At a glance
The absolute threshold is quoted as a constant but measured as a rate. Ten papers, from Hecht's 1942 photon count to Currie's detection limits, explain why.
On a July afternoon in 1942, three researchers in Columbia University's biophysics laboratory published a number nobody has since improved on. Selig Hecht, Simon Shlaer and Maurice Henri Pirenne dark-adapted their observers, flashed a small disc of blue-green light, and asked whether anything had been seen. Energy arriving at the cornea came to between 54 and 148 quanta of light. Allow for what the eye loses on the way in — roughly half is absorbed in the ocular media alone — and the figure that did the work was 5 to 14 quanta captured by the rods.
Five photons. That is the absolute threshold of human vision as textbooks give it, and it holds up. It is also a rate dressed as a constant: the intensity at which an observer reports a flash on half the trials where one appears. Shift the observer's criterion, widen the target, lengthen the flash, and the same eye will register three times less light, or ten times more, without anything in the retina changing.
That gap between a clean number and a measured one runs through the ten papers below. Most published thresholds are not physical constants at all; they are agreements. The exceptions are the ones where physics, not committee, sets the floor.
The line that turned out to be a slope
Gustav Theodor Fechner gets the credit and some of the blame. In Elemente der Psychophysik, published in 1860, he set out three procedures for locating the smallest detectable stimulus — the method of limits, the method of constant stimuli and the method of adjustment — and used the result as the zero point of a sensory scale. The procedures survived. The zero did not.
Run the constant-stimuli version properly and no step appears. You get a sigmoid: at very low intensities the observer says yes now and then, at high intensities almost always, and somewhere in the middle the curve crosses over. The threshold is wherever you decide to put it, which by convention is the 50 per cent point. The reason is ordinary biology. Sensory neurons fire without any stimulus at all, so every detection is a judgement made against a noisy background.
Fechner himself allowed for what he called negative sensations, responses to stimuli too weak to reach consciousness. The modern version of that concession is that below threshold does not mean unprocessed. It means unreported.
Ten papers, in the order they arrived
The sequence is chronological, not a citation ranking. A defensible ranking would need a field-by-field count this article cannot honestly supply, and influence in this corner of science does not track neatly onto reference lists anyway. What the list does show is a philosophical question about the smallest detectable thing turning into an engineering problem, then into a statistical one.
Photo by little plant on Unsplash
- Sivian and White, 1933, Journal of the Acoustical Society of America. Fourteen ears, 100 to 15,000 cycles per second, Bell Telephone Laboratories. The first trustworthy map of the quietest audible sound field, and the argument that a more sensitive ear would be pointless because thermal noise in the air sets the floor.
- Hecht, Shlaer and Pirenne, 1942, Journal of General Physiology. Five to fourteen quanta absorbed across roughly 500 rods, plus the finding that the trial-to-trial flicker in what observers saw came from Poisson variation in the light rather than from the observers.
- Albert Rose, 1948, Journal of the Optical Society of America. Treated the eye as a signal-to-noise device. Quantum efficiencies between 0.5 and 5 per cent covered the whole range of human visual performance, making the threshold a ratio rather than a quantity.
- Hardy, Wolff and Goodell, 1948, Journal of Clinical Investigation. The dolorimeter, a focused beam of heat, and the dol scale built on it. Claimed a sensory pain threshold uniform across neurologically normal people while the reaction to pain varied wildly.
- Tanner and Swets, 1954, Psychological Review. Detection is a decision, they argued. Sensitivity and criterion could be separated, and the classical threshold concept needed re-evaluation.
- Barlow, 1956, Journal of the Optical Society of America. Put the visual limit in retinal noise rather than optics: what stops you seeing the last photon is the retina's own spontaneous activity.
- Currie, 1968, Analytical Chemistry. Replaced one detection limit with three — the critical level, the detection limit and the determination limit. Chemistry has used his definitions ever since.
- Sakitt, 1972, Journal of Physiology. Asked observers to rate blanks and dim flashes on a 0 to 6 scale. One subject's ratings behaved as if he were counting every rod signal, noise included.
- Baylor, Lamb and Yau, 1979, Journal of Physiology. A suction electrode on single toad rod cells: one photoisomerisation produced a current event of about 1 picoamp, 5 per cent of the dark current.
- Dey, Zele, Feigl and Adhikari, 2021, Vision Research. Full-field threshold vision needed about a thousand photons spread over roughly 92 million rods — more sensitive than the classic focal measurement, and a reminder that stimulus geometry is part of the answer.
Why vision is the one sense with a hard number
Hecht's team did more than count photons. They showed that at threshold the variability sits in the stimulus: a flash of nominally fixed energy delivers a Poisson-distributed number of quanta, and the observer's yes or no tracks that distribution. The 1942 paper is also scrupulous about its own arithmetic, listing the corrections for corneal reflection, media absorption and retinal transmission separately — the kind of bookkeeping that makes a result reusable eighty years later.
Barlow's contribution was to explain why the number is not zero. Rhodopsin occasionally isomerises without a photon, so the retina carries a background of photon-like events. Detection becomes a signal-to-noise problem, and the observer's willingness to guess is a variable separate from the observer's sensitivity. Sakitt tested that by letting people rate their confidence on blank trials as well as real ones. Her forced-choice simulation put the threshold around 0.4 log units below the ordinary yes/no value — a factor of about two and a half, bought entirely by changing the instructions.
Then physiology caught up. Baylor, Lamb and Yau sucked the outer segment of a toad rod into an electrode and recorded the current. One photon, one event, with a quantum efficiency of 0.5 plus or minus 0.1 — about as close to a physical limit as a biological detector gets.
The argument since has been about geometry rather than mechanism. The 2021 full-field study compiled published estimates running from 43 photons at the cornea to just over a million, depending on target size, duration and where on the retina the light fell. Their own answer, roughly a thousand photons across 92 million rods, is a summation model rather than a new physical constant. Rods pool signals, and larger stimuli recruit more noise along with more signal.
Hearing's zero is a committee decision
The reference for sound pressure level is 20 micropascals, and it is defined as 0 decibels at 1,000 hertz for a young, otologically normal listener. That phrasing gives the construction away: average ear, specific frequency, defined age band. Sivian and White measured the minimum audible field on 14 ears at Bell Labs in 1933 and found the average at 1,000 cycles per second sat about 71 decibels below one bar of pressure. Their conclusion was that a keener ear would gain nothing, since the thermal motion of air molecules would swamp it.
Sensitivity is not flat. It peaks between roughly 2 and 5 kilohertz, where standardised thresholds fall to about minus 9 decibels — below the nominal zero, which is why negative dB SPL readings are legitimate rather than an error. The number has been revised repeatedly: a full rewrite in 2003 after 18 years of work, then a 2023 edition that lowered the 20 hertz threshold by 0.4 decibels to match a revised reference in ISO 389-7. The two editions differ by at most 0.6 decibels, so nothing clinical turns on it.
What does turn on it is the audiogram. Because the reference is standardised and the transducers are calibrated against it, a hearing level of 60 dB at 4 kilohertz means the same thing in Brisbane and in Bremen. A threshold everyone has agreed to stop arguing about is a useful instrument, even when it is not a fact about ears.
Currie drew three lines where everyone had drawn one
Chemistry had its own version of the problem, and it was worse. By the late 1960s the literature carried a tangle of competing recipes for the detection limit, several of them statistically indefensible. Lloyd Currie took a worked example — counting a radioactive source for ten minutes with a 10 per cent efficient detector against a background of 20 counts per minute — and applied eight common definitions to it. The answers spanned a factor of nearly a thousand.
His 1968 paper replaced the single number with three. The critical level is the instrument reading above which you may call something detected. The detection limit is the true signal level that will, a priori, produce a detection. The determination limit is where you can actually quantify, at a stated precision. With 5 per cent risks of both false positives and false negatives, and a blank whose standard deviation is known, the working expressions come out as multiples of that deviation: 1.64 for the critical level, 3.29 for the detection limit, roughly 10 for the determination limit. The framework is now standard, and the NIST account of the paper's influence records more than a thousand citations by the time it was written up as a landmark.
Photo by Brett Jordan on Unsplash
Pain and smell never settled on a number
Hardy, Wolff and Goodell came out of Cornell with a device and a unit. The dolorimeter delivered a focused beam of infrared heat to a small patch of skin, and the dol scale was built from just-noticeable differences in intensity. Their 1948 paper on the dol scale argued that the sensory threshold for pain was constant across neurologically normal people, and that what varied was the emotional reaction to it. It was an attractive claim. It also asked subjects to be detached, attentive observers of a pure sensation, which is not how anyone meets pain in a clinic, and the instrument never became clinical furniture.
Smell shows the same pattern from another direction. Punter's 1983 study of 58 structurally related compounds, run with a standardised olfactometric procedure and a consistent panel, found pooled group thresholds reliable to within a factor of four for the same compound. Compare absolute values across different laboratories and the agreement collapses. The relative ordering of compounds holds up; the numbers do not travel.
For both senses the honest summary is a range with a method attached. Anyone quoting a single odour threshold without saying which olfactometer, which panel and which staircase produced it has quoted nothing usable.
What this means for your lab
A lab I visited — the PI, call her Dr. R — had a detection limit that moved by a factor of six between the first draft and the response to reviewers, with no record anywhere of which criterion produced which version. Nobody was cheating. The number had been recalculated by three people who each assumed the definition was obvious.
The avoidance is cheap. Four things belong next to any threshold you publish:
- The criterion: what proportion of trials, what confidence rating, what false-positive rate.
- The blank: how it was measured, how many replicates, and its standard deviation.
- The trial count and the stimulus geometry — duration, area, frequency, retinal location for vision, presentation order for anything behavioural.
- Whether the figure is a detection limit or a determination limit, because they differ by roughly a factor of three and reviewers notice.
Currie's three levels and Tanner and Swets's two parameters amount to the same instruction: separate what the instrument can do from what your decision rule lets you claim. Labs measuring weak signals — spectrometry, or any assay counting rare events — get this wrong more often than others, simply because the numbers are small enough that a factor of four looks like noise.
One concrete next step. Before the next threshold leaves your lab, write a single sentence into the methods naming the criterion, the blank and the trial count, then paste that sentence into the figure legend. If you cannot write it, the number is not ready to publish — and it is better to learn that from your own notebook than from a reviewer.
- absolute threshold
- psychophysics
- signal detection theory
- detection limit
- vision science

