On this page
Theory ↔ Experiment
A strong technical reference should not stop at an equation or a plot. The useful chain is
\[\boxed{ \text{physical hypothesis} \rightarrow \text{model} \rightarrow \text{observable} \rightarrow \text{instrument transfer function} \rightarrow \text{data} \rightarrow \text{residuals} }\]1. Start with the observable
Ask first: what does the instrument actually output?
Examples:
- VNA → complex wave ratio;
- photodetector → photocurrent/voltage, not atomic population directly;
- spectrum analyzer → filtered/detected spectral power;
- oscilloscope → bandwidth-limited voltage waveform;
- atomic receiver → optical transmission/phase/polarization encoding an RF-driven susceptibility.
2. Separate physics from readout
For a Rydberg current experiment, for example:
\[\text{atomic excitation} \rightarrow \text{ionization} \rightarrow \text{particle transport} \rightarrow \text{Shockley–Ramo current} \rightarrow Z_T(\omega) \rightarrow V_{scope}.\]If the final voltage disagrees with theory, the disagreement can live in any layer.
3. Add realism in controlled steps
Do not begin with every complication. Use a ladder:
- ideal analytic model;
- measured geometry;
- measured drive amplitudes;
- finite linewidth/loss;
- spatial averaging;
- instrumental bandwidth;
- temperature/drift;
- interactions/nonlinearity;
- uncertainty propagation.
At each step, ask whether the change improves the particular residual seen in data.
4. Compare shapes before amplitudes
A useful debugging order is:
- peak position;
- symmetry/asymmetry;
- linewidth;
- number of peaks;
- relative amplitudes;
- absolute scale.
Absolute amplitude is often the last thing to trust because it accumulates calibration errors from many layers.
5. Plot residuals
For data $y_i$ and model $m_i$,
\[r_i=y_i-m_i.\]Residual structure is diagnostic:
- random scatter → noise-dominated;
- slope → calibration/background error;
- alternating structure → frequency-axis or phase issue;
- broadened central residual → missing linewidth/spatial averaging;
- systematic side features → missing states/modes/coupling pathways.
6. Use dimensionless comparisons
Before fitting many parameters, compare regime ratios such as
\[\Omega/\Gamma,\qquad \Omega/\Delta,\qquad ka,\qquad \omega_c/\nu_{coll},\qquad R/R_{FF}.\]These ratios often tell you what physics can plausibly matter.
7. Calibrate independently where possible
A parameter measured independently should not also be freely fitted unless you are explicitly testing its calibration.
Examples:
- RF field from AT calibration;
- laser power/waist → optical Rabi estimate;
- photodetector transfer function measured separately;
- magnetic field from atomic resonance;
- cable loss from VNA/power calibration.
8. Falsification tests
A good model should predict what happens when you deliberately change a control parameter that was not used to fit it.
Examples:
- reverse $B$;
- rotate polarization;
- change vapor-cell position;
- change beam waist;
- vary detector bandwidth;
- switch a Floquet replica order/basis size;
- change antenna distance from near to far field.
9. Theory ↔ experiment templates
Antenna
Maxwell/full-wave model → current distribution → far-field pattern → chamber transfer/calibration → measured pattern.
RF receiver
signal source → channel → antenna/front end → gain/noise/nonlinearity → demodulator → EVM/BER.
Atomic sensor
Hamiltonian → density matrix/Floquet → susceptibility → propagation → photodetection → electronics → measured spectrum/beat note.
Charged-particle detector
field map → particle transport → weighting field → induced current → amplifier response → waveform.
10. A reusable comparison figure
For publication-quality theory/experiment comparisons, show whenever possible:
- same axes and units;
- shared color scale;
- measured and simulated linewidths;
- residual panel or difference map;
- independently measured parameters listed in caption/table;
- fitted parameters explicitly labeled;
- uncertainty or repeatability indication.
See Measurements & Instruments, Simulation Library, and Scaling Laws.
Advanced topic — identifiability and experiment design
An excellent numerical fit does not prove the physical mechanism. If measured observables (\mathbf y=f(\boldsymbol\theta)) depend on parameters (\boldsymbol\theta), the sensitivity matrix is (\mathbf J=\partial f/\partial\boldsymbol\theta). Nearly dependent columns reveal parameters that cannot be separately estimated.
With Gaussian measurement noise covariance (\mathbf\Sigma), the local Fisher information is
\[\mathbf F=\mathbf J^T\mathbf\Sigma^{-1}\mathbf J.\]Small eigenvalues of (\mathbf F) reveal weakly identifiable parameter combinations. Deliberately varying field polarization, RF frequency, detuning, temperature or geometry can increase identifiability rather than merely increasing the number of repeated measurements.
Distinguish random measurement noise, calibration uncertainty, parameter uncertainty and model discrepancy. Plot residuals against the independent variables and against time; structured residuals can reveal unmodeled field gradients or drift.
Advanced exercise: fit competing models to one dataset, then select a new measurement condition where their predictions diverge maximally. Use held-out data for model validation.
Further reading: JCGM 100, Guide to the Expression of Uncertainty in Measurement; A. Tarantola, Inverse Problem Theory.