# Experimental physics **Experimental physics** is the part of [[Physics|physics]] that observes and measures, as distinct from [[Theoretical_physics|theoretical physics]], which builds the models the measurements test. Its subject matter is not only the phenomenon under study but the apparatus, the [[Observational_error|errors]] and the [[Statistics|statistics]]: an experimental result is a number *with* an uncertainty, and a number without one is not a result.[^openstax-v1-ch1] The distinguishing skill is knowing which part of a disagreement between measurement and theory belongs to nature and which part belongs to the instrument. In the microsim below the reader runs a laboratory classic — measuring g by timing objects in [[Free_fall|free fall]] — with two dials: N, the number of drops, on a log scale from 3 to 1,000, and σ, the timing error, from 1 to 100 ms. A [[Least_squares|least-squares]] fit of fall distance s against t² returns the slope g/2, and its standard error behaves as `SE_g ∝ sigma/sqrt(N)`.[^spec-p44] For a 1 m drop the fractional error from a single timing is 2σ/t, which is 4.43 % at σ = 10 ms; averaging 100 drops brings that to 0.44 %, or ±0.044 m/s², and reaching ±0.01 m/s² would take about 1,900 drops.[^derived-ep] The third control is the lesson. A "reaction time" toggle adds a fixed 0.15 s to every reading, and no number of drops removes it: the fitted g collapses to 5.53 m/s² from a 1 m drop and does not converge on 9.81 as N grows.[^derived-ep] Random error averages out; systematic error does not. The run is driven by a seeded random-number generator, so reloading the page reproduces the same drops.[^spec-p44] On the [[Physics]] flagship this article is the *Theory and experiment* section of Part III — Research, and it is the root of its own sim rather than a variant of another. Its companions are [[Measurement_uncertainty|measurement uncertainty]], [[Least_squares|least squares]] and [[Observational_error|observational error]] as their own pairs, and [[Computational_physics|computational physics]], which is the third mode alongside theory and experiment. ## Overview Physics is conventionally divided into theory and experiment, and the division is real but leaky. An experimentalist designs an apparatus using theory, and a theorist chooses which calculations to do partly by what can be measured. What makes the experimental side a discipline of its own is that its hardest problems are rarely the ones the textbook poses. The measurement in the sim, timing a falling object, is trivially described by s = ½gt²; everything difficult about actually doing it lies in when the clock starts, how the object is released, whether air resistance matters at the heights used, and how many repetitions are worth the effort. The output of that work is a quoted value and an interval, and the [[Accuracy_and_precision|distinction between accuracy and precision]] governs how the interval is built. Precision is repeatability, and it improves as more data are collected; accuracy is closeness to the true value, and it is limited by effects that repeat identically every time.[^openstax-v1-ch1] A precise and inaccurate result is the characteristic failure mode of a careful experiment with an unnoticed bias, and it is exactly what the sim's reaction-time toggle produces: the scatter shrinks as 1/√N while the centre stays 44 % low.[^derived-ep] ## History Systematic experiment is younger than physics. [[Galileo_Galilei|Galileo]]'s inclined planes were an early example of the central experimental move — arranging a phenomenon so that the quantity of interest can be measured with the instruments available. Timing a body in free fall was beyond a seventeenth-century clock, so the fall was slowed by rolling a ball down a shallow incline until the intervals could be measured by water clock. That is the same reasoning a modern experimenter uses when a signal is moved into a frequency band where the detector is quiet. The [[Galileo's_Leaning_Tower_of_Pisa_experiment|Leaning Tower]] story is the popular version of the work, and rather less certainly historical than the inclined planes. [[Isaac_Newton|Newton]]'s prism work established another pattern: an experiment designed to discriminate between two explanations rather than to display an effect. Through the eighteenth and nineteenth centuries the instrument became the limiting factor and instrument-making became part of the discipline, culminating in interferometry, [[Spectroscopy|spectroscopy]] and precision electrical measurement. By the twentieth century the apparatus had grown beyond what one person could build, and the collaboration — hundreds or thousands of authors, with specialists in detectors, electronics and analysis who may never see the phenomenon directly — became the normal unit of experimental work. ## Current experiments Contemporary experimental physics runs at two extremes of scale. At the large end are facilities built around a single question: the [[Large_Hadron_Collider|Large Hadron Collider]], whose detectors found the [[Higgs_boson|Higgs boson]] and continue to test the [[Standard_Model|Standard Model]]; kilometre-scale interferometers that detect [[Gravitational_wave|gravitational waves]] as length changes far smaller than a proton; and [[Neutrino|neutrino]] observatories buried under rock to shield them from everything except the particle they are built to see. At the small end are tabletop experiments whose precision rivals the large ones: atomic [[Spectroscopy|spectroscopy]], where a caesium transition at 852 nm has a natural width of 5.22 MHz on an optical frequency of 3.5 × 10¹⁴ Hz — about one part in 10⁸, and the line centre can be placed far more sharply than its width — and torsion balances and atom interferometers testing [[Gravity|gravity]] at short range.[^raven-lines][^derived-ep] The two ends share a design logic. Both are limited by noise rather than by signal, and both spend most of their effort on rejecting backgrounds. A Portal Book's particle-data table shows the resolution now routine: the Z boson's width is quoted as 2.4952 GeV and the muon's mean life as 2.19703 μs, five- and six-figure numbers for quantities that cannot be observed directly at all.[^schiller-particles] Numbers of that quality are what allow a theory to be excluded rather than merely disfavoured. They also share a characteristic difficulty: the interesting width is usually far narrower than the width the apparatus imposes. The caesium line at 852 nm has a natural width of about 5 MHz, but thermal motion of the atoms smears the observed profile to roughly 1.2 GHz — some 230 times wider — so the experiment's real subject is removing the broadening rather than measuring the line.[^raven-lines] Almost every current experiment has a version of that problem, and the instrument is designed around it rather than around the phenomenon. ## Method Experimental method is a sequence: decide what to measure, estimate in advance how well it can be measured, take the data, propagate the errors, and compare. The sim compresses that sequence into two sliders, and it is worth walking through what each of them does. The measurement is of g from timed drops. Since s = ½gt², a plot of s against t² should be a straight line through the origin of slope g/2, and fitting that line by [[Least_squares|least squares]] is the standard estimator. A timing error σ propagates into the time-squared axis with a factor 2t, so the fractional error in g from a single drop is 2σ/t. That is 4.43 % for a 1 m drop at σ = 10 ms, 6.26 % from 0.5 m and 3.13 % from 2 m — which is the quantitative form of the advice to drop from as high as the ceiling allows.[^derived-ep] Repetition then buys precision at the standard rate. With σ = 10 ms and a 1 m drop, the error on the fitted g falls from 2.56 % at N = 3 to 1.40 % at N = 10, 0.44 % at N = 100 and 0.14 % at N = 1,000; the residual scatter closes on the 9.81 line as 1/√N, and the sim draws the shrinking band around it.[^derived-ep] The rate is the point. Each extra decimal place costs a hundredfold increase in effort, which is why precision measurement is expensive and why the first question asked of a proposed experiment is what its error budget will be. The reaction-time toggle breaks the rule on purpose. Adding 0.15 s to every reading leaves the scatter unchanged and moves the centre: the fitted g is 4.54 m/s² from 0.5 m, 5.53 from 1 m and 6.43 from 2 m.[^derived-ep] Two lessons follow. Averaging is powerless against it — 1,000 drops give the same wrong answer as three, more precisely. And the diagnostic is not the error bar but the residual plot: because the bias depends on height, the residuals show structure rather than noise, and the apparent g depends on which drops were used. A systematic error is found by making it vary, not by collecting more data. ## Famous experiments The experiments that changed physics mostly share a shape: a simple apparatus, a result that could not be argued away, and a theory that had to move. The interferometer experiment that failed to find the Earth's motion through the supposed light-carrying medium is the classic null result, and it mattered precisely because its sensitivity was well enough understood that "no effect" was a measurement rather than a failure. The [[Double-slit_experiment|double-slit experiment]] gave interference for light and, much later, for electrons, establishing wave behaviour for objects with mass. The [[Photoelectric_effect|photoelectric effect]] gave the opposite result for light — a threshold frequency that intensity could not overcome — and forced the quantum. Scattering experiments form a second family. Firing a beam at a target and measuring the angular distribution of what comes out is how the [[Atomic_nucleus|atomic nucleus]] was found, how nucleon substructure was found, and how [[X-ray_crystallography|crystal structures]] are still determined. The technique's power is that it measures a Fourier transform of the target's structure without requiring a lens, which is why it works at wavelengths where no lens exists. Cosmological measurement is a third: the [[Cosmic_microwave_background|cosmic microwave background]] and the redshift–distance relation of [[Hubble's_law|Hubble's law]] are observations rather than experiments, since the apparatus cannot be adjusted, and they carry their own characteristic problem of separating a signal from foregrounds that cannot be switched off. ## Experimental techniques A working experiment is usually a chain: something converts the phenomenon into an electrical signal, something amplifies it without adding more noise than it must, and something records it. The [[Sensor|sensor]] sets the ceiling on what is possible, and much of the craft lies in choosing one whose noise is small in the band where the signal lives — which is why [[Signal_processing|signal processing]] and [[Noise_(electronics)|electronic noise]] are core experimental subjects rather than peripheral ones. [[Cryogenics|Cooling]] is the most general noise-reduction technique, since thermal noise falls with temperature, and [[Superconducting_magnet|superconducting magnets]] and detectors are among the reasons large facilities are cold. Interference is the standard route to extreme precision in length. Counting fringes as a mirror moves gives displacement in units of half a wavelength, and the same instrument measures a refractive index by counting the fringes a sample introduces: the Portal Book's worked example finds n = 1.0019 for a gas from 19 fringes.[^openstax-v3-ch3][^manual04-optics] Thin-film interference gives thickness on the same scale, with bright and dark bands 122 nm apart for 650 nm light in a soap film.[^manual04-optics] [[Spectroscopy|Spectroscopic]] techniques push further still, turning a frequency measurement into the most precise kind of measurement available, with line shapes whose width is itself the physics — a natural linewidth is the reciprocal of a lifetime, and [[Doppler_broadening|Doppler broadening]] carries the temperature of the sample.[^schiller-particles] [[Simulation|Simulation]] is now part of the apparatus rather than a separate activity, and it brings its own errors. A numerical experiment has an accuracy floor set by the integrator's order and, below that, by the arithmetic: a Portal Book's error study shows the error of each method falling with a slope set by its order until it flattens on a roundoff floor, machine epsilon being 2.2 × 10⁻¹⁶ in double precision and 1.2 × 10⁻⁷ in single.[^anagnostopoulos-error][^manual04-integrator] The floor arrives near 50,000 steps in double precision but at about 18,000 in single, so a simulated measurement can be limited by its own arithmetic exactly as a real one is limited by its instrument.[^anagnostopoulos-error] [[Monte_Carlo_method|Monte Carlo]] methods have the same character: the statistical error of a sampled estimate falls as 1/√N, which is the same rate the sim's drops obey, for the same reason. ## Prominent experimental physicists The people who are remembered as experimentalists tend to be remembered for an instrument as much as for a result. Galileo built telescopes and an inclined plane; Michael Faraday built apparatus for electromagnetic induction from materials he prepared himself; Marie Curie's work on radioactivity rested on an electrometer sensitive enough to turn ionisation into a quantitative assay. In the twentieth century the pattern continued with cloud and bubble chambers, the cyclotron, and the detectors that made [[Standard_Model|particle physics]] possible. The pattern is not an accident of biography. A new instrument opens a region of parameter space that nobody has looked at, and the discoveries follow almost automatically; a new result obtained with an existing instrument is usually harder and less durable. This is also why the boundary between experimental physics and engineering is thin, and why a great deal of experimental work is published as instrumentation papers that never mention a discovery. The modern version of the role is distributed. On a large collaboration the person who designed the sensor, the person who wrote the reconstruction software and the person who set the statistical procedure are different people, and no single author could reproduce the result alone. What has not changed is the criterion by which the work is judged: whether the quoted uncertainty is believable, which in practice means whether the systematic effects have been searched for as hard as the statistical ones have been reduced. ## Timelines A timeline of experimental physics is in practice a timeline of measurement precision. Each era is characterised by the smallest quantity it could detect: the water clock and the inclined plane, the pendulum clock, the interferometer, the photographic plate, the photomultiplier, the transistor amplifier, the charge-coupled device, and the digital acquisition systems now standard. The constants track the same progression, and the endpoint of that progression is a change in status rather than in precision — the [[Speed_of_light|speed of light]] is now fixed by definition at 299,792,458 m/s, so measuring it means calibrating a length standard rather than measuring nature.[^murphy-c] Set against that is the timeline of what was measured: falling bodies, planetary orbits, the [[Electric_charge|electron's charge]], the spectrum of [[Black-body_radiation|black-body radiation]], the nucleus, the [[Laser|laser]], the microwave background, the Higgs boson and gravitational waves. The two lists are the same list read from different ends, because each new measurement was made possible by the instrument immediately before it. The sim sits at the beginning of both lists. Timing a falling body with a hand-held clock is the seventeenth-century experiment, and its error budget — a few per cent per drop, closing as 1/√N, with a reaction-time bias that averaging cannot touch — is the same budget every later experiment has, written at a scale a reader can check by hand.[^derived-ep] What changed over four centuries is the size of σ, not the structure of the argument built on it. ## See also - [[Measurement_uncertainty]] — the interval that makes a number a result - [[Least_squares]] — the estimator the sim fits with - [[Theoretical_physics]] — the other half of the division - [[Observational_error]] — random against systematic, the sim's central lesson - [[Accuracy_and_precision]] - [[Computational_physics]] — the third mode, with its own error floor - [[Scientific_method]] - [[Falsifiability]] ## References [^openstax-v1-ch1]: Sanny, Jeff; Ling, Samuel (2016). *University Physics Volume 1* (OpenStax). Chapter 1, "Units and Measurement" (pp. 17–52): units and dimensional analysis, significant figures, and the treatment of accuracy, precision and measurement uncertainty (page to pin). https://open.umn.edu/opentextbooks/textbooks/university-physics-volume-1 [^anagnostopoulos-error]: Anagnostopoulos, Konstantinos (2016). *Computational Physics: A Practical Introduction to Computational Physics and Scientific Computing (using C++)*, 2nd edition. Chapter 6, "Motion of a Particle": error behaviour of the Euler, Euler–Cromer, Verlet and Runge–Kutta integrators, with machine epsilon 2.2 × 10⁻¹⁶ for double precision and 1.2 × 10⁻⁷ for single, and the roundoff floor setting in near 50,000–100,000 steps in double precision and about 18,000 in single (pp. 202–209; the per-decade error slopes at pp. 224–225). https://open.umn.edu/opentextbooks/textbooks/computational-physics-a-practical-introduction-to-computational-physics-and-scientific-computing-using-c [^manual04-integrator]: Wikitube MICROSIM_GUIDE sub-manual 04, *Atomic, Quantum, Statistical and Electromagnetic Physics*, §6.1 "Integrator error: Euler, Euler–Cromer, Verlet, RK4": the log–log error study against step count with slopes 1, 1, 2 and 4 and the roundoff floor, and the note that a Float32Array state quietly reproduces the book's single-precision failure near 18,000 steps. [^manual04-optics]: Wikitube MICROSIM_GUIDE sub-manual 04, §8.2 "Thin-film interference colour" and §8.3 "Michelson interferometer: fringe counting": the soap-film example at λ = 650 nm giving dark at 0 nm, bright at 122, dark at 244, bright at 366, dark at 488 and bright at 610, with the derived band spacing λ/(4n) = 122.2 nm; and the fringe-counting relations `dd = m*lambda/2` and `N = 2*t*(n - 1)/lambda`. The sub-manual records that most displayed equations in the source extract were lost and standard forms were supplied, so the printed pages should be checked. [^openstax-v3-ch3]: Sanny, Jeff; Ling, Samuel (2016). *University Physics Volume 3* (OpenStax). Chapter 3, "Interference" (pp. 119–144): the Michelson interferometer, fringe counting, the refractive-index example giving n = 1.0019 from 19 fringes, and thin-film interference (page to pin). https://open.umn.edu/opentextbooks/textbooks/university-physics-volume-3 [^schiller-particles]: Schiller, Christoph. *Motion Mountain, Vol. IV: The Quantum of Change*. Particle data: Z width 2.4952 GeV, W width 2.124 GeV, muon mean life 2.19703 μs, tau lifetime 290.6 fs, and the width–lifetime relation Γτ = ħ with t½ = τ ln 2 (pp. 128–129). https://open.umn.edu/opentextbooks/textbooks/the-adventure-of-physics-vol-iv-the-quantum-of-change [^raven-lines]: Raven, Will (2025). *Atomic Physics for Everyone*. Chapters 3 and 4: the low-power Lorentzian lineshape and natural linewidth, with the caesium 852 nm line at γ ≈ 5.22 MHz (pp. 55–70), and the Monte-Carlo construction of a Doppler profile from N atoms, giving a thermally broadened width of roughly 1.2 GHz (pp. 79–86). https://open.umn.edu/opentextbooks/textbooks/atomic-physics-for-everyone-an-introduction-to-atomic-physics-quantum-mechanics-and-precision-spectroscopy-with-no-college-level-prerequisites [^murphy-c]: Murphy, Thomas (2021). *Energy and Human Ambitions on a Finite Planet*. Chapter 15, the speed of light c = 2.99792458 × 10⁸ m/s (p. 266). https://open.umn.edu/opentextbooks/textbooks/energy-and-human-ambitions-on-a-finite-planet [^derived-ep]: Computed for this article from s = ½gt² with g = 9.81 m/s², propagating a timing error σ into g through the fractional relation 2σ/t. Fall times: 0.3193 s from 0.5 m, 0.4515 s from 1.0 m, 0.6386 s from 2.0 m. Single-drop fractional error at σ = 10 ms: 6.26 % (0.5 m), 4.43 % (1.0 m), 3.13 % (2.0 m); at σ = 1 ms, 0.626 %, 0.443 % and 0.313 %; at σ = 100 ms, 62.6 %, 44.3 % and 31.3 %. Averaging N drops from 1.0 m at σ = 10 ms: 2.557 % (N = 3), 1.401 % (10), 0.443 % (100), 0.140 % (1,000), i.e. ±0.251, ±0.137, ±0.0435 and ±0.0137 m/s²; reaching ±0.01 m/s² requires N ≈ 1,890. Caesium line at 852 nm: f = c/λ = 3.518 × 10¹⁴ Hz, so the natural width of 5.22 MHz recorded in [^raven-lines] is 1.48 × 10⁻⁸ of the frequency, and the Doppler-broadened 1.2 GHz is 230 times that width. Systematic offset of 0.15 s added to every reading: the apparent g = 2s/(t + 0.15)² is 4.541 m/s² from 0.5 m, 5.527 from 1.0 m and 6.433 from 2.0 m, errors of −53.7 %, −43.7 % and −34.4 % that do not shrink with N and that show as height-dependent structure in the residuals. [^spec-p44]: Matter & Energy Cluster contract, `_registry/plans/PHYSICS_SECTIONS.md` row P44: new root sim, a law fitted to noisy data. Free-fall timings carry a timing error σ over N drops; the reader sets N on a log scale from 3 to 1,000 and σ from 1 to 100 ms. The least-squares slope of s against t² gives g with its standard error `SE_g ∝ sigma/sqrt(N)`, the residual plot and the 9.81 line close the gap as 1/√N, and a "reaction time" toggle adds a 0.15 s offset that no number of drops removes — random error averages out, systematic error does not. The generator is seeded (`hydro.util.rng(seed)`) so a reload reproduces the run, and the least-squares step is closed form, so the sim runs live. ## Further reading - Sanny, Jeff; Ling, Samuel (2016). *University Physics Volume 1*, Chapter 1 — units, significant figures and uncertainty at first-course level. https://open.umn.edu/opentextbooks/textbooks/university-physics-volume-1 - Anagnostopoulos, Konstantinos (2016). *Computational Physics*, 2nd edition, Chapter 6 — integrator error, roundoff floors and how a numerical experiment acquires its own instrument limit. https://open.umn.edu/opentextbooks/textbooks/computational-physics-a-practical-introduction-to-computational-physics-and-scientific-computing-using-c ## External links - The Wikipedia pair's *External links* section is the place to look for laboratory guides, collaboration sites and timelines; this page lists only the open texts above, which it has read. <!-- MATTERSIM:BEGIN g24 — Matter & Energy Cluster microsim (framework build, specs/sims/Experimental_physics.json); do not hand-edit inside --> **Microsim — three.js (Wikitube framework), pending deploy:** *Experimental physics* will play here once `https://wikitube-3d-microsims.netlify.app/matter/Experimental_physics.html` is live. <!-- pending: <div class="wt-sim" data-src="https://wikitube-3d-microsims.netlify.app/matter/Experimental_physics.html" data-title="Experimental physics"></div> --> <!-- MATTERSIM:END --> ## Wikipedia : Wikitube **Strict pair:** [Wikipedia](https://en.wikipedia.org/wiki/Experimental_physics) : [Wikitube](https://en.wikitube.io/wiki/Experimental_physics) · pinned revision [1368387879](https://en.wikipedia.org/w/index.php?oldid=1368387879) · 2026-09-11 ## Previous hub tags Hubs: `Life_Physics`. Portals: [[PORTAL_Physics]]. --- *Matter & Energy Cluster child articles, wave 1 · 2026-09-11 · drafted · Physics row P44 · sim pending (matter/Experimental_physics).*