When a column of about a thousand water molecules is stretched between an atomically sharp tip and a flat mica surface, its stiffness does not fall smoothly. It drops in regular steps, roughly 0.5 N/m tall and about 1 nm apart, because the rim of the water bridge sticks to atomic ledges on the tip and then slips to the next one. Formation, Manipulation, and Elasticity Measurement of a Nanometric Column of Water Molecules reported these steps at 2%, 15% and 31% relative humidity. Above 40% humidity they were gone, and the stiffness curve looked like ordinary bulk capillarity. That paper, and the reading since, treat this as a physical crossover: the contact line stops pinning. There is a second possible reading. Wetter bridges are larger and more mobile, and their thermal fluctuations are bigger. The steps may still be present and simply too small, compared with the noise, to see by eye or with a threshold. No one has measured the detection limit, so the two readings cannot currently be told apart.
We propose to settle this with statistical tools from rare-event physics, adapted to one specific fact about the signal. Stick-slip steps are not arbitrary change points. They sit on a nearly regular ladder set by the atomic ledge spacing on the tip. We will build a likelihood model that treats a stiffness trace as a smooth capillary background plus a periodic comb of steps with unknown spacing, phase and height. The model is fit to the raw, unbinned time series and pooled across repeated retractions. A comb gathers evidence from all of its steps together instead of testing each one separately. It should therefore detect steps well below the level that per-step change-point methods or visual inspection can reach. It also gives a calibrated sensitivity curve: the smallest step height that could have been seen at a given humidity and noise level. That curve is what decides between "no pinning" and "pinning below our sensitivity."
Core Claim
The claim is that, for stick-slip in one-dimensional liquid bridges, the right detection statistic is a likelihood ratio. It compares a smooth-background-plus-periodic-step model against a smooth-background-only model, evaluated event by event on the continuous signal and calibrated by simulating null traces. Written out, the statistic for a set of $R$ retraction traces is
$$\Lambda = 2\sum_{r=1}^{R}\Big[\ln p(y_r \mid b_r, h, a, \phi_r) - \ln p(y_r \mid b_r)\Big],$$
where $y_r$ is the measured force-gradient trace of retraction $r$ as a function of tip height, $b_r$ is a smooth capillary background (a monotone spline with a small number of knots), $h$ is the step height shared across retractions, $a$ is the shared step spacing, and $\phi_r$ is a phase for each retraction. The phase is needed because each bridge starts at a random position on the ledge ladder. Averaging traces without a phase parameter smears the steps into a smooth curve. That is one likely reason the steps seem to disappear whenever noise forces experimenters to average. The spacing and phases are search parameters, so the statistic has a look-elsewhere penalty, as in the diluted cosmic-ray hotspot of the Telescope Array anisotropy study. We will calibrate that penalty by parametric bootstrap under the null model rather than assume asymptotic chi-square behavior.
The method draws on the rare-event physics literature on unbinned extended maximum likelihood. Unbinned Maximum Likelihood Estimation for Signal Extraction, J. Lee's rare-event likelihood analysis and D'Amico et al. show that evaluating densities at each event's continuous coordinates beats binned counting when signal-to-noise is low. The same literature also constrains the design: unbinned fits give biased estimates and overconfident intervals when the background shape is slightly wrong. For that reason the background here is flexible rather than a fixed capillary formula, and every significance is checked for coverage on simulated traces. The motivating result is still Choe et al. We do not dispute their low-humidity observation. What we question is how much the absence of steps at high humidity can support.
Research Program
The work has three parts.
Forward model. The first part is a generative model of the measurement. The physical layer is a Langevin model of the contact-line position in a tilted periodic potential, driven by the 0.15 nm/s retraction. It has three parameters: barrier height, ledge spacing near 1 nm, and a friction coefficient. The model produces stick-slip at low temperature and mobility, and smooth sliding when thermal energy exceeds the barrier. It is exactly the kind of model in which the steps can shrink continuously instead of switching off. The instrument layer is based on the quartz tuning-fork oscillator described by Choe et al. (stiffness about $5.4\times10^5$ N/m, amplitude 0.01 nm). It turns the bridge stiffness into a frequency-shift signal with thermal and detection noise. The noise level will be estimated from segments of real or simulated traces where no bridge is present.
Physical ground truth. The second part checks the physics against molecular simulation of water bridges between a stepped silicon-like tip and mica, at bridge sizes matching low and high humidity. We will use classical water models and a foundation interatomic potential from MACE. These simulations answer whether contact-line pinning, measured directly as the jump statistics of the rim atoms, survives in larger and wetter bridges. They also tell us how the expected step height should scale with bridge size. That prediction is essential for interpreting the sensitivity curve. We treat the simulations cautiously: Emergence of Debye Scaling in the Density of States of Liquids under Nanoconfinement found that classical molecular dynamics underestimates confinement effects compared with neutron experiments. So the simulations guide the scaling but do not decide the answer.
Detector and sensitivity curves. The third part is the detector itself: the comb likelihood, its bootstrap calibration, and injection-recovery sensitivity curves. Injection-recovery means planting steps of known height into null traces and measuring how often they are found at a fixed false-alarm rate. The detector will be run on simulated traces and, through collaboration with an atomic force microscopy group able to repeat the Choe et al. protocol across humidity, on real retraction series. We state this experimental dependency openly. The file lists no public trace dataset, so the real-data arm needs new measurements.
Baselines
The naive reference point is the procedure used in the original work: visually identifying plateaus in a single trace, or thresholding the derivative of a smoothed trace. We will implement it as a derivative threshold with the smoothing width and threshold tuned on simulated low-humidity traces, then frozen. Choe et al. report step height about 0.5 N/m and plateau length about 1 nm. They report no detection rate or false-alarm rate, so there is no reported result on Lacuna for step-detection power. The comb detector beats this baseline if, at the same false-alarm rate, its minimum detectable step height is clearly lower on identical simulated traces at high-humidity noise levels.
The strongest competitor is generic change-point detection: penalized step-fitting that finds an unknown number of step locations in a noisy piecewise-constant signal, run on the same raw traces. It is strong because it does not need the periodicity assumption and so cannot be misled by it. We will give it the same background model, the same data and a matched hyperparameter tuning budget, choosing penalties by the same bootstrap false-alarm calibration. I found no reported result on Lacuna for change-point methods on liquid-bridge force data. The comb must beat it at matched false-alarm rate by a margin larger than the bootstrap uncertainty of the sensitivity curve.
A learned classifier trained on simulated traces is included as a third comparison. It is the analogue of the neural discriminator in Search for Neutrinoless Double-Beta Decay with the Complete EXO-200 Dataset, which reported about a 25% gain in half-life sensitivity over single-site/multi-site cuts. It gets the same training simulations and compute as the comb detector.
The key ablation isolates the ingredient we claim matters. We run the full likelihood with the periodic constraint removed (independent step heights and positions), and separately with the per-retraction phases removed (steps forced to align across traces). If the independent-steps version matches the comb, the periodic ingredient adds nothing and the claim fails.
Why not just…?
Why not just average many retractions to beat down the noise? Averaging works only if the steps line up across retractions. They do not, because the first pinning ledge a new bridge meets is random, so the phase varies from trace to trace. Averaging without phase alignment produces a smooth curve even when every individual trace has steps. The phase-ablation experiment measures how much sensitivity this costs. If phases turn out to be nearly fixed in practice, for example set by a reproducible tip apex, plain averaging becomes a serious rival, and we would report it as the simpler method.
Why not just use an off-the-shelf change-point detector? This is the real threat. On clean low-humidity data, a well-tuned step-fitter probably does as well as the comb. The advantage we claim appears only near the noise floor. There, each isolated step is individually insignificant, but a regular ladder of them is not. The head-to-head sensitivity curves in the proposal settle this directly. If they coincide, the contribution shrinks to the calibration and injection-recovery methodology alone, which still answers the physical question.
Why not just simulate the wet bridge and see whether pinning survives? We will run these simulations, but they cannot decide the question alone. The force fields are known to misjudge confinement strength (Jin et al.). The simulated tip is an idealized stepped surface, not the real silicon apex. And a simulation says nothing about what the instrument could have detected. Simulation supplies the expected step-size scaling. The measurement-side sensitivity curve is still needed to compare that expectation with what was observed.
Why not just build a quieter instrument? Better hardware would help and is complementary. But the existing and planned measurements need an honest detection limit whatever their noise level. Without a detection limit, a better instrument would only move the unanswered question to a higher humidity.
Proposed Experiments
The first experiment establishes a calibrated sensitivity map on synthetic data. We generate traces from the forward model over a grid of barrier heights, step heights from 0.05 to 0.5 N/m, and noise levels spanning the low-to-high humidity range. For each detector we measure detection power at a bootstrap-calibrated false-alarm rate matching $3\sigma$ and $5\sigma$ global significance. We also check coverage of the fitted step-height intervals, because unbinned fits can be overconfident. Coverage is repeated with deliberately misspecified backgrounds: a wrong capillary curvature, and drift in the oscillator's resonance frequency. This tests robustness to the known weakness of unbinned methods.
The second experiment is the physical check in simulation. We measure contact-line jump statistics and bridge stiffness in molecular simulations of small (low-humidity) and large (high-humidity) bridges. From these we extract the expected step height as a function of bridge size and compare it with the sensitivity map.
The third and decisive experiment runs the detector on retraction series measured at about 30% and 45% humidity with the Choe et al. protocol. At 30% the detector must recover steps near 0.5 N/m with regular spacing near 1 nm; this is a sanity check against the published observation. At 45% there are three possible outcomes, and each is interpretable. If steps are detected at global significance, the reported crossover is a detection artifact. If no steps are found but the sensitivity curve shows that steps of the simulation-predicted size would have been detected, the crossover is real, and for the first time it has a quantified bound on residual pinning. If the predicted steps lie below sensitivity, the honest conclusion is that current data cannot decide, and we report the exposure (number of retractions) needed to decide. Under the Poisson-like scaling of pooled evidence, that exposure grows as the inverse square of the ratio of step height to noise.
References
- Formation, Manipulation, and Elasticity Measurement of a Nanometric Column of Water Molecules
- INDICATIONS OF INTERMEDIATE-SCALE ANISOTROPY OF COSMIC RAYS WITH ENERGY GREATER THAN 57 EeV IN THE NORTHERN SKY MEASURED WITH THE SURFACE DETECTOR OF THE TELESCOPE ARRAY EXPERIMENT
- Dalitz plot analysis of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msubsup><mml:mi>D</mml:mi><mml:mi>s</mml:mi><mml:mo>+</mml:mo></mml:msubsup><mml:mo>→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>K</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup></mml:math>
- Update of the search for the neutrinoless decay<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mrow><mml:mover><mml:mrow><mml:mi>τ</mml:mi></mml:mrow><mml:mrow><mml:mo>→</mml:mo></mml:mrow></mml:mover></mml:mrow></mml:mrow><mml:mi>μ</mml:mi><mml:mi>γ</mml:mi></mml:math>
- Signal estimation in on/off measurements including event-by-event variables
- Emergence of Debye scaling in the density of states of liquids under nanoconfinement
- Search for Neutrinoless Double-<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi>β</mml:mi></mml:mrow></mml:math>Decay with the Complete EXO-200 Dataset