inner residuals to estimate their covariance:
For , let be with its eigenvalues truncated below at . The auxiliary bandwidths satisfy
The perturbation weights account for the stabilized denominator. Put
and define the position and response perturbations by
Here , so the position perturbations are centered. Let transport along back to , and set . Draw independent and standard Gaussian vectors , independently of the data. Keeping these draws fixed across time, solve
This ordinary differential equation propagates both sources of variation through the fitted curve dynamics. Its solution is in unscaled curve-error units; write for its covariance conditional on the data. Let be the conditional -quantile of . The simultaneous tube is .
The following theorem connects the simulation to the curve error: the scaled covariance consistently approximates , and the conditional law of approximates the same Gaussian law as . The supremum quantile therefore gives simultaneous coverage. Estimating only would miss the cross-time dependence that determines a maximum over the trajectory. The supremum over pairs below controls that dependence even though propagation is evaluated on the same random curve that is being estimated.
Theorem 3.5. Under the conditions of Theorem 3.4 and (20), the preceding estimators satisfy
The conditional distribution function of approximates that of uniformly over thresholds, in probability. Consequently,
Thus the CLT identifies the error dynamics, and the finite-bandwidth Gaussian approximation allows quantiles of the maximum curve error to be estimated. Taking the maximum of each simulated error path accounts for uncertainty over the whole time interval and gives one radius for the simultaneous tube.
Algorithm 1 Passive curve estimation and simultaneous tube
Require: Data, , , stabilization , level , and bandwidths satisfying (7) with and (20).
Fit (6) and integrate from to obtain .
Simulate (21), keeping each Gaussian draw fixed across time; record .
Take its -quantile ; return and .
A numerical implementation approximates the supremum on an evaluation grid and the quantile with finitely many Gaussian draws.
Active sampling for Q2
When measurement locations can be controlled, new responses can be concentrated near the route of interest. The route is unknown, so an initial passive sample first provides a pilot curve. We then use that curve to choose sampling neighborhoods and distribute the remaining measurements throughout them. The neighborhoods allow the final curve to depart from the pilot while keeping the new observations relevant to its reconstruction. The total response budget is : an independent pilot uses observations, and the final batch uses new responses. All acquired responses count toward the budget, including those in rejected fits.
The design
The true route is unknown, so the pilot indicates where new measurements will be informative. Its error must be small relative to the sampling neighborhoods; otherwise those measurements may miss the true route. We require the pilot to locate the true curve to within , where is the time-segment length and the reconstruction neighborhoods have radius proportional to . Fit (3.3) to the first observations with bandwidth , and let be its integral curve from . Use a kernel as in Assumption 2.2 that is positive near zero, take below the common normal radius, and fix .
For with fixed , put
Every fixed moment of is bounded by the corresponding power of , plus an exponentially small term in . Thus the budget and bandwidth can make the pilot accurate relative to ; the final fit need not reproduce the pilot’s estimation error.
The pilot determines the sampling neighborhoods and the allocation of the remaining observations. Within these neighborhoods, we collect new location-response pairs under the random sampling strategy below.
Dividing the time interval into short segments lets us reconstruct each part of the route from measurements in a small neighborhood. Allocating observations to every segment supplies information along the whole curve. Let be a positive integer, , , and , closing the last cell at . For each cell, center a sampling neighborhood at and use orthonormal coordinates on . We acquire responses at positions of the form with bounded scaled coordinates , where is fixed. Thus the pilot specifies the regions to which a mobile sensor is sent or a scanning instrument is directed. For sufficiently small , all these positions lie inside their designated charts by construction. The pilot determines the coordinate charts and the conditional sampling laws within them.
Choose a fixed reconstruction radius in scaled coordinates large enough to cover the movement during one segment and the pilot’s location error. On , each true segment lies inside its reconstruction neighborhood. Allocate the remaining responses so that the cell counts satisfy and uniformly over cells, with comparison constants independent of and . This allows nonuniform allocation, including proportions chosen from the pilot, while giving each segment a sample size of the same order. Each observation retains its acquisition cell label even when geometric neighborhoods overlap.
Within each cell, sample positions independently along the pilot chord and in the directions transverse to it. Measurements only on the pilot can leave transverse field changes undetermined. Spreading observations throughout the neighborhood identifies these changes, allowing the reconstructed curve to depart from the pilot. We fit a low-degree polynomial to the local coordinate velocities, map the fitted velocities back to the manifold, and integrate them to reconstruct the curve.
Write for the pilot, final cell labels and recorded positions. The observed pairs satisfy . Conditional on , the errors are independent, centered and uniformly sub-Gaussian. In the coordinates of the assigned cell, the observation has position and velocity
Here denotes the differential of at . This transformation expresses each response as the rate of change of its local coordinates, so the regression estimates the velocity needed to reconstruct a trajectory in that chart. With , fit a degree- polynomial by least squares:
Here is a multi-index and each coefficient is a vector in .
Each segment needs comparable sample counts, bounded scaled positions and a well-conditioned polynomial Gram matrix. These ensure that the local approximation is accurate and that the fit distinguishes its polynomial terms without amplifying noise. Let have entries for . For fixed positive constants , the required properties are
for every cell, with all positions in their designated charts. Denote the event in (26) by . The Gram lower bound ensures that (25) has a unique solution and controls how observation noise affects the fitted polynomial.
Map the fitted velocities back to the manifold to obtain a local field . Within the reconstruction neighborhood ,
Starting from , integrate the first field and use its endpoint to initialize the next segment:
The field is evaluated at the current estimated position, allowing the reconstructed curve to move away from the pilot.
To continue the flow outside the reconstruction neighborhood, choose a smooth function equal to one on and zero beyond , and put . Let be a smooth cutoff supported inside the chart and equal to one on . Extend the fitted field by
and set it to zero outside the chart. This agrees with the original fit on and gives a field defined along the entire flow.
We also check the fitted coefficients to control the field’s magnitude and first derivative. For a coefficient matrix , define
The fixed model and design bounds give constants and bounding the constant coefficient and of the conditional mean fit whenever the pilot localizes the true curve and the design satisfies (26). Accept the response fits when, for every cell,
These bounds keep every accepted extended field uniformly bounded together with its first derivative. The factor compensates for the derivative of the scaled coordinate .
Choose the count thresholds in (26) to contain the allocated counts. Fix and so that whenever holds, the eigenvalues of each conditional population Gram matrix lie inside by a common positive margin. These bounds hold for sampling uniformly along and across each pilot chord. Algorithm 2 summarizes the procedure.
Convergence rate
The next theorem bounds the expected maximum distance between the estimated and true curves over . We use this bound to choose the pilot bandwidth and segment length so that the error is , where counts observations from both stages.
Theorem 4.1. Use the random neighborhood design and reconstruction of § 4.1 with recorded sampling positions. Under Assumptions 2.3 and 2.4, for every fixed , the neighborhood estimator satisfies, for sufficiently small admissible and ,
The constants depend on the fixed geometry, field class, noise bound, reconstruction radius and Gram bounds.
Remark 4.2. For the random neighborhood sampling in § 4.1, the design failure probability satisfies
under the stated count and Gram bounds. This gives the first exponential term in the last part of (30).
The two-stage estimator, using a kernel pilot with , , and attains expected risk in dimensions two and three for , under the random sampling strategy above. In higher dimensions, the same nonnegative kernel can be used with a fixed pilot fraction and . Choose with
This choice yields expected risk whenever . The sufficient smoothness threshold grows with dimension, with degree used in the final local fits; the pilot remains a second-order kernel estimate. The allocation calculation is given in the supplementary material.
Simultaneous confidence tubes
Section 4.2 bounds the expected maximum distance between the estimated and true curves. A tube around the estimated curve covers the true position at every time in precisely when this maximum is at most its radius. We therefore estimate the quantile of the maximum distance to choose the radius.
An error in the velocity fitted on one segment can shift the starting position of the next, so errors at different times are dependent. We use the residuals from the second stage velocity fits to simulate this propagation. In each draw, we give each residual an independent Gaussian weight and follow its effect along the fitted curve, keeping the weight fixed over time. We record the largest displacement along each simulated path and use the resulting quantile as the tube radius.
Order the polynomial coefficients with the constant term first and write , using columnwise vectorization. Set
The influence matrix maps residual to its propagated displacement at time . To compute it, let transport back along , identify with using a fixed orthonormal basis, and solve
For , put . The coefficient derivative of (4.4) is
and is zero outside the chart. Thus
This combines the change in the local least-squares fit with its propagation through the fitted curve dynamics. With coefficients per response component, require in each accepted cell and use the degrees-of-freedom correction, for
The correction accounts for the number of fitted coefficients in each cell without changing the fitted field or curve. There is one scalar multiplier per vector response, shared at all times; its law is conditional on all observed data. For , let be the exact conditional quantile of . Report the continuous-time tube
If a chart, design or response fit is rejected, report no tube and count the attempt as noncoverage.
To construct confidence tubes, we require the local approximation bias to be negligible relative to , and the noise variance to remain bounded away from zero in every tangent direction. The following conditions also specify the required pilot accuracy and sample sizes.
Assumption 4.3. For a fixed , every final error satisfies . As , assume , , and
The condition makes approximation bias negligible on the curve-error scale, following the undersmoothing principle for confidence bands in Chernozhukov et al. [6]. The covariance lower bound provides nondegeneracy for Gaussian supremum calibration, whose positive-variance requirement is explicit in Chernozhukov et al. [7]. The remaining rates control pilot localization and local-fit stability in this sampling design.
To describe the approximation, replace the fitted field and curve in the influence matrices by the true field and curve, giving . Let be the response error in the local coordinates and write . The leading error and its conditional covariance are and . The covariance determines the possible Gaussian limits of the normalized error. The following result states these limits and the coverage of the residual-based tube. For formal probability statements, give the theoretical comparison objects zero values on . Set if the short logarithm between the true and estimated curves is not available throughout . On an observed chart, design or response-fit failure, give undefined feasible covariances, multiplier processes and critical values zero values and report no tube. Conditioning on always precedes response-fit acceptance.
Theorem 4.4 (Active Gaussian limits and simultaneous inference). For the reconstruction in § 4.1, under Assumptions 2.3, 2.4 and 4.3, put . For every deterministic sequence of admissible recorded histories , every subsequence admits a further subsequence along which uniformly on and, in
Here is a continuous centered Gaussian process with covariance . The same convergence holds along any subsequence on which the normalized covariance converges uniformly.
For the tube in (33), let denote successful reporting and coverage of at every . Under the random sampling design of § 4.1, . Thus every conditional weak limit of the normalized curve error is Gaussian. Its covariance can depend on the recorded designs and the subsequence; convergence of along the full sequence gives a full sequence limit. These are conditional limits: averaging over random designs need not give a Gaussian law. The feasible tube does not require a unique limiting covariance, because the residuals consistently estimate the covariance for the current design. Its radius has order in probability.
The coverage conclusion is uniform over model and design families with the stated common bounds. The pilot determines where to sample, while the final responses estimate the curve and its uncertainty. The method assumes that actual measurement positions are recorded; position error and physical deployment costs are separate modeling questions.
Algorithm 2 Active sampling and simultaneous tube
Require: , , , , , , , , , reconstruction and acceptance constants, and level .
Fit (6) to the pilot observations with bandwidth , and integrate from to obtain .
Construct pilot charts; collect location–response pairs per cell, sampling along and transverse to each pilot chord. Record cell labels and check chart admissibility and (26).
Extend by (28), integrate (27), and compute residuals and (31).
Simulate the Gaussian process (32) with one Gaussian multiplier per response fixed across time. Return and (33) using the simulated supremum quantile.
On a chart, design, fit or integration failure, return the prescribed constant curve and no tube; count the attempt as noncoverage and count all allocated responses in the budget. No step tests the unobservable event .
Numerical results
We compare passive kernel regression with pilot-guided neighborhood sampling on , and . The response budgets are 200, 400, 600, 800, 1000 and 2000, with 500 independent evaluation datasets per setting and horizon . Evaluation data are separate from development data. The two methods use different sampling and calibration rules, so these are comparisons at a common response count rather than on identical observations. Additional torus and projective experiments are reported in Appendices A.4–A.5.
For both passive and active methods, the kernel function is
where normalizes the kernel to integrate to one. The final active fits use quadratic local polynomials. In the active simulations the pilot uses ordinary kernel weights, giving
Here the normalizing constant cancels. The implementation reports failure if the sum of the unnormalized weights is below .
Coverage is the proportion of all 500 attempts whose tube contains the true curve at every evaluation node; a failed fit counts as noncoverage. Bias in the tables is the average, over the 500 repetitions, of the maximum intrinsic error norm on the evaluation grid ,
including the prescribed constant-curve output after failure. Radius is the average tube radius over successfully constructed tubes. These are finite-grid coverage summaries.
Figure 2 and Figure 3 illustrate one evaluation dataset for each method at budget .
Figure 2. Passive sampling on . Left panel: noisy vector observations and the true curve. Right panel: the estimated curve and simultaneous confidence band, with the true curve for comparison.
Figure 3. Active sampling on . Left panel: noisy vector observations and the true curve. Right panel: the estimated curve and simultaneous confidence band, with the true curve for comparison.
Both use an oblique Mercator chart centred at longitude 1.1 and latitude 0.4 radians, with the map rotated counterclockwise.
Passive kernel method
The passive experiments implement Algorithm 1, with the uncorrected curve as the tube center. The fields include rotation with changing latitude on , rotation at unequal angular speeds on , and position-dependent angular velocity on .
Bandwidths were fixed before evaluation. All fits completed, so each mean radius uses 500 tubes. We compute the curve, parallel frame and covariance factor using a Runge–Kutta solver with continuous interpolation. Two solves with successively tighter tolerances check numerical stability; the second supplies the curve and factor used for Gaussian simulation. All 15000 passive fits, including the additional manifolds in Appendix § 9.§ 4, passed these checks. Appendix § 9.§ 2 gives the numerical settings.
On the unit sphere , the density relative to Riemannian volume is , and the field combines rotation with changing latitude, . The initial point is ; the reference curve is computed by numerical integration of this field. Responses have independent tangent noise , . The main bandwidth is at all six sample sizes.
On the unit sphere , use
The initial point is , and independent tangent noise is with . For all three-dimensional passive experiments, is the larger of and the bandwidth giving 12 expected kernel neighbors at the minimum design density; the support radius is .
On with the rotation-angle metric, use density and vector field with left-body angular velocity , starting at the identity.
Independent left-body noise has coordinate standard deviations .
Distances and radii are physical rotation angles in radians. The main bandwidth follows the three-dimensional rule above.
| Coverage (%) | Bias | Radius | Coverage (%) | Bias | Radius | Coverage (%) | Bias | Radius | |||
| 200 | 95.2 | 0.156 | 0.390 | 89.6 | 0.103 | 0.158 | 87.8 | 0.127 | 0.188 | ||
| 400 | 96.0 | 0.098 | 0.212 | 92.6 | 0.086 | 0.140 | 93.0 | 0.103 | 0.166 | ||
| 600 | 96.6 | 0.080 | 0.166 | 93.8 | 0.078 | 0.131 | 93.0 | 0.093 | 0.157 | ||
| 800 | 96.6 | 0.069 | 0.133 | 96.8 | 0.070 | 0.125 | 94.4 | 0.087 | 0.149 | ||
| 1000 | 95.6 | 0.064 | 0.122 | 96.4 | 0.067 | 0.121 | 93.6 | 0.086 | 0.145 | ||
| 2000 | 94.8 | 0.046 | 0.086 | 96.0 | 0.057 | 0.100 | 95.6 | 0.072 | 0.132 |
Table 1. Passive sampling on three manifolds; is the sample size.
Active neighborhood sampling
On all three manifolds, Algorithm 2 uses the same field, starting point and noise law as the corresponding passive setting. The density describes the independent pilot locations.
We allocate responses to the pilot and to the final batch. The pilot bandwidth is , with , and on , and , respectively. Final observations are allocated as evenly as possible across time cells and sampled uniformly along the pilot chords, with independent transverse coordinates. All acquired responses count toward . On , use , with independent uniform transverse coordinates of half-widths . Calibration uses the 951st ordered maximum of 1000 Gaussian paths on , with each residual multiplier shared across time. Confidence tubes use the degrees-of-freedom correction in (4.8).
Across these three manifolds and six common budgets, the active mean radius is smaller in most scenarios. The exception is at budget , where active and passive radii are and . At budget on , active and passive coverage are and , with radii 0.097 and 0.140. On at budget 200, active and passive coverage are 94.0% and 95.2%, with mean radii 0.139 and 0.390. These comparisons evaluate both sampling and inference: the sampling rules, quantiles and evaluation grids differ. The effect of sampling alone therefore cannot be separated from the other differences.
Real-data applications
Regional ocean circulation on the sphere
How much of a regional mean flow trajectory can existing ocean-drifter observations support?
We study summer circulation in the North Atlantic box degrees longitude by degrees latitude. The Global Drifter Program hourly product Elipot et al. [10] supplies 11977 eligible daily records from 221 buoys during June–August, 2010–2019. Entire buoys define training, validation and test sets of sizes 142, 33 and 46. The target is the cohort’s average summer velocity, giving each buoy equal total weight, and its integral curve. Eastward and northward velocities are tangent vectors on : the local constant estimator in (6), with these buoy weights, uses parallel transport to average them, and the fitted curve is integrated on the sphere.
Spatial support limits the length of the reconstructed route. Figure 4 shows the training positions, the supported domain at bandwidth 400 kilometres and the selected starting point. The quarter-degree shading displays spatial support, which is checked directly along the curve.
Figure 4. North Atlantic drifters and spatial support. Daily training locations, the shaded supported domain and the selected starting point show where the mean flow route can be reconstructed.
| SO(3) | |||||||||
| Coverage (%) | Bias | Radius | Coverage (%) | Bias | Radius | Coverage (%) | Bias | Radius | |
| 200 | 94.0 | 0.078 | 0.139 | 97.2 | 0.118 | 0.237 | 97.6 | 0.091 | 0.172 |
| 400 | 95.2 | 0.048 | 0.083 | 97.6 | 0.054 | 0.097 | 95.8 | 0.050 | 0.088 |
| 600 | 95.2 | 0.032 | 0.060 | 96.8 | 0.061 | 0.115 | 97.8 | 0.053 | 0.098 |
| 800 | 94.0 | 0.028 | 0.050 | 97.0 | 0.044 | 0.080 | 96.2 | 0.039 | 0.070 |
| 1000 | 94.2 | 0.025 | 0.045 | 97.2 | 0.036 | 0.064 | 96.0 | 0.033 | 0.058 |
| 2000 | 93.8 | 0.017 | 0.031 | 96.0 | 0.027 | 0.047 | 97.4 | 0.023 | 0.042 |
Table 2. Active sampling; is the total sample budget.
We restrict reconstruction to a training-defined domain satisfying minimum distinct-buoy and effective-sample-size requirements in every training-fold omission. The original starting position (, ) degrees is unsupported; its nearest admissible quarter-degree grid point is (, ) degrees. From this point, the estimated route remains supported for 3.79 days, short of the intended five-day horizon. This illustrates the trade-off between spatial resolution and observational support in drifter climatologies [17]. Appendix § 10 gives the start selection, support criteria, inclusion rules and temporal sensitivity analysis.
Within , the fitted field reduces test velocity loss by 6.6% relative to a transported constant-vector baseline, using 615 records from 28 buoys. This improvement translates only weakly into prediction of individual drifter motion: endpoint errors are 24.40 versus 24.60 kilometres at one day and 127.91 versus 132.74 kilometres at five days. These comparisons involve only 17 and 12 completed test windows, respectively, each from four buoys; validation results show no consistent advantage. Appendix § 10 reports the errors and completed windows. The curve describes regional mean flow, while individual motion also contains temporal and eddy variability [20, 17]. Dependence between observations precludes interpreting the present analysis as a validation of nominal confidence-tube coverage.
Mean flow through a model canopy
How does the local mean flow route beside a canopy element change between positions within and above the canopy? Spatially averaged wind directions can conceal departures around individual elements. Using three dimensional velocimetry, Harman et al. [11] found strong spatial variation in the mean flow within a model canopy, diminishing rapidly above it. We use active reconstruction to express this heterogeneity in terms of the lateral and vertical displacement of a finite mean streamline.
The public CSIRO Detailed Spatial Sampling archive ([12]) contains three-component laser Doppler anemometry (LDA) measurements at 1540 positions around a canopy of 60 mm-high elements. Each location was sampled for 35 seconds, and the released velocities are residence-time-weighted means. In a corridor beside one element, 169 of 504 positions supply reference reconstructions and 335 form the acquisition pool. The two mean streamlines start at the same horizontal location and heights 42 and 84 mm; horizons of 10 and 5 milliseconds give similar downstream advances of about 26 mm. The target is an integral curve of the three-dimensional mean velocity, not an individual turbulent particle path.
Pointwise LDA permits measurement locations to be chosen, and adaptive acquisition has been demonstrated experimentally ([30]). The present archive was collected on a prescribed grid, so we implement active selection retrospectively by revealing actual measurements at selected locations. A quarter of the total budget supplies a kernel pilot, which guides one further batch of observations to neighborhoods along the estimated route. The final reconstruction uses the unweighted local least-squares fits in (25) and their integral curves. The main display uses , including 40 pilot observations; Appendix § 11 gives the joint allocation rule and results for , 160, 240, with 500 sampling repetitions at each height and budget.
Figure 5 shows horizontal projections of the mean of 500 three dimensional reconstructions, with measured horizontal velocities on a common arrow scale. The paths may change height. At , mean outward displacement is 4.22 mm within the canopy and 0.27 mm above it; downward displacements are 2.17 and 0.38 mm. Thus the within-canopy route bends outward and downward, whereas the route above it remains nearly streamwise. The standard deviations across sampling designs are 0.19 and 0.06 mm for outward displacement, and 0.15 and 0.05 mm for downward displacement. These summarize variation in selected locations conditional on the archive. The height contrast also appears directly in the measurements: the median absolute lateral velocity in the corresponding local strips is 0.367 metres per second within the canopy and 0.039 metres per second above it. Appendix § 11 gives the full results.
Figure 5. Local mean flow reconstruction within and above a model canopy. Arrows show measured horizontal velocities, and curves show horizontal projections of the mean of 500 reconstructions.
The greater within-canopy deflection gives a spatial interpretation of the heterogeneity reported by Harman et al. [11]. The contrast persists across all three budgets and reference reconstruction orders. For the reserved observations, the quadratic reference gives outward displacements of 3.72 and 0.31 mm, whereas a linear reference gives 1.86 and 0.19 mm. The magnitude depends on reconstruction order. The selected locations vary across repetitions, conditional on the archive; shared calibration errors and unavailable raw time series prevent validation of nominal confidence-tube coverage.
Discussion
Sampling design changes how accurately a fixed response budget can recover an integral curve. Passive observations support reconstruction and simultaneous inference; directing measurements toward a pilot can improve accuracy under the stated localization and smoothness conditions. The applications show how these curves describe local mean transport and how observational support limits their reconstruction.
The applications suggest two extensions. First, repeated measurements along a drifter track contain temporal dependence, and seasonal or eddy variation can separate an individual trajectory from a mean flow route [20, 17]. For dependent responses, propagating the field error would also propagate cross-covariances between observations; the present independent Gaussian multipliers would need to be replaced by a construction that preserves that dependence. A time-varying field would further require local information in time as well as space, reducing the observations available near each part of the route. These issues explain why improved mean-velocity prediction alone does not establish coverage for individual drifter paths. Second, the active method treats the response count as the sampling budget. Adaptive LDA and autonomous underwater sampling show that directing measurements is practical [30, 14], but moving between locations also consumes time. A design that balances localization against travel cost would need to retain enough spread across the route to identify transverse changes in the field. The canopy analysis shows a related constraint even in a fixed archive: increasing the quota can force observations farther from the pilot. Accounting for these costs and spatial constraints would connect the statistical gain from active sampling more directly to measurement effort.
Acknowledgment
Qirui Hu’s research was supported by National Natural Science Foundation of China (NSFC) (Grant Nos.12601520), the Shanghai Engineering Research Center of Finance Intelligence (Grant No. 19DZ2254600) and by TRR 391 Spatio-temporal Statistics for the Transition of Energy and Transport (Project number 520388526) funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation).
Numerical implementation details
Passive calibration: residual pilots and forcing columns
This section specifies the auxiliary estimates used in § 3.2. Their consistency is established in the supplementary proof of Theorem 3.5. Choose outer folds by fixed observation indices or by random labels independent of all observations. Put and partition each into inner validation sets in the same way. The numbers of folds are fixed and all training and validation sizes are at least for a fixed . For an index set , let be (6) with sample , denominator normalized by , and bandwidth . For and , set
Pool the inner residuals over and define
Let be with each eigenvalue replaced by its maximum with . Fold-size-dependent thresholds tending to zero are also allowed, as shown in the supplementary proof of Theorem 3.5. The bandwidths obey (20). No derivative of the mean pilot is required, because propagation uses the derivative of .
The following weights account for the smooth denominator stabilization. At a query point , put
and define
Thus exactly. Where stabilization is inactive, and . The design and noise columns follow from perturbing observation weights and responses, respectively, as verified below.
Let be parallel transport along to , and set . Draw mutually independent and standard Gaussian , independently of the data. Use the same weights at every time and solve the ordinary differential equation
Equation (A.4) has the same linear propagation structure as the limiting SDE, with an estimated drift and Gaussian forcing that retains the kernel overlap at bandwidth . Conditional on the Gaussian draws, it is an ordinary differential equation. The process is in unscaled curve-error units, so its supremum quantile can be used directly as a geodesic tube radius.
In orthonormal frames at the observations, concatenate the forcing columns into and solve , . Then , with one standard Gaussian vector , and its entire conditional covariance kernel is
Changing an observation frame leaves this law unchanged.
To obtain the design-forcing columns in (A.3), perturb the empirical observation weights by and differentiate the normalized kernel ratio with responses . Differentiating its numerator and denominator gives precisely . Independent Gaussian response perturbations with covariances give the columns .
Passive integration and refinement
The experiments use five outer folds for conditional means and five inner folds for residual covariances. On , auxiliary means use local linear fits and covariances use smoothed residual squares. On and , the auxiliary bandwidths are and , respectively. Gaussian paths follow (A.4), with independent design and response-noise weights across observations, shared across time. We simulate 2000 paths on .
The numerical settings aim to keep integration error small relative to the statistical curve-error scale . Using the Runge-Kutta solver described in § 5.1, we limit every curve, frame, factor and numerical-reference step to . Coefficients are evaluated at the integration times, retaining all nonzero kernel contributions.
Stability is checked by comparing solutions at two tolerance levels.
Active sampling and reconstruction
Partition into equal time cells and allocate the final observations as evenly as possible. On and , use at the six increasing budgets. Final positions are uniform along each pilot chord, with independent transverse coordinates uniform on . On , use , with independent uniform transverse coordinates of half-widths .
For and , a position-only rule selects the first admissible reconstruction window among four ordered pairs of longitudinal margin and transverse half-width: , , and , each scaled by . An observation retains its acquisition cell label and enters its assigned fit only when its recorded position lies in that cell’s selected window. Excluded responses still count toward .
Each cell is checked for sufficient sample size, Gram conditioning and leverage; chart and fitted-field checks follow Algorithm 2.
The active calibration uses the 951st ordered maximum among 1000 Gaussian paths on , with one residual multiplier per response shared across time.
Additional passive experiments
The torus experiments use the sample sizes, repetition count, kernel, quantile rule and evaluation grid of Appendix § 9.§ 2. All fits completed. Distances are shortest wrapped displacements in radians.
On , use
The true curve has lift . Noise is , , independently of locations. The bandwidths are , and . The constant in was selected from by held-out prediction error on 24 independent development datasets and fixed before validation.
On , use
Independent noise coordinates have standard deviations , 0.15 and 0.25. The main bandwidth follows the three-dimensional rule in § 5.1; .
The three-dimensional support floor is and eventually smaller than the logarithmic bandwidth, giving . The two-dimensional choices give , with on and on , verifying the asymptotic undersmoothing condition.
The result at combines high coverage with 59 local-support warnings. The smaller-budget undercoverage on and is discussed with the main results.
| Coverage (%) | Bias | Radius | Coverage (%) | Bias | Radius | ||
| 200 | 98.6 | 0.070 | 0.162 | 95.2 | 0.106 | 0.175 | |
| 400 | 96.6 | 0.057 | 0.120 | 94.4 | 0.096 | 0.163 | |
| 600 | 95.0 | 0.051 | 0.103 | 95.4 | 0.087 | 0.156 | |
| 800 | 96.6 | 0.044 | 0.091 | 97.4 | 0.081 | 0.151 | |
| 1000 | 96.4 | 0.040 | 0.083 | 95.0 | 0.081 | 0.148 | |
| 2000 | 96.6 | 0.032 | 0.065 | 96.0 | 0.075 | 0.138 |
Table 3. Passive sampling on two tori; is the sample size.
Additional active experiments
The active torus and projective experiments follow the and sampling and reconstruction settings in Appendix A.3. The torus pilot densities, fields, starting points and noise laws are those in Appendix A.4. Their pilot bandwidth constants are for and for .
For , represent a point by on the unit sphere , with distance . Relative to Riemannian quotient volume, the pilot density is
The normalization follows from the uniform-sphere identity . The vector field has antipodally equivariant lift , starting at the initial point in Section 5.1. Tangent noise is isotropic Gaussian with standard deviation 0.2. The pilot bandwidth is , keeping kernel support inside the injectivity radius. The metric on differs from the physical rotation-angle metric on by a factor of two.
Each row has 500 successful tubes, with bias and radius as defined in Section 5. Across all five manifolds shared by the two methods, the active mean radius is smaller in
| Coverage (%) | Bias | Radius | Coverage (%) | Bias | Radius | Coverage (%) | Bias | Radius | |
| 200 | 95.6 | 0.035 | 0.064 | 96.2 | 0.068 | 0.129 | 95.8 | 0.079 | 0.147 |
| 400 | 94.2 | 0.024 | 0.043 | 97.0 | 0.042 | 0.074 | 97.2 | 0.052 | 0.093 |
| 600 | 96.4 | 0.020 | 0.036 | 96.8 | 0.041 | 0.075 | 96.6 | 0.057 | 0.104 |
| 800 | 93.6 | 0.018 | 0.031 | 97.4 | 0.032 | 0.058 | 96.6 | 0.041 | 0.075 |
| 1000 | 96.2 | 0.015 | 0.027 | 95.8 | 0.028 | 0.049 | 95.6 | 0.035 | 0.060 |
| 2000 | 95.2 | 0.011 | 0.019 | 96.6 | 0.021 | 0.037 | 96.0 | 0.025 | 0.044 |
Table 4. Active sampling on two tori and ; is the total sample budget.
29 of 30 settings; the exception is the setting identified in § 5.2. Across all six active manifolds, mean radii are smaller at than at and coverage ranges from 93.6% to 97.8%. Intermediate radii need not decrease monotonically because the number of time cells changes with the budget.
Regional drifter analysis: cohort, support and completion
Data and supported reconstruction
Drifter climatologies separate mean circulation from seasonal and eddy variation [20]. Irregular sampling and correlated observations make both spatial resolution and uncertainty sensitive to how observations are pooled [17]. We therefore split by buoy, weight buoys equally, and restrict reconstruction to regions supported by several distinct buoys.
We use the Global Drifter Program hourly location, velocity and temperature data.1 We retain 00:00 UTC records during June–August, 2010–2019 in the box of § 6.1, with valid positions and velocities, interpolation gaps of at most 12 hours, and documented positive drogue-loss dates later than the observation. Zero or missing dates are excluded.
There are 11977 records and 221 eligible buoys: 7556 records from 142 training buoys, 1813 from 33 validation buoys and 2608 from 46 test buoys. Buoys are assigned deterministically by identifier to training, validation and test sets; the training folds use a separate identifier hash.
For a fit containing eligible buoys, buoy with retained records receives per-record weight . Weights are recomputed for the eligible training subset. Recorded eastward and northward velocities in metres per second are represented in their corresponding unit tangent directions by
in radians per day on the unit sphere, using Earth radius 6371000 metres. GDP velocity uncertainty fields are not substituted for the conditional covariance of the regression experiment. The comparisons use transported local constant regression, transported local linear regression and a common-anchor transported constant-vector baseline. The local constant method is the primary reconstruction in § 6.1; the local linear fit checks sensitivity to a first-order spatial term.
Spatial support. None of the candidate bandwidths, 100, 200 and 400 kilometres, supports the original full-domain analysis in every training fold. Before evaluating held-out errors, we therefore fix the bandwidth at 400 kilometres and restrict reconstruction to the training-defined domain . Its compact kernel neighborhoods have radius 360 kilometres. Membership requires at least ten distinct buoys, row effective sample size at least 30 and buoy effective sample size at least ten, both in the full training fit and in every training-fold omission. At a query point, let include the fitting weight, volume correction and kernel weight, and let be the sum for buoy . The row and buoy information summaries are
A zero denominator is unsupported. These thresholds use locations and identifiers, without held-out responses.
The selected curve starting point is , obtained from the search described in § 6.1. The transported constant-vector baseline retains the original reference point : it averages responses transported there and transports that mean to each query point.
Integration stops on leaving , failing the one-bandwidth box margin or the land screen,2 or falling below speed 0.02 metres per second. Each fitted subset must also supply support within the original domain . Intrinsic midpoint integration is checked at a finer step size.
Temporal sensitivity and prediction assessment
We first examine how the reconstructed route changes when parts of the observation period are omitted, then assess velocity and displacement prediction on held-out buoys. The omissions probe the temporal composition of this pooled mean-flow estimate; seasonal and interannual components are treated explicitly in the climatology of Lumpkin and Johnson [20].
The primary local constant, local linear and baseline curves stop for insufficient support after 3.79, 3.97, 2.69 days. Each of ten leave-one-year-out fits and the two half-period fits retains the fixed bandwidth, anchor and original , while also requiring support from its own subset. All twelve subsets, including unsupported starts, are shown in Table 5 and Figure 6. Durations are in days; zero denotes an unsupported starting point.
| Training subset | Local constant | Local linear | Baseline |
| Omit 2010 | 1.99 | 1.91 | 2.41 |
| Omit 2011 | 3.83 | 3.98 | 2.63 |
| Omit 2012 | 2.49 | 2.56 | 2.89 |
| Omit 2013 | 3.79 | 3.97 | 2.32 |
| Omit 2014 | 3.20 | 3.26 | 3.78 |
| Omit 2015 | 5.00 | 5.00 | 2.78 |
| Omit 2016 | 5.00 | 5.00 | 2.64 |
| Omit 2017 | 3.12 | 3.23 | 2.55 |
| Omit 2018 | 0 | 0 | 0 |
| Omit 2019 | 3.96 | 4.18 | 2.76 |
| 2010–2014 only | 0 | 0 | 0 |
| 2015–2019 only | 0 | 0 | 0 |
Table 5. Passive drifter observations on : supported duration.
Figure 6. Supported regional mean-flow reconstruction on . (a) The three fitted prefixes in east/north normal coordinates at the starting point; markers denote complete days and crosses denote support stops. “Kernel” denotes transported local constant regression. (b) Supported durations after omitting each training year or retaining only 2010–2014 (A) or 2015–2019 (B), with the bandwidth, starting point and original support domain fixed. Zero denotes an unsupported start; five days denotes the intended horizon. Symbols identify the same estimators in both panels.
Held-out prediction. Within the supported region, both local regressions improve held-out velocity prediction over the constant-vector baseline (Table 6). Squared vector errors are averaged first within buoy and then across eligible buoys. Local constant regression has the lowest validation loss; on test data the two local losses are nearly equal and about 6.6%–6.7% below the baseline. This suggests that spatially varying mean velocity is useful for prediction in the sampled region, while adding a local linear term yields little further test improvement at this bandwidth. The result concerns 615 supported test records from 28 buoys, out of 2608 records from 46 test buoys.
| Split | Local constant | Local linear | Baseline | Supported records/buoys | All |
| Validation | 148.635 | 152.143 | 152.283 | 735/26 | 1813/33 |
| Test | 146.954 | 146.761 | 157.313 | 615/28 | 2608/46 |
Table 6. Passive drifter observations on : velocity prediction, with losses in .
For displacement prediction, a window consists of six consecutive daily records spanning five days. A greedy chronological rule accepts the next window from a buoy only when its start is at least five days after the preceding start. Time interiors are disjoint, although neighboring windows may share an endpoint. The 811 windows comprise 334 from 30 validation buoys and 477 from 43 test buoys. Of these, 131 validation windows from 24 buoys and 113 test windows from 25 buoys start in .
Completion requires reaching the horizon at both numerical step sizes, with endpoint discrepancies at most one metre. A common comparison requires all three methods to complete the same window. Only 17 test windows from four buoys qualify at one day, and 12 from four buoys at five days. Table 7 reports geodesic endpoint errors in kilometres, averaged first within buoy and then across buoys in each common completion subset. Gains over the baseline are smaller and less consistent than for velocity prediction. Table 8 records method-specific and common completion counts; Table 9 accounts for every window, including those excluded from the endpoint-error comparison.
| Split | Horizon | Windows/buoys | Local constant | Local linear | Baseline |
| Validation | 1 day | 14/4 | 38.05 | 38.39 | 36.51 |
| Validation | 5 days | 10/3 | 125.34 | 122.93 | 123.41 |
| Test | 1 day | 17/4 | 24.40 | 24.58 | 24.60 |
| Test | 5 days | 12/4 | 127.91 | 129.34 | 132.74 |
Table 7. Passive drifter observations on : displacement prediction.
| Split | Horizon | Local constant | Local linear | Baseline | Common |
| Validation | 1 day | 14/4 | 14/4 | 15/5 | 14/4 |
| Validation | 5 days | 10/3 | 12/4 | 14/4 | 10/3 |
| Test | 1 day | 17/4 | 17/4 | 18/4 | 17/4 |
| Test | 5 days | 13/4 | 13/4 | 15/4 | 12/4 |
Table 8. Passive drifter observations on : numerical completion.
| Split | Method | Box margin | Horizon | Support | Low speed | Unresolved crossing | Total |
| Validation | Local constant | 266 | 10 | 56 | 1 | 1 | 334 |
| Validation | Local linear | 266 | 12 | 55 | 1 | 0 | 334 |
| Validation | Baseline | 266 | 14 | 54 | 0 | 0 | 334 |
| Test | Local constant | 340 | 13 | 122 | 1 | 1 | 477 |
| Test | Local linear | 340 | 13 | 122 | 1 | 1 | 477 |
| Test | Baseline | 340 | 15 | 122 | 0 | 0 | 477 |
Table 9. Passive drifter observations on : stopping reasons.
Table 9 lists the five-day stopping reasons for all windows. Each row sums to all windows in its split, with categories defined by the algorithm’s stopping conditions.
Only 17/477 = 3.56% of all test windows enter the common one-day comparison and 12/477 = 2.52% the common five-day comparison. Both proportions use all test windows as the denominator, whereas method-specific completion and support-restricted proportions use different denominators. The endpoint errors therefore describe only the small subset on which every method completes.
Canopy LDA analysis: active design and sensitivity
Data and active reconstruction
The strong within-canopy spatial variation documented by Harman et al. [11] motivates reconstructing local routes at two heights, rather than summarizing the flow by one prevailing direction. The observations are version 1 of the CSIRO Detailed Spatial Sampling collection, doi:10.25919/xs2b-be57. The experiment and processing are described by Harman et al. [11] and the collection’s data overview. The archive provides three mean velocities and turbulent moments, but no raw time series or formal uncertainties for the means. Turbulent moments are not standard errors of those means. The overview identifies systematic and dependent error sources associated with the common instrument setup. Each measured position represents 35 seconds of acquisition, excluding repositioning and calibration.
Let , where is the source coordinate. We use , and mm, beside the central 10 mm-wide canopy element and away from its solid interior. The starts are and mm, with horizons and seconds. Recorded velocities in metres per second are multiplied by 1000 for integration in millimetres. All three velocity components are retained. Outward displacement is , downward displacement is , and downstream advance is , all in millimetres.
Source indices , and identify the eight downstream sections, ten lateral positions and heights in descending order. The condition reserves 169 corridor observations; the remaining 335 form the acquisition pool. Missing positions are not imputed. Reference observations enter neither the pilot nor the final budgeted fits or selection decisions. The subsets share experimental calibration and conditions.
Acquisition from the measured pool. Adaptive LDA can allocate measurements according to local curvature, statistical error and sample spacing [30]. Here the allocation targets neighborhoods along a pilot route in the existing archive. For each , a uniform sample of distinct pool positions supplies a kernel pilot. The smooth compact radial kernel is with radius 36 mm. Pilot positions at times 0 and define two anchors, and each segment receives final responses. The final locations are chosen using positions and pilot responses alone; their responses are revealed together after selection.
To allocate distinct points jointly, let and be the unqueried candidate sets within a common radius of the two anchors. Quotas of size are feasible when , and . We start at mm and, when necessary, choose the smallest radius satisfying these conditions among actual candidate distances. For candidate and segment , let be its distance in millimetres to the anchor. With independent standard Gumbel variables , we minimize the total assignment cost subject to the two quotas and no reused point. Thus overlapping neighborhoods are allocated simultaneously, with a preference for nearby positions. The spatial costs govern acquisition only.
For each segment, the ten monomials of total degree at most two are fitted by the unweighted criterion (25), separately for the three velocity components and using only that segment’s assigned final responses. Coordinates are centered at the anchor and scaled by 20 mm for numerical conditioning. This rescaling leaves the polynomial space and fitted field unchanged. In Euclidean space the exponential-map differential is the identity. The two fitted fields are integrated successively as in (27). Every pilot and final observation counts toward ; the nominal acquisition times are 46.7, 93.3 and 140 minutes at the three budgets.
The finite grid entails sampling without replacement and joint allocation, rather than the independent continuous-location experiment of § 4. Together with the shared errors of archived means, this distinction makes the exercise a retrospective assessment of reconstructed routes, without a claim of theoretical rate or confidence-coverage validation. Each budget and height has 500 sampling repetitions, with parameters fixed after separate development runs. The ODE uses fourth-order Runge-Kutta on 100 steps. Infeasible allocation, fewer than 14 regression rows, a design singular-value condition number above , or integration leaving the corridor returns the initial constant curve, which remains in every summary. Failed pilots likewise return a constant pilot. All 3000 active runs succeeded. Above the canopy at , four runs require radius expansion, to at most 40.67 mm; no other setting requires expansion. Assignment, distinct budgets, deterministic replay and numerical integration were checked.
Route features and reconstruction sensitivity
Table 10 shows that the stronger outward and downward departure within the canopy persists at all budgets. Mean outward displacement ranges from 3.75 to 4.26 mm within the canopy and from 0.21 to 0.44 mm above it, over similar downstream advances. The main figure averages the 500 active reconstructed paths at each common time and plots the resulting mean downstream and transverse coordinates. It summarizes the reconstructions; it is not a separately integrated velocity field or a confidence tube. The displayed setting was fixed before the formal runs. Shading marks the horizontal projection of the analysis corridor. The bar represents the canopy element with schematic thickness and is dashed when the element lies below the displayed height. Dots mark measured positions, the cross marks the common horizontal start, and dotted horizontal segments mark the initial streamwise direction.
| Flow | Downstream | Outward | Downward | |
| Within canopy | 80 | 26.08 (0.55) | 4.26 (0.30) | 2.13 (0.22) |
| Within canopy | 160 | 26.29 (0.36) | 4.22 (0.19) | 2.17 (0.15) |
| Within canopy | 240 | 26.29 (0.27) | 3.75 (0.16) | 2.02 (0.12) |
| Above canopy | 80 | 25.79 (0.21) | 0.21 (0.08) | 0.39 (0.06) |
| Above canopy | 160 | 25.82 (0.17) | 0.27 (0.06) | 0.38 (0.05) |
| Above canopy | 240 | 25.99 (0.16) | 0.44 (0.07) | 0.41 (0.05) |
Table 10. Endpoint displacements of active mean-flow reconstructions (mm): mean over 500 sampling designs, with the between-design standard deviation in parentheses. The standard deviations describe location-selection variability conditional on the archive, not measurement-error confidence intervals.
At each query position, the reference fits a total-degree-two polynomial to the reserved observations using Gaussian distance weights of scale 15 mm, then integrates the fitted local velocity. Sensitivity reconstructions use scales 12 and 18 mm, the complementary acquisition subset at scale 15 mm, and a local linear fit at scale 15 mm. These are checks on reconstructed route features, not additional acquisition methods or observed curve truth. The primary reference paths have positive streamwise velocity throughout and remain within 4.28 and 3.03 mm of actual measurement locations. Direct velocity summaries use and , with heights 36–48 mm within the canopy and 78–96 mm above it. These strips contain 50 and 54 measured positions, respectively.
The reserved-data quadratic reference gives outward and downward displacements of 3.72 and 2.09 mm within the canopy, versus 0.31 and 0.40 mm above it. Alternative quadratic references differ from these paths by at most 0.33 and 0.12 mm. The linear reference changes the within-canopy path by as much as 2.42 mm and gives outward displacements of 1.86 and 0.19 mm. Thus the direction and height contrast are more stable than the displacement magnitudes.
The linear-reference discrepancies in Table 11 are larger than those from changing the quadratic scale. These differences reflect reconstruction dependence and are not calibrated errors against physical truth. Finite spatial support also matters: above the canopy, the mean distance of selected locations from their assigned anchors increases from 17.13 to 19.68 and 23.99 mm as the budget grows. Disagreement with the primary reference decreases from 0.285 to 0.257 mm and then rises to 0.411 mm. A larger quota in a fixed archive therefore need not improve localization, even when allocation is feasible. The height contrast remains present despite this variation.
| Flow | 15 mm | 12 mm | 18 mm | Complementary | Linear | |
| Within canopy | 80 | 0.882 | 0.767 | 1.055 | 0.872 | 2.784 |
| Within canopy | 160 | 0.666 | 0.554 | 0.837 | 0.604 | 2.812 |
| Within canopy | 240 | 0.375 | 0.392 | 0.498 | 0.506 | 2.372 |
| Above canopy | 80 | 0.285 | 0.272 | 0.366 | 0.246 | 0.907 |
| Above canopy | 160 | 0.257 | 0.265 | 0.336 | 0.196 | 0.942 |
| Above canopy | 240 | 0.411 | 0.462 | 0.446 | 0.313 | 1.133 |
Table 11. Mean maximum three-dimensional disagreement between the active reconstructions and alternative reference curves (mm), evaluated at 101 times. Quadratic reference scales are indicated; the complementary subset and linear reference use scale 15 mm.
References
- [1]Ery Arias-Castro and Wanli Qiao. Confidence bands for the gradient lines of a density function. arXiv:2606.14087, 2026.
- [3]C. Banerjee, L. A. Sakhanenko, and D. C. Zhu. Global rate optimality of integral curve estimators in high order tensor models. Theory of Probability and Its Applications, 68(2):250–266, 2023.
- [4]C. Banerjee, L. A. Sakhanenko, and D. C. Zhu. Global rate optimality of integral curve estimators in high order tensor models: Supplemental material. Theory of Probability and Its Applications, 69(3):404–424, 2024.
- [6]Victor Chernozhukov, Denis Chetverikov, and Kengo Kato. Anti-concentration and honest, adaptive confidence bands. The Annals of Statistics, 42(5):1787–1818, 2014.
- [7]V. Chernozhukov, D. Chetverikov, and K. Kato. Comparison and anti-concentration bounds for maxima of Gaussian random vectors. Probability Theory and Related Fields, 162:47–70, 2015.
- [9]Ming-Yen Cheng and Hau-Tieng Wu. Local linear regression on manifolds and its geometric interpretation. Journal of the American Statistical Association, 108(540):1421–1434, 2013.arxiv.org/abs/1201.0327
- [10]Shane Elipot, Rick Lumpkin, Renellys C. Perez, Jonathan M. Lilly, Jeffrey J. Early, and Adam M. Sykulski. A global surface drifter data set at hourly resolution. Journal of Geophysical Research: Oceans, 121(5):2937–2966, 2016.DOI
- [11]Ian N. Harman, Margi Böhm, John J. Finnigan, and Dale Hughes. Spatial variability of the flow and turbulence within a model canopy. Boundary-Layer Meteorology, 160:375–396, 2016.DOI
- [12]Ian Harman, Margi Bohm, John Finnigan, and Dale Hughes. “Detailed Spatial Sampling” Wind Tunnel Experiment: Primary mean flow and turbulence data. Version 1, CSIRO Data Collection, 2021. doi: 10.25919/xs2b-be57.
- [13]Bolesław Kacewicz. Almost optimal solution of initial-value problems by randomized and quantum algorithms. Journal of Complexity, 22(5):676–690, 2006.
- [14]Kemna, S., Heidarsson, H. K., and Sukhatme, G. S. (2018). On-board adaptive informative sampling for AUVs: a feasibility study. OCEANS 2018 MTS/IEEE Charleston.DOI
- [15]Vladimir Koltchinskii and Lyudmila Sakhanenko. Asymptotics of statistical estimators of integral curves. In High Dimensional Probability V: The Luminy Volume, volume 5 of IMS Collections, pages 326–337. Institute of Mathematical Statistics, 2009.DOI
- [16]Vladimir Koltchinskii, Lyudmila Sakhanenko, and Songhe Cai. Integral curves of noisy vector fields and statistical problems in diffusion tensor imaging: Nonparametric kernel estimation and hypotheses testing. The Annals of Statistics, 35(4):1576–1607, 2007.DOI
- [17]Lucas C. Laurindo, Arthur J. Mariano, and Rick Lumpkin. An improved near-surface velocity climatology for the global ocean from drifter observations. Deep-Sea Research Part I: Oceanographic Research Papers, 124:73–92, 2017.DOI
- [18]Zhenhua Lin and Fang Yao. Intrinsic Riemannian functional data analysis. The Annals of Statistics, 47(6):3533–3577, 2019.DOI
- [19]Zhenhua Lin and Fang Yao. Functional regression on the manifold with contamination. Biometrika, 108(1):167–181, 2021.DOI
- [20]Rick Lumpkin and Gregory C. Johnson. Global ocean surface velocities from drifters: Mean, variance, El Niño–Southern Oscillation response, and seasonal cycle. Journal of Geophysical Research: Oceans, 118(6):2992–3006, 2013.DOI
- [21]Bruno Pelletier. Kernel density estimation on Riemannian manifolds. Statistics & Probability Letters, 73(3):297–304, 2005.
- [22]Bruno Pelletier. Non-parametric regression estimation on closed Riemannian manifolds. Journal of Nonparametric Statistics, 18(1):57–67, 2006.
- [23]Xavier Pennec. Intrinsic statistics on riemannian manifolds: Basic tools for geometric measurements. Journal of Mathematical Imaging and Vision, 25(1):127–154, 2006.DOI
- [25]Lyudmila A. Sakhanenko. A global optimal convergence rate in a model for the diffusion tensor imaging. Theory of Probability and Its Applications, 55(1):77–90, 2011.DOI
- [26]Lyudmila Sakhanenko. Rate acceleration for estimators of integral curves from diffusion tensor imaging (DTI) data. Statistics & Probability Letters, 107:286–295, 2015.DOI
- [27]D. J. Simmons, F. O. Thomas, T. C. Corke, and F. Hussain. Experimental characterization of smooth body flow separation topography and topology on a two-dimensional geometry of finite span. Journal of Fluid Mechanics, 944:A42, 2022.DOI
- [28]Charles J. Stone. Optimal global rates of convergence for nonparametric regression. The Annals of Statistics, 10(4):1040–1053, 1982.DOI
- [29]Runlong Tang, Moulinath Banerjee, and George Michailidis. A two-stage hybrid procedure for estimating an inverse regression function. The Annals of Statistics, 39(2):956–989, 2011.
- [30]Raf Theunissen, Jesse S. Kadosh, and Christian B. Allen. Autonomous spatially-adaptive sampling in experiments based on curvature, statistical error and sample spacing with applications in LDA measurements. Experiments in Fluids, 56:116, 2015.DOI
- [33]Susan Wei and Victor M. Panaretos. Empirical evolution equations. Electronic Journal of Statistics, 12(1):249–276, 2018.DOI
- [34]Wildmann, N., Vasiljevic, N., and Gerz, T. (2018). Wind turbine wake measurements with automatically adjusting scanning trajectories in a multi-Doppler lidar setup. Atmospheric Measurement Techniques 11, 3801–3814.DOI





