Glossary: the notions of this site
Each notion used on the site is defined here, in the words of the pages that detail it.
67 notions, from persistence to the p-value. Each entry gives the definition, the unit, related notions and the page that explains it. The same glossary exists as machine-readable files.
Definitions follow the wording of the site pages. They describe a research and teaching use, not a medical recommendation.
Domains
Forecasting and evaluation
An error only makes sense against a reference: describe the signal, choose simple references, measure the gain, then add complexity only if the gain justifies it.
AR(2) (second-order autoregression)
A regression on the last two known values. Raw: without removing the cycle. It measures what removing the cycle brings.
ARTU (second-order reference)
A second-order reference on the ratio to trend: two autocorrelations and a table give its three weights, on the last value, the value h steps earlier and the mean.
BLEND (cyclostationary persistence)
The convex combination of persistence and cyclic persistence: two positive weights adding up to 1, λ and 1 − λ, computed at each moment of the cycle.
Brownian motion
Each value is the previous one plus a random draw: memory φ = 1. Persistence wins on it.
Clear-sky index (ratio to the clear-sky trend)
In solar forecasting, the ratio to the clear-sky trend: irradiance divided by clear-sky irradiance.
CLIPER
The convex combination of persistence and the mean. Same idea as BLEND, another partner.
Convex combination (positive weights summing to 1)
Two forecasts weighed by positive weights adding up to 1, computed on the past. Persistence and the mean: CLIPER. Persistence and cyclic persistence: BLEND.
Cyclic persistence (cyclic persistence)
The forecast copies the value one cycle earlier, at the same moment of the cycle.
Deseasonalization (removing the cycle, seasonal adjustment)
A known cycle is not forecast: remove it, forecast the rest, put it back.
Difference to trend (difference to trend)
We study y − c, the gap to the cycle c. Suited to additive noise.
Extreme learning machine (ELM) (ELM, frugal learner)
The Extreme Learning Machine retains a random hidden layer and fits its output coefficients analytically, avoiding iterative deep-network training.
Forecast evaluation (forecast verification)
An error only makes sense against a reference: same targets, stated scope, several error norms, then a test of the difference.
Forecast horizon (horizon)
The number of steps between the last known value and the value to forecast.
Forecastability F (F indicator, forecastability)
F = 1 + sCV × (ρmax − 1), between 1 − sCV and 1. The higher it is, the easier that horizon is to forecast.
γ-index (gamma index)
For each forecast pixel, the distance to the nearest observation in space, time and intensity, each divided by its tolerance. γ ≤ 1: accepted.
MAE (mean absolute error)
The mean absolute error: the average of the absolute gaps between forecast and observation.
MALR (mean absolute log return) (mean absolute log return)
A parameter borrowed from econometrics that describes the size of relative changes. This empirical diagnostic does not replace testing the model at the intended site and horizon.
Mean (mean of the past, climatological mean)
The forecast is always the mean of the past.
Memory φ (noise memory)
The correlation between two neighbouring values of the noise. 0: none. 1: total.
MIMO (multiple-input multiple-output) (MIMO-ELM)
A multiple-input multiple-output architecture: one model reads the recent history of several signals and forecasts them together.
NICE (NICEᵏ, NICEΣ, NICE index)
The forecast error divided by that of persistence, at the same dates and horizon. 0: perfect. 1: as much error as persistence. Above 1 it is written “> 1”: worse than persistence.
nMAE (normalized MAE)
The mean absolute error divided by the mean of the signal.
“No free lunch” (no free lunch theorems)
Under a uniform average over all functions of a finite problem, the average performance of the search algorithms considered coincides. This precise framework is not the distribution of real solar problems.
nRMSE (normalized RMSE)
The RMSE divided by the mean of the signal (0 for a perfect forecast).
P10–P90 band (P10 and P90, median and P10–P90 band)
P10 and P90 are the 10th and 90th percentiles of the simulated results: the P10–P90 band contains the central 80 %. It describes variation between runs under the chosen parameters, not all uncertainty about a real system.
Persistence (simple persistence, naive forecast)
The forecast copies the last known value. It is the minimal reference of any forecast.
Ratio to trend (ratio to trend)
We study y / c. Suited to multiplicative noise. With c as clear sky: the clear-sky index.
RMSE (root mean squared error)
The root mean squared error. It gives large errors more weight than the MAE.
sCV (stochastic coefficient of variation)
The root of the mean squared departure from clear sky, divided by the root of the mean square of clear sky. 0: clear sky. About 1: each hour clear or overcast on a coin toss. It is not a maximum.
Simulated data (synthetic data)
When there are no measurements yet, no model can be trained or validated, but what is known can be simulated: a first estimate, not a measurement.
Smart persistence (clear-sky-index persistence)
Persistence applied to the clear-sky index (SP: persistence of the clear-sky index), rather than to raw irradiance.
Tunnel Load Simulator (TLS) (TLS, tunnel load simulator)
An open-source application that builds the electricity demand of a road tunnel from its geometry, traffic, lighting and ventilation, says how uncertain it is and states its limits.
Radiotherapy
When the link between a radiotherapy schedule and its biological effect is known, it is computed directly with an explicit model, with no training data.
Biologically effective dose (BED) (BED, biological effective dose)
Dose in grays that makes different fractionations comparable. For n fractions of dose d, the LQ model gives BED = n · d · (1 + d / (α/β)).
Doubling time (Td)
The time, in months, for the number of tumour cells to double without treatment: 9.8 months for the refractory profile, 66.6 for the responder profile in the paper.
Equivalent dose in 2 Gy fractions (EQD2) (EQD2)
The translation of the BED into a dose given in 2 Gy fractions: EQD2 = BED / (1 + 2/(α/β)).
Fractionation (fractions, treatment schedule)
For n fractions (the number of treatment sessions), each delivering a dose d in grays, physical dose is not enough to predict the biological effect, which also depends on the dose received at each fraction.
Gray (Gy)
The unit of absorbed physical dose: the energy deposited per kilogram of tissue.
Linear-quadratic-linear model (LQL) (LQL)
The same curve as LQ up to the transition dose, then a straight line. Below the transition dose, the two models are identical.
Linear-quadratic model (LQ) (LQ)
The BED of a fraction grows along a curve with dose, at all doses: the LQ curve keeps bending upward as the dose per fraction grows.
LQL-Equiv (BED and EQD2 calculator)
A free, open-source (MIT) calculator of BED and EQD2 under the LQ and LQL models, in the browser. A research and teaching tool, not a medical device.
α/β ratio (α/β)
A tissue-specific parameter: it describes how sensitive a tissue is to the size of each fraction. The lower it is, the more sensitive the tissue is to the dose per fraction (about 3.9 Gy for the rectum).
Thyroglobulin (Tg) (Tg)
The blood marker followed by the radioiodine therapy model.
Transition dose dT (dT)
The dose per fraction at which the LQL model changes from a curve to a straight line (7.8 Gy for the rectum). In the LQL-Equiv library it equals 2 α/β.
Medical physics
How a beam of photons or neutrons crosses matter and deposits energy in it, from simple models to transport simulations.
Attenuation (exponential decay)
Each thickness equal to the half-value layer halves what remains: it is an exponential decay, not a subtraction.
BNCT (boron neutron capture therapy) (boron neutron capture therapy)
A compound carries boron-10 to the target tissue; when this boron captures a neutron, an alpha particle and a lithium-7 nucleus deposit their energy over a few micrometres.
Compton scattering (Compton effect)
The scattering of a photon on an electron: the photon loses energy and the electron receives a kinetic energy K. The rest energy of the electron is 511 keV.
Cross section (neutron cross section, σ(E))
It says how readily a nucleus reacts with a neutron of energy E. It is expressed as an area, often in barns, without being the size of the nucleus or, by itself, the probability of capture.
Half-value layer (HVL) (HVL, half-value thickness)
The thickness of material that halves the intensity transmitted without scattering: HVL = ln 2 / μ, with μ the linear attenuation coefficient.
Macroscopic cross section (Σ)
The cross section multiplied by the number of nuclei per unit volume: an inverse length. For a mixture, Σ adds these products nucleus by nucleus.
Mean free path (L = 1/Σ)
The inverse of the macroscopic cross section, L = 1/Σ: the average distance between two collisions.
Monte Carlo (Monte Carlo method, Monte-Carlo)
Simulating many times with different random draws. More histories reduce the statistical error, as 1/√N for N histories; they correct neither a wrong geometry nor wrong physical data.
Statistics
Compare a difference with its uncertainty, measure an association or a prediction, and read a result without over-interpreting it.
Chi-square test (chi-squared test, χ²)
The statistic is the sum of the contributions (O−E)²/E, with O the observed count and E the count expected under the tested hypothesis. A small p-value says the departure is hardly compatible with the hypothesis, not that it is large.
Coefficient of determination R² (R²)
The share of variation described by the line (1 for a perfect forecast). It does not measure the importance of the association.
Confidence interval
It brackets the mean of y at x₀ (95 % band, Student quantile with n − 2 degrees of freedom).
Influential point (leverage)
A point that moves the least-squares line; leverage is its potential weight, greater the farther it lies in x.
Normal distribution (Gaussian distribution, z-score)
The z-score is the number of standard deviations between x and the mean; Φ(z) is the probability that a standard normal variable stays below z: the area to the left of z.
p-value (p value)
The probability, if the means are equal (or ρ = 0) and the model is correct, of a result at least as extreme. It is not the probability that the hypothesis is true; it does not measure the importance of an association.
Pearson correlation (correlation coefficient r)
The Pearson correlation in the scatter (r) and in the population (ρ).
Prediction interval
It brackets a new observation at x₀: add 1 under the root, hence a band wider than the confidence interval.
ROC curve (AUC, sensitivity and specificity)
Sensitivity against the false-positive rate 1 − Sp as the threshold varies. The area under the curve (AUC) summarizes the test without choosing a threshold.
Student t-test (t-test)
The test divides the difference between two means by its standard error: the same difference can stand out from sampling chance or drown in it. Significant does not mean important.
Thermal physics and thermodynamics
Heat exchange, engine and refrigeration cycles and Carnot limits, to size a cooling, a freezing or a heat engine.
Biot number (Bi)
Internal conduction resistance compared with surface convection resistance.
Carnot efficiency (Carnot engine, Carnot limit)
Efficiency η is the work W divided by the heat received Qhot; ηCarnot is its limit, reached by a reversible cycle, with no friction and no temperature gap during exchanges.
Coefficient of performance (COP) (COP)
Heat extracted from the cold reservoir per unit of work supplied. COP > 1 is normal: it is not a percentage efficiency. Carnot limit: the maximum COP of a reversible machine between two reservoirs.
Freezing (freezing energy)
The energy to remove to freeze a mass of water: sensible heat (mcΔT) plus latent heat, with no temperature change in this model; average power is energy divided by duration.
Latent heat
The heat removed to freeze the water, with no temperature change in this model.
Sensible heat
The heat exchanged when temperature changes: the term mcΔT, with c the specific heat capacity and ΔT the temperature change.
Files for machines
Turtle (SKOS) · JSON-LD (SKOS) · See the ontology · French version