An event sensor's transfer function depends on its probe
An event vision sensor is silent in front of a static scene, so the classical way of measuring a system's transfer function, hold the input constant and read the output, does not exist for it. Nimrod Kruger and Gregory Cohen of Western Sydney University work through what replaces it: the probe must move, the probe's shape becomes part of the measurement, and different event statistics invert to transfer functions with wildly different fragility. Their analytical result is blunt. The popular step probe, which reads amplitude through event counts, collapses under a five percent variation in pixel thresholds, while linear and exponential ramp probes, which read time scales, still deliver usable transfer functions at fifteen to twenty percent mismatch. For anyone characterising a spiking system, camera or cortical organoid, this is the difference between measuring the device and measuring its heterogeneity.
Source: Modelling dynamic systems transfer functions from events in computational neuromorphic imaging, arXiv:2609.29863, September 2026. Primary source. Read: the full 21-page PDF including the pixel model, the closed-form inverses for all three probes, the noise and mismatch analysis of Section IV, and the discussion of moving-point-source calibration.
What the work claims
This is an analytical paper with simulated verification: closed-form derivations for an idealised event sensor, checked against a specified ground-truth optical transfer function in one dimension. The central claim is about the order of operations. An event-measured point spread function is not the optical one; it is a pointwise nonlinear warp of it, set jointly by the probe, the pixel model, and the statistic read from the event stream. Because the warp is pointwise, it must be inverted per pixel before any Fourier transform is taken: transforming the raw statistic and inverting afterwards gives the transform of a warped function, which does not converge to the true transfer function even as noise vanishes. The authors write the inversion chain out for three temporal probes, a step, a linear ramp, and an exponential ramp, and benchmark each by reconstructing a known modulation transfer function from simulated event streams.1
The verified numbers: for a 101-pixel simulated array with a contrast threshold of 0.25 log units and a 300 microsecond refractory period, step-probe reconstructions reach 7.7 percent RMS error using event counts, 10.8 percent using onset latency, and 5 percent using the first inter-event interval, with comparable total event counts around 173. Ramp probes do better per event: the linear ramp reaches roughly 5 percent RMS with all three of its statistics across 174 events, and the exponential ramp reaches 4.4 percent with its first-interval inverse and 6.1 percent with first-event delay across 172 events. Then the fragility ordering appears. Under per-pixel threshold mismatch, the step probe blows past 20 percent reconstruction error once mismatch exceeds about 5 percent, because its inversion is exponential in the mismatch and amplified by the probe's dynamic range; the linear ramp stays under 20 percent error at 15 percent mismatch; the exponential ramp performs well at around 20 percent. Under photon shot noise all probes fail only when the noise reaches the threshold itself, a fluctuation-to-threshold ratio near one to two.
How it works
The pixel model is deliberately minimal: a logarithmic photoreceptor, a single-pole front-end filter with time constant, a symmetric contrast threshold, and a refractory period during which the reference is frozen. Each probe is a point source whose intensity follows a temporal profile. Under a step, the excursion in log-irradiance at a pixel is proportional to the optical weight there, and the number of events in the resulting burst inverts to that weight through two cascaded exponentials, one from the logarithm and one from the refractory-limited burst. Under a linear ramp, the weight stops setting amplitude and instead sets a characteristic time, so the inversion becomes affine in a measured interval followed by a reciprocal: one additive offset, no exponential. Under an exponential ramp, matched to the logarithmic photoreceptor, the intervals settle to a constant that carries only sensor parameters, the threshold and refractory period, while the onset latency alone carries the optical weight; recovering the transfer function then requires no sensor parameter at all under the normalisation the authors use, and the stationary interval doubles as a null test for model mismatch.
Two structural facts do the heavy lifting. First, the warp is pointwise only because the probe supplies the entire temporal variation and every pixel shares a single onset time; that is what makes per-pixel inversion valid, and it is exactly what a natural dynamic scene does not provide. Second, threshold mismatch hurts the step probe exponentially because the threshold sits inside the inversion as a multiplicative factor on the measured log-range, so a per-pixel threshold error becomes a lognormal weight error; in the ramp inversions the threshold appears only in factors shared by all pixels, which normalise away, leaving a residual sensitivity of order unity.
Where a skeptic should push
The load-bearing assumption is the pixel model. A single front-end pole, symmetric thresholds, monotonically rising probes carried entirely by ON events, and one spatial dimension are all chosen for tractability, and the authors say plainly that a second pole breaks the interval cancellation the linear-ramp inverse relies on. Real devices bias the two polarities independently, often by tens of percent, and real scenes are superpositions with diverse onsets, which voids the pointwise property and with it every closed-form inverse in the paper; the authors state this themselves and reframe scene reconstruction as an open problem needing a composition rule for statistics. The authors also flag that their results are indicative rather than a noise analysis, and that their error metric is bounded above, so it cannot distinguish degrees of failure once a reconstruction has failed. Finally, the mismatch study emulates only static threshold spread; time-varying mismatch, the kind that dominates real sensors and all biological tissue, is not modelled.
One transparency note deserves weight in a skeptical reading: the authors disclose that an AI assistant was used as an interactive collaborator for literature retrieval, independent derivation checking, simulation debugging, and prose drafting, with the author verifying all derivations against the stated model and all claims numerically. That is a more honest accounting than most event-vision papers give, but a reader should still treat the closed forms as author-certified rather than independently refereed; the paper is a preprint.
Probing organoid transfer functions from spike trains
For organoid intelligence this paper is a mirror held up to a ubiquitous experimental practice. Characterising a cortical organoid's input-output properties means stimulating, electrically or optogenetically, and reading the resulting spikes; the implicit hope is that the measured response curve is a property of the tissue. Kruger and Cohen prove, for a system far simpler than tissue, when that hope is structurally justified and when it is not. The mapping is direct. A step probe with a rate readout is the organoid field's workhorse: pulse amplitude up, count spikes, fit a curve. That is precisely the fragile configuration, an amplitude-coded statistic whose inversion passes measured quantities through exponentials, so per-cell threshold variability, which in tissue spans many-fold rather than the simulated tens of percent, contaminates the inferred transfer function exponentially. The paper's equation for the step-probe error says the damage scales with the probe's dynamic range: the harder you drive tissue to get a clean response, the more heterogeneity your measurement amplifies. A great deal of published organoid input-output characterisation may be measuring cell-to-cell variability dressed as tissue dynamics.
The opportunity is the alternative the paper hands over. Ramp-style probes, graded sustained stimuli whose strength rises over seconds, move the signal into event timing rather than event counts, and timing-based inversions shed the exponential and with it most of the heterogeneity sensitivity. For tissue this suggests a concrete experimental protocol: characterise organoids with ramped stimulation and read out latencies and inter-spike intervals, not counts, and report the transfer function only after the per-site inversion step. The exponential-ramp construction adds a second gift: its statistic separates into a stationary component carrying only the sensor's parameters and a phase component carrying the signal. Translated to neurons, that is a formal recipe for separating intrinsic excitability, thresholds and refractoriness, from synaptic drive, by stimulus design rather than by pharmacological dissection, and for using the stationary component as a running null test that flags when the tissue has drifted out of stationarity mid-experiment. The invert-first-transform-second rule is the cheapest takeaway of all: any pipeline that averages spike rates across sites before analysing spectral properties is computing a warped transfer function, and the warp does not average away.
The threat is honest asymmetry. Event pixels are stationary, independent, and reset on a fixed schedule; neurons adapt, burst, and drift on minutes-to-days timescales, so the mismatch the paper treats as fixed-pattern noise is in tissue a moving, history-dependent, correlated process that no additive-offset inversion removes. The probe-sharing premise also fails in one-electrode-many-neurons recordings, where a site is not a pixel. And there is a hype-correction angle: robustness of ramp statistics is robustness to static heterogeneity, not a licence to skip controls, since a ramp that outruns adaptation stops probing the tissue you think it probes. Used with those caveats, the paper converts a scattered methodological anxiety into two design rules, match the excitation to the statistic, and invert before you transform, that organoid characterisation studies can adopt this year.
The bottom line
Established, by derivation plus simulation against a known ground truth: event-stream statistics are probe-dependent warps of the underlying transfer function; inversion must precede transformation; step-like amplitude coding is exponentially fragile to threshold mismatch, failing past roughly 5 percent, while linear and exponential ramp coding tolerate roughly 15 and 20 percent; and all schemes fail when photon noise approaches the contrast threshold. Not established: behaviour on real sensors or any scene beyond the single-probe regime, robustness to time-varying non-idealities, or any biological validation. For organoid intelligence the paper's value is methodological discipline: ramped stimulation with timing readouts and per-site inversion is the defensible way to measure tissue transfer functions, and count-based step characterisations should be presumed contaminated by heterogeneity until shown otherwise. The confirming experiment would be an organoid characterisation that recovers the same transfer function under deliberate cell-to-cell variability manipulation only with the ramp protocol; the breaking observation would be adaptation dynamics so fast that no ramp stays quasistatic, at which point even these inversions need a non-stationary extension the paper does not attempt.
Frequently asked questions
What is an event vision sensor?
A camera whose pixels emit asynchronous events when the log of local irradiance changes by a fixed contrast threshold, rather than integrating frames. Static scenes produce no output at all, which is why transfer-function measurement needs a probe with a temporal profile.
What is the probe-statistic correspondence?
The pairing of a chosen temporal excitation, such as a step or a ramp, with a specific reading of the event stream, such as counts, latencies, or intervals. The paper's point is that each pair has its own closed-form inverse with its own fragility, so excitation and statistic must be designed together and reported together.
Why does the step probe fail under threshold mismatch?
Its inversion is exponential in the measured event count, and a per-pixel threshold error multiplies the inferred log-range, producing a weight error that is exponential in the mismatch and grows with the probe's dynamic range. Around 5 percent threshold spread pushes reconstruction error past 20 percent.
What does invert first, transform second mean?
Because the sensor's nonlinearity acts per pixel before any spatial mixing, the correct pipeline is to invert the per-pixel statistic to a weight, then take the Fourier transform. Transforming raw statistics first yields the transform of a warped point spread function, which does not converge to the true transfer function even with zero noise.
Is the exponential ramp probe practical?
It is the least convenient to generate, requiring a logarithmic intensity driver or a fine high-bit-depth staircase, but it is the most informative: its intervals carry only sensor parameters while its onset latency carries the signal, so it can even measure per-pixel thresholds and refractory periods under unstructured illumination.
What should an organoid experimentalist take from this?
Three rules: characterise with ramped rather than step stimulation; read out timings and intervals rather than counts; and apply any nonlinear inversion per recording site before aggregating. Count-based amplitude response curves in heterogeneous tissue are the step-probe configuration and inherit its exponential fragility.
References
- N Kruger, G Cohen. Modelling dynamic systems transfer functions from events in computational neuromorphic imaging. arXiv:2609.29863. 2026. https://arxiv.org/abs/2609.29863. Accessed 2026-11-21.