The slow approach: what VO2 oscillator jitter teaches every threshold detector

What the work claims

This is a primary experimental result: fabricated devices, measured spectra, a proposed mechanism, and a demonstrated control knob.¹ The authors, a collaboration between the Universities of Gothenburg and Oakland, UC San Diego, and Tohoku University, study the Pearson-Anson relaxation oscillator built from a 500 by 100 nanometer VO2 film: a resistor-capacitor charging circuit in which the vanadium dioxide film plays the role of the comparator, snapping from insulator to metal when the voltage across it crosses a threshold. Such oscillators are candidates for neuromorphic networks and oscillator-based Ising machines, and their known weakness is a spectral linewidth on the order of tens of kilohertz, far broader than comparable CMOS relaxation oscillators achieve, which the paper notes can be compensated to practical figures of merit close to theory.

The paper makes two claims. First, the dominant source of that broadening at low oscillation frequencies is not the switching event itself but the incubation phase before it: as the capacitor voltage approaches the insulator-to-metal transition threshold with a shallow slope, small thermal fluctuations produce large timing errors. Second, the waveform of an external synchronization signal matters as much as its amplitude: a square wave at twice the oscillation frequency locks the oscillator at lower injected amplitude, and suppresses phase noise more effectively, than a sinusoid of the same size, because its sharp edges steepen the approach to threshold and shorten the stochastic incubation window.

How it works

A relaxation oscillator does not resonate; it charges and fires. The external capacitor charges through a resistor toward a steady-state voltage until the drop across the VO2 film reaches the insulator-to-metal transition threshold, the film collapses to its metallic state, the capacitor dumps its charge through it, and when the voltage falls below the reverse threshold the film reverts to an insulator and the cycle restarts. The oscillation period is essentially the charging time, and both thresholds fluctuate because the transition is thermally driven: the film switches when Joule heating and ambient fluctuations carry it over a critical temperature, not when a clean electronic comparison flips.

The measurements tie the spectral and time-domain pictures together consistently. At a supply of 3.5 volts the phase noise decomposes into a flicker-frequency region falling as one over frequency cubed below roughly a kilohertz of offset, a white-frequency region falling as one over frequency squared up to about ten megahertz, and a flat thermal floor beyond. As the supply drops below 2.2 volts, the white-frequency component rises steeply while the flicker component stays constant, until it dominates the spectrum down to 10 hertz offsets; the authors quantify the increase as approximately 24 rad^2/Hz of phase noise power. The time domain says the same thing more vividly: period jitter, measured as the full width at half maximum of the period histogram, is 9.8 nanoseconds at 3.4 volts, 50.7 nanoseconds at 2.0 volts, and 280.2 nanoseconds at 1.81 volts, where the histogram also becomes visibly right-skewed. That skew is the signature of first-passage-time statistics: when the drift toward threshold is shallow, the crossing waits for a rare fluctuation, and the waiting-time distribution grows a long tail. The supplementary analysis condenses this into a scaling law: jitter is inversely proportional to the threshold overdrive, the excess of supply voltage over the extracted threshold of about 1.79 volts, with a high-overdrive jitter floor near 3.5 nanoseconds. One metric note: the scaling law is fitted to the standard deviation of the period distribution, while the headline jitter values are histogram widths, and the two track each other only if the distribution's shape holds steady across overdrive.

The underlying ruler is one line: timing uncertainty is voltage noise divided by slew rate at the threshold. The square-wave synchronization result follows directly. A sharp edge arriving at the right phase drives the film across its threshold quickly, so the device spends less time in the fluctuation-sensitive regime; the paper reports the square wave producing visible noise reduction at 5 to 15 millivolts and partial injection locking at 25 millivolts, where the sinusoid reaches a full lock only near 40 millivolts; those are different lock regimes, so the amplitude comparison is suggestive rather than strictly like-for-like. Square-wave drive at 100 millivolts also speeds the amplitude-restoration response, raising the amplitude relaxation frequency to one to two hundred kilohertz.

Where a skeptic should push

The load-bearing assumption is that a single device geometry, from a single film growth, characterizes the class. Every number above comes from one 500 by 100 nanometer device fabricated from one sputtered 100 nanometer film; VO2 thresholds are known to be strongly hysteretic and film-dependent, and the paper offers no device-to-device spread. The mechanism argument does not depend on that spread, but the quantitative anchors, the 1.79 volt threshold and the 3.5 nanosecond floor, do. The transition is also hysteretic and memristive, so the thresholds may carry cycle-to-cycle memory; the analysis treats each crossing as a fresh first-passage draw, and correlated threshold dynamics from one cycle to the next would complicate that model in ways the paper does not test.

Second, for a device whose entire switching physics is thermal, the thermal environment is underspecified in the main text. The scaling with electrical overdrive is convincingly demonstrated; the equivalent scaling with ambient temperature, which is what a deployment environment would actually impose, is not measured.

Third, there are small internal inconsistencies a careful reader trips over. The full-lock plateau under 0.04 volt sinusoidal drive is stated once as extending below an offset of five times ten to the five hertz and, in the adjacent sentence, below five times ten to the four hertz. And the quoted white-noise increase of approximately 24 rad^2/Hz reads more naturally as a decibel figure from the logarithmic plots; as printed, the unit does not describe a dimensionless gain. Neither slip undermines the mechanism, but both should be resolved before anyone designs to these numbers. A methodological note belongs here too: the phase-noise decomposition rests on a Hilbert-transform phase extracted from a strongly non-sinusoidal waveform, a representation in which harmonics can fold amplitude fluctuations into the extracted phase; the time-domain period histograms are the more assumption-free of the paper's two views.

Finally, the framing claims about Ising machines and neuromorphic networks are extrapolations: everything measured here is a single oscillator under injection from a signal generator. The step from one locked oscillator to a network of coupled ones, where phase slips between oscillators are the errors that matter, is asserted, not demonstrated.

Level crossings between tissue and silicon

The source never mentions biopotentials, electrodes, or arrays; what follows is this analysis's extrapolation, grounded in the paper's mechanism. It matters for microelectrode array instrumentation twice over: once for anyone contemplating VO2 oscillators as near-array compute, and once, more broadly, because an event-driven acquisition chain is a stack of level-crossing detectors, and this paper is a clean experimental study of what level-crossing timing does when the approach to threshold is slow.

Take the direct reading first. Oscillator networks built from phase-transition devices are regularly proposed as compact neuromorphic co-processors, exactly the sort of thing one might co-locate with an array to close a stimulation loop without a host computer. This paper hands the integrator two hard specifications. A free-running linewidth of tens of kilohertz puts a short coherence budget on any phase-encoded state, so a working network will need injection locking or some other phase-restoring mechanism much of the time, and the locking signal this paper recommends is a harmonic-rich square wave at twice the oscillation frequency: precisely the kind of aggressor, sharp edges and a comb of harmonics, that a microvolt-level recording front end sharing a board, a supply, or a substrate cannot ignore. And because the device switches thermally, an incubator environment is unlikely to be neutral to it: the same fluctuation sensitivity that sets the jitter floor should make the operating point a function of ambient temperature. The paper never sweeps ambient temperature, so this is an unquantified deployment hypothesis, not a demonstrated failure mode; but a compute element whose switching physics is itself thermal, sitting millimeters from tissue held at 37 degrees, plausibly couples temperature regulation to computational stability more tightly than conventional logic does, and that coupling deserves its own line in any feasibility analysis.

The broader reading is the more useful one. Spike detection by comparator, event-driven delta modulation at the sensor, time-domain and oscillator-based converters: all of these decide when an analog voltage crosses a level, and all of them obey the paper's ruler, timing spread equals voltage noise over slew rate at the crossing. An earlier version of this analysis compressed the consequence into a single claim of systematic late bias for small events; the cross-check rightly rejected that, and the corrected version is worth spelling out, because three distinct mechanisms make a threshold detector mistime small events and they behave differently. The first is deterministic time walk: a smaller event of the same shape reaches a fixed threshold later in its rise, a plain geometric shift that calibration can remove. The second is stochastic jitter, which widens inversely with slew; and because a waveform in additive noise triggers on the first noise excursion above threshold, this term actually biases the mean crossing slightly early, not late, so the VO2 paper's right-skewed tail must not be imported wholesale. The third is where the incubation physics does have an electronic cousin: a comparator's regeneration time stretches logarithmically as overdrive shrinks, so events that barely clear threshold acquire a genuine, diverging late tail. What survives all three is the operational point: timing error and its shape become functions of event amplitude, so analyses that lean on fine timing across channels, latency ordering, propagation mapping, synchrony estimation, are silently mixing error distributions whose width, walk, and tails depend on signal size, and the effect is largest for small, slow, near-threshold events, the regime of delta modulators and field-potential crossings rather than of large, fast spike edges. The prior entry on this stream about timing metadata dealt with the timebase; this is the layer beneath it, the detector physics, and no amount of clock discipline removes it.

The genuine opportunity is that the same physics is a calibration model, not just a hazard. Jitter inversely proportional to overdrive is a measurable, parameterizable law: characterize a detector's timestamp error against crossing slope and amplitude, and the walk and the tails become correctable in analysis rather than invisible. The waveform lesson transfers too: where an acquisition chain does control an edge, a clock, a stimulus marker, a sampling strobe, edge steepness buys timing stability at fixed amplitude, a cheap lever. The genuine threat is the design conflict the paper makes explicit: the cure for threshold jitter is to steepen the approach to threshold, and a biosignal front end only half-controls that. The tissue sets the source waveform; gain and filtering can reshape the slope a comparator actually sees, but linear gain scales signal and in-band noise together, leaving the noise-over-slew ratio untouched, and buys relief only against noise, drift, and regeneration effects added after the gain. Oversampled digital detection with interpolation sidesteps the analog level-crossing problem entirely, at a price in power and latency. Event-encoding front ends that timestamp threshold crossings of small, slow signals in the analog domain are therefore working in the regime this paper identifies as pathological, and their datasheets, which quote timing resolution as a clock period, will not say so.

The bottom line

As device physics, this is a solid single-device result, mechanistically coherent despite the two wording slips noted above: the incubation-phase mechanism is demonstrated three independent ways, spectrally, in time-domain jitter, and through the overdrive scaling law, and the square-wave-beats-sinusoid finding comes with a physical explanation that predicts it. Established: threshold-crossing timing in this device degrades inversely with overdrive and is governed by first-passage statistics near threshold. Hypothesis: that this transfers quantitatively across devices, films, and operating temperatures, and that locked networks of such oscillators inherit the single-device improvement. For array instrumentation, the transferable content is the ruler, not the device: amplitude-dependent timing error, deterministic walk plus slew-dependent jitter plus late tails from vanishing overdrive, is the default behavior of analog level-crossing detectors fed by small, slow signals, and it can be either characterized or ignored. Replication across devices, an explicit temperature sweep, and a two-oscillator locking experiment would each move the framing claims from asserted to demonstrated; a failure of the overdrive scaling law in other geometries would break the paper's central mechanism.