# Measurement uncertainty (GUM and Monte Carlo)

> One page of the phonometry documentation, as Markdown. Index: https://docs.jmrp.io/phonometry/llms.txt

Canonical: https://docs.jmrp.io/phonometry/signals/metrology/gum-uncertainty/
Language: en
Alternate: https://docs.jmrp.io/phonometry/es/signals/metrology/gum-uncertainty/index.md
Version: phonometry 4.0.0
Install: pip install "phonometry[full]"
Author: José Manuel Requena Plens (https://orcid.org/0000-0003-1250-6212)
Published: 2026-07-09
Updated: 2026-10-21T12:42:48Z
Standards: JCGM 100:2008, the GUM; JCGM 101:2008
Verified: 7 conformance checks (JCGM 100:2008; JCGM 101:2008) https://docs.jmrp.io/phonometry/reference/conformance/#measurement-uncertainty-gum-supplement-1
License: MIT (https://github.com/jmrplens/phonometry/blob/main/LICENSE)
Summary: GUM law of propagation (ISO/IEC Guide 98-3:2008) and Monte Carlo (Supplement 1): combined and expanded uncertainty, Type B, coverage intervals.

---

Every measurement result is incomplete without a statement of its uncertainty.
The *Guide to the Expression of Uncertainty in Measurement* gives two ways to
propagate the uncertainties of the inputs of a measurement model
$y = f(x_1, \dots, x_N)$ to the output: the **law of propagation of
uncertainty** (**JCGM 100:2008**, the GUM, published also as
ISO/IEC Guide 98-3:2008, clause 5) and the **Monte Carlo method**
(**JCGM 101:2008**, Supplement 1, published also as ISO/IEC Guide 98-3-1:2008,
clause 7). Four designations, two documents; the rest of this page says "the
GUM" and "Supplement 1".
The first combines standard uncertainties analytically through the sensitivity
coefficients; the second propagates the whole probability distributions
numerically. They agree for linear models and diverge, informatively, when the
model is non-linear or the inputs are far from Gaussian.

![From a shared measurement model and its input estimates and standard uncertainties, two parallel lanes: the GUM law of propagation (sensitivity coefficients, combined in quadrature, Welch-Satterthwaite effective degrees of freedom, expanded uncertainty U equals k times uc) and the Monte Carlo method (draw each input from its PDF, propagate M trials, take the probabilistically symmetric 95 percent coverage interval)](https://docs.jmrp.io/phonometry/media/diagram_uncertainty.svg)

## 1. The law of propagation of uncertainty (GUM clause 5)

For a model $y = f(x_1, \dots, x_N)$ with uncorrelated inputs, the combined
standard uncertainty is the quadrature sum of the input contributions
$u_i(y) = |c_i|\,u(x_i)$ weighted by the sensitivity coefficients
$c_i = \partial f / \partial x_i$:

$$
u_\mathrm{c}^2(y) = \sum_{i=1}^{N} \left(\frac{\partial f}{\partial x_i}\right)^2 u^2(x_i).
$$

`uncertainty_propagation` evaluates the sensitivities by central differences, so
almost any callable model works, with no hand-derived partials. Input quantities
are described by a `Quantity` (best estimate, standard uncertainty, PDF, degrees
of freedom).

The step of that difference is the input's *own* standard uncertainty, which is
what the GUM recommends (5.1.4 NOTE 2), so the coefficient is the average slope
over the range the input actually explores rather than the tangent at the
estimate. For a smooth model that is a feature. It breaks on models that are not
smooth over that range, and the failure is silent: a `max()`, a `min()`, or a
threshold such as a clamp at a physical bound or a pass/fail correction has one
derivative on one side and another on the other; a table lookup or a rounded
intermediate can return a sensitivity of exactly zero; a model that draws its own
random numbers returns pure noise. Two defences: compare `result.sensitivities`
against the analytic partials wherever you know them, and reach for `monte_carlo`
whenever the model is non-smooth, because propagating distributions needs no
derivative at all. The third panel of the figure in section 2 is exactly such a
model.

### Type A: uncertainty from repeated observations

The GUM's first half is the one this page's API makes easiest to skip. Given $n$
independent readings of the same quantity with experimental standard deviation
$s$, the best estimate is their mean and the standard uncertainty **of that
mean** is

$$
u = \frac{s}{\sqrt{n}}, \qquad \nu = n - 1
$$

(GUM 4.2.3): exactly the pair a `Quantity` takes as `uncertainty` and `dof`.
The Position input of the budget below is that: ten measurements at different
microphone positions scattering with $s = 1.1$ dB give $u = 1.1/\sqrt{10} =
0.35$ dB with `dof=9`.

Two consequences worth internalising. First, $s$ and $u$ describe different
things: $s$ is a property of the sound field and does not shrink when you
measure more positions, while $u$ does, so *more positions* is the only lever
on this term, and it is a $\sqrt{n}$ lever. Second, a small $n$ is expensive
twice: once through $\sqrt{n}$, and once through the coverage factor, because
9 degrees of freedom already lift $k$ from 1.96 to 2.11 and 4 degrees of freedom
lift it to 2.78. Entering the raw $s$ instead of $s/\sqrt{n}$, or leaving `dof`
at its infinite default, are the two commonest ways to get this term wrong, and
they pull in opposite directions.

### Type B: uncertainty from a stated bound

Type B evaluations start from a half-width $a$ (a tolerance, a resolution, a
manufacturer's limit) and a choice of distribution over it (GUM clause 4.3):

| Constructor | $u$ | Use when |
| :--- | :---: | :--- |
| `rectangular(x, a)` | $a/\sqrt{3}$ | only bounds are known and nothing inside them: a class tolerance, a display resolution, a stated limit. The honest default |
| `triangular(x, a)` | $a/\sqrt{6}$ | values near the centre are demonstrably more likely than values at the edges: the sum of two comparable rectangular influences, a quantity that was centred by adjustment |
| `u_shaped(x, a)` | $a/\sqrt{2}$ | the influence *cycles* through its range and spends most of its time at the extremes: a thermostatted room swinging between switching points, a standing-wave field sampled at an arbitrary position |

The two Type B inputs of the budget below are both bounds, hence both
rectangular: the 0.20 dB calibration half-width is the calibrator's class
tolerance together with the pre/post drift observed in the field check, and the
0.30 dB instrument half-width is the meter's class tolerance at the reference
frequency. The choice of distribution reaches the GUM result *only* through $u$,
so the same half-width read as triangular would simply divide by $\sqrt{6}$; but
it sets the shape of the Monte Carlo output entirely, which is why the two
methods agree on this budget and would not if one rectangular term dominated it
(shown in the first panel of the figure in section 2).

```python
from phonometry import metrology

# A-weighted level: a reading plus zero-mean calibration, instrument and
# positional corrections. The model is their sum.
quantities = [
    metrology.Quantity(value=74.0, uncertainty=0.0, name="Reading"),
    metrology.rectangular(0.0, 0.20, name="Calibration"),
    metrology.rectangular(0.0, 0.30, name="Instrument"),
    metrology.Quantity(value=0.0, uncertainty=0.35, dof=9, name="Position (Type A)"),
]
result = metrology.uncertainty_propagation(lambda a, b, c, d: a + b + c + d, quantities)

print(round(result.value, 2))                 # 74.0
print(round(result.combined_uncertainty, 3))  # 0.407 dB
print(result.contributions.round(3))          # [0.    0.115 0.173 0.35 ]
print(round(result.effective_dof, 1))         # 16.5

expanded = result.expanded(0.95)
k, U = expanded.coverage_factor, expanded.expanded_uncertainty
print(round(k, 2), round(U, 2))               # 2.11 0.86  ->  Y = 74.0 ± 0.9 dB
```

### Reading the budget you just computed

A budget is a decision aid, not a certificate, and this one says something
specific in one glance. The position term contributes 0.350 dB of the 0.407 dB
total, which is **74 % of the variance** on its own; instrument and calibration
together are the remaining 26 %; the reading itself contributes nothing, because
it carries no uncertainty of its own in this model. The measurement is limited by
*where the microphone was*, not by the instrument.

Quadrature makes that lesson quantitative and slightly brutal. Removing the
calibration term altogether (assuming a perfect calibrator with no drift) moves
$u_\mathrm{c}$ from 0.407 to 0.390 dB, a gain of 0.017 dB for an impossible effort. Halving the
position term instead, which means four times as many positions, moves it to
0.272 dB. Budgets are improved at the top of the bar chart and never at the
bottom, and the bar chart is there to tell you which end is which.

Finally, calibrate your expectations for the number itself. An expanded
uncertainty near 1 dB is the ordinary outcome for a class 1 chain and a handful
of positions. Standardized budgets for environmental and occupational
measurements land between 1 and 3 dB once source, meteorological and sampling
terms enter. Anything much below half a decibel should prompt a search for the
term that was left out rather than satisfaction.

![An outdoor A-weighted level measurement drawn to scale: a source, a sound level meter on a tripod with the microphone at 1.5 m above the ground and 4 m from a façade, three alternative microphone positions with a spread bracket, the calibrator beside the meter and a weather station; leader lines connect each element to the rows of a budget table naming its evaluation type, distribution and degrees of freedom, with the calibrator term marked as the one shared between channels](https://docs.jmrp.io/phonometry/media/diagram_uncertainty_sources.svg)

Every row of a budget is a piece of hardware or a decision about geometry, and
budgets go wrong by omission far more often than by arithmetic: the forgotten
meteorological term, the forgotten mounting or directivity term, the calibrator
term counted twice as though the two channels had been calibrated independently.

### Coverage factor and effective degrees of freedom

The **expanded uncertainty** $U = k\,u_\mathrm{c}$ scales the combined uncertainty by a
coverage factor $k = t_p(\nu_\text{eff})$ from the $t$-distribution at the
**effective degrees of freedom** given by the Welch–Satterthwaite formula
(Annex G.4). Here the single Type A input (9 degrees of freedom) pulls
$\nu_\text{eff}$ down to 16.5, so $k = 2.11$ rather than the large-sample 1.96.
This chain reproduces the GUM's own worked examples end to end: the Annex H.1
end-gauge budget ($u_\mathrm{c} = 31.7$ nm, $U_{99} = 92$ nm against the printed 32/93)
and the Annex H.2 correlated resistance measurement ($u_\mathrm{c}(R) = 0.071\ \Omega$,
$u_\mathrm{c}(X) = 0.295\ \Omega$, $u_\mathrm{c}(Z) = 0.236\ \Omega$ from Table H.3).

`expanded(0.95)` implements the GUM's own rule: a two-sided interval with $k$
taken from the $t$-distribution at $\nu_\text{eff}$. Acoustics standards
routinely fix $k$ by convention instead: $k = 2$ for a two-sided 95 % statement
when the degrees of freedom are large, and $k = 1.65$ for the **one-sided** 95 %
statement used when a level is compared against a limit and only exceedance
matters. The same $u_\mathrm{c} = 0.407$ dB then gives three different expanded
uncertainties:

| Convention | $k$ | $U$ |
| :--- | :---: | :---: |
| GUM, two-sided, $t$ at $\nu_\text{eff} = 16.5$ | 2.11 | 0.86 dB |
| Conventional two-sided, large sample | 2 | 0.81 dB |
| One-sided, against a limit | 1.65 | 0.67 dB |

A compliance decision can turn on that choice alone, so the convention belongs
next to every $U$ you quote. `coverage_factor_override` is how to impose a
standard's prescribed $k$, and it is also the only route when the budget is
correlated, for the reason below.

### Correlated inputs

Correlated inputs are handled by passing a correlation matrix; a fully
correlated sum then adds linearly instead of in quadrature. Correlation is the
classic silent error in a budget: two corrections traceable to the *same*
calibrator, or two channels sharing one instrument, do not average away the way
independent terms would, and combining them in quadrature as if they were
uncorrelated typically understates $u_\mathrm{c}$. With sensitivities of opposite sign
the bias points the other way, and a fully correlated pair can cancel exactly.

![Left: the combined standard uncertainty of two 0.3 dB terms against their correlation coefficient, rising from zero at minus one through the quadrature value 0.424 at zero to 0.600 at plus one for same-sign sensitivities, and the mirrored curve for opposite signs. Right: the same budget as two bars, 0.424 dB assumed independent against 0.600 dB traceable to the same calibrator, annotated as 41 percent understated](https://docs.jmrp.io/phonometry/media/uncertainty_correlation.svg)

*Two terms of 0.3 dB each, combined by `uncertainty_propagation` with a correlation
matrix. Treating a fully correlated pair as independent understates $u_\mathrm{c}$ by
41 % (0.424 dB instead of 0.600 dB) and the error grows with the number of
terms that share a source. Reading the same curve backwards: a pair with
$\rho = -1$ and same-sign sensitivities cancels to zero, which is what a
differential measurement through one instrument buys you.*

Because the GUM defines Welch–Satterthwaite for *independent* inputs only, a
correlated budget with finite input degrees of freedom carries no effective
degrees of freedom at all. What to do depends on the budget:

| The budget | `effective_dof` | What to do |
| :--- | :--- | :--- |
| Uncorrelated | Welch–Satterthwaite value | nothing: `expanded(p)` takes $k$ from $t$ at $\nu_\text{eff}$ |
| Correlated, every input Type B with infinite dof | $\infty$ | nothing: $k$ comes from the normal distribution |
| Correlated, any input with finite dof | `nan`, with a warning | pass an explicit $k$: `expanded(0.95, coverage_factor_override=2.0)` |

```python
# Same four inputs, now with the calibration and instrument terms fully
# correlated because both are traceable to the same calibrator.
correlation = [
    [1.0, 0.0, 0.0, 0.0],
    [0.0, 1.0, 1.0, 0.0],
    [0.0, 1.0, 1.0, 0.0],
    [0.0, 0.0, 0.0, 1.0],
]
correlated = metrology.uncertainty_propagation(  # warns: no Welch-Satterthwaite here
    lambda a, b, c, d: a + b + c + d, quantities, correlation=correlation)
print(round(correlated.combined_uncertainty, 3))            # 0.454 dB (was 0.407)
print(correlated.effective_dof)                             # nan
print(round(correlated.expanded(0.95, coverage_factor_override=2.0).expanded_uncertainty, 2))  # 0.91
```

Counting the shared calibrator once instead of twice costs 0.047 dB here, on a
budget where the two Type B terms are small. On a budget where they dominate
(such as a two-channel intensity measurement or a level difference between two
rooms measured with one meter), the same mistake is the 41 % of the figure above.

## 2. The Monte Carlo method (Supplement 1)

When the model is non-linear or the inputs are markedly non-Gaussian, the GUM
Gaussian assumption for the output can be inaccurate. `monte_carlo` instead
draws samples of each input from its PDF, evaluates the model over all trials
and reports the mean, the standard deviation and the **probabilistically
symmetric coverage interval** (equal probability in each tail, clause 7.7).
The inputs are sampled independently; the Supplement's multivariate-Gaussian
path for non-independent quantities (6.4.8) is not implemented, so correlated
budgets belong to `uncertainty_propagation`. The number of trials is fixed (no
adaptive 7.9 procedure, at least 2 trials) and the interval is the symmetric
one, not the 5.3.4 shortest interval.

```python
from phonometry import metrology

quantities = [
    metrology.Quantity(value=74.0, uncertainty=0.0, name="Reading"),
    metrology.rectangular(0.0, 0.20, name="Calibration"),
    metrology.rectangular(0.0, 0.30, name="Instrument"),
    metrology.Quantity(value=0.0, uncertainty=0.35, dof=9, name="Position (Type A)"),
]
mc = metrology.monte_carlo(lambda a, b, c, d: a + b + c + d, quantities,
                    n_trials=1_000_000, coverage=0.95, seed=1)

print(round(mc.value, 2))                 # 74.0
print(round(mc.standard_uncertainty, 3))  # 0.407 dB  (matches uc above)
print([round(x, 2) for x in mc.interval]) # [73.2, 74.8]
```

For this near-linear model the Monte Carlo standard uncertainty reproduces the
GUM $u_\mathrm{c}$ to three digits and the 95 % interval matches $Y \pm U$. The two
methods are validated against the Guides' own worked examples: the additive
model of four unit inputs gives $u_\mathrm{c} = 2.0$ (Supplement 1 clause 9.2), and four
rectangular inputs give a Monte Carlo interval of $[-3.88,\, 3.88]$
(Supplement 1 clause 9.2.3).

When does the Monte Carlo method earn its extra cost? Whenever either of the
GUM's two simplifications fails: the model is replaced by its first-order
expansion, and the output distribution by a Gaussian (or a $t$). Both hold
well for the additive level model above, which is why the two methods agree
to three digits. They stop holding when the model is strongly non-linear over
the span of the input uncertainties (energy-to-level conversions with wide
inputs, products and quotients with large relative uncertainties), when a
single non-Gaussian input dominates the budget (one large rectangular term
makes the output nearly rectangular, and a Gaussian $\pm\,2u_\mathrm{c}$ interval
overcovers it), or when the output sits near a physical bound (an absorption
coefficient near 0 or 1, a level correction that cannot cross zero), where
the true coverage interval is asymmetric and no $Y \pm U$ statement can
represent it. In those regimes the Monte Carlo interval is the reference:
Supplement 1 (clause 8) treats the GUM framework as validated precisely when
it agrees with the Monte Carlo result, and as superseded by it when it does
not.

Those three regimes are worth seeing rather than being told about, because a
reader who has never watched a GUM interval fail cannot recognise the failure in
a budget of their own:

![Three stacked panels, each a Monte Carlo histogram with the GUM Gaussian overlaid, the Monte Carlo coverage interval shaded and the GUM Y plus or minus U interval drawn as dashed rules. Top: a dominant rectangular input gives a nearly rectangular output that the Gaussian interval overcovers. Middle: the energy sum of two levels gives a skewed distribution whose mean is offset from the GUM value. Bottom: an absorption coefficient clamped at one gives an asymmetric interval that no symmetric statement can represent](https://docs.jmrp.io/phonometry/media/uncertainty_gum_vs_mc.svg)

*The three regimes, each computed with `uncertainty_propagation` and
`monte_carlo` on the same inputs. **Top**, one dominant rectangular term
($a = 1$ dB) beside a small Gaussian one: the output is nearly rectangular, and
the GUM interval $\pm 1.15$ dB overcovers the Monte Carlo $[-0.98, 0.98]$ by
17 %. **Middle**, the energy sum $10\log_{10}(10^{L_1/10} + 10^{L_2/10})$ of two
levels with $u = 3$ dB each: the Monte Carlo distribution is skewed and its mean,
73.48 dB, sits 0.47 dB above the value the GUM propagates through the model,
73.01 dB: the first-order expansion has lost a real bias. **Bottom**, an
absorption coefficient of 0.95 with $u = 0.06$ against the physical bound at 1:
the GUM interval reaches 1.06, which does not exist, while the Monte Carlo
interval $[0.83, 1.00]$ is asymmetric and no $Y \pm U$ statement can express it.
In all three, Supplement 1 clause 8 says the Monte Carlo result is the reference.*

![Two panels for the A-weighted level example. Left: the GUM uncertainty budget, a horizontal bar chart of each input's contribution to the combined uncertainty with a dashed line at uc of 0.407 dB. Right: the Monte Carlo output histogram overlaid with the GUM Gaussian and the shaded 95 percent coverage interval; the title reads Y equals 74.00 dB, U equals 0.86 dB, k equals 2.11](https://docs.jmrp.io/phonometry/media/uncertainty_budget.svg)

*Left is where the money goes: the ten positions carry 73.9 % of the variance
on their own, the instrument 18.1 % and the calibrator 8.0 %, so $u_\mathrm{c}$ =
0.407 dB is a statement about the survey design and not about the meter.
Halving the instrument term would move $u_\mathrm{c}$ by 0.02 dB; three more positions
move it by ten times that. Right is the same budget resampled a million times:
the Monte Carlo standard uncertainty is 0.4073 dB against the GUM's 0.4072, and
its 95 % interval $[73.20, 74.80]$ matches $Y \pm U$ = 74.00 ± 0.86 dB with
$k = 2.115$ at $\nu_\text{eff}$ = 16.5: this is the near-linear case where the
two methods agree, and the panel above shows the three where they do not.*

**Show the code for this figure**

```python
import matplotlib.pyplot as plt
import numpy as np
from phonometry import metrology

quantities = [
    metrology.Quantity(value=74.0, uncertainty=0.0, name="Reading"),
    metrology.rectangular(0.0, 0.20, name="Calibration"),
    metrology.rectangular(0.0, 0.30, name="Instrument"),
    metrology.Quantity(value=0.0, uncertainty=0.35, dof=9, name="Position (Type A)"),
]
model = lambda a, b, c, d: a + b + c + d
result = metrology.uncertainty_propagation(model, quantities)
mc = metrology.monte_carlo(model, quantities, n_trials=1_000_000, coverage=0.95, seed=1,
                    keep_samples=True)
expanded = result.expanded(0.95)
k, U = expanded.coverage_factor, expanded.expanded_uncertainty

# One line per panel — the budget bars, and the Monte Carlo histogram with
# its coverage interval (the committed figure also overlays the GUM Gaussian):
result.plot()
mc.plot()
plt.show()

# By hand, both panels — budget bars and the Monte Carlo output distribution:
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(12.5, 5.4))
ax1.barh(result.names, result.contributions)
ax1.axvline(result.combined_uncertainty, ls="--",
            label=rf"$u_\mathrm{{c}}$ = {result.combined_uncertainty:.3f} dB")
ax1.invert_yaxis(); ax1.legend()

rng = np.random.default_rng(1)
samples = model(np.full(200_000, 74.0),
                rng.uniform(-0.20, 0.20, 200_000),
                rng.uniform(-0.30, 0.30, 200_000),
                rng.normal(0.0, 0.35, 200_000))
ax2.hist(samples, bins=120, density=True, alpha=0.35, label="Monte Carlo")
ax2.axvspan(*mc.interval, alpha=0.12, label="95 % coverage interval")
ax2.set_title(f"Y = {result.value:.2f} dB,  U = {U:.2f} dB (k = {k:.2f})")
ax2.legend()
plt.show()
```

The `UncertaintyResult` carries the `value`, `combined_uncertainty`, the
`sensitivities`, the per-input `contributions` and the `effective_dof`; its
`.plot()` draws the budget and `.expanded(coverage)` returns   an
`ExpandedUncertainty` (`coverage_factor`, `expanded_uncertainty`). The `MonteCarloResult` carries the `value`, `standard_uncertainty`,
the coverage `interval` and its `coverage`; with `keep_samples=True` it also
retains the output `samples`, and its `.plot()` draws the output histogram
with the coverage interval marked (the right panel above). The building-acoustics uncertainty
of ISO 12999-1, which combines reproducibility terms for a single-number
rating, is a separate, domain-specific budget.

A budget is only as honest as the record behind it: every averaged input
assumes stationarity, and
[Data qualification](https://docs.jmrp.io/phonometry/signals/metrology/data-qualification/) provides the
Bendat & Piersol tests that check it before the propagation starts.

## Where the library consumes this machinery

The domain pages carry their own, standard-prescribed uncertainty clauses,
and each is an instance of the clause-5 sum this page implements in
general form; use this page whenever your measurement model is not one of
those standardized cases:

- **Calibration.** The calibrator's class tolerance and the pre/post drift
  bound of [Calibration](https://docs.jmrp.io/phonometry/signals/metrology/calibration/) are textbook
  Type B inputs. Both are stated as **half-widths**, so build them with
  `rectangular`, which applies the $a/\sqrt{3}$ divisor for you:
  `metrology.rectangular(0.0, 0.4, name="Calibrator class 1")` for the IEC 60942
  tolerance limit, and `metrology.rectangular(0.0, d / 2, name="Drift")` for a
  pre/post difference $d$. Do not pass a half-width to `Quantity`, whose second
  argument is already a *standard* uncertainty, which overstates the term by a
  factor 1.73 and nothing downstream will notice.
- **Environmental levels.** The ISO 1996-2 combined uncertainty of the
  [Environmental Levels](https://docs.jmrp.io/phonometry/environment/assessment/environmental-levels/) page is
  Formula (2) of that standard: the same quadrature sum with the standard's
  own sensitivity coefficients.
- **Building acoustics.** The per-band and single-number uncertainties of
  [field insulation measurements](https://docs.jmrp.io/phonometry/buildings/insulation/insulation-field/)
  apply the tabulated reproducibility terms of ISO 12999-1, a standardized
  budget for one specific model.
- **Absorption.** The [absorption-measurement page](https://docs.jmrp.io/phonometry/materials/absorbers/absorption-measurement/)
  carries the ISO 12999-2 absorption uncertainty, the same construction
  for the reverberation-room method.
- **Occupational exposure.** The ISO 9612 uncertainty of
  [occupational measurements](https://docs.jmrp.io/phonometry/perception/hearing/occupational-exposure/)
  budgets sampling, instrument and position contributions in exactly this
  way.

## What this guide covers

- **Covered.** From the GUM: the clause 5 law of propagation with sensitivity coefficients
  evaluated numerically, the clause 4.2 Type A evaluation as the
  `uncertainty`/`dof` pair of a `Quantity`, the clause 4.3 Type B evaluations
  (`rectangular`, `triangular`, `u_shaped`), correlated inputs through a
  correlation matrix, the Annex G.4 Welch–Satterthwaite effective degrees of
  freedom and the Annex G coverage factor, validated against the Annex H.1
  end-gauge and Annex H.2 resistance budgets. From Supplement 1: Monte Carlo
  propagation of distributions with the clause 7.7 probabilistically symmetric
  coverage interval, checked against the clause 9.2 worked examples.

- **Not covered.** The adaptive trial-count procedure of Supplement 1 clause 7.9 (the number of
  trials is fixed, with a minimum of 2), the shortest coverage interval of
  clause 5.3.4, and the multivariate-Gaussian sampling of clause 6.4.8 are not
  implemented: correlated budgets stay with `uncertainty_propagation`. Supplement 2 (multiple
  output quantities) is not implemented at all. The clause 8 validation of the
  GUM framework against the Monte Carlo result is a comparison you make by
  reading the two results, not an automated verdict. Domain budgets prescribed
  by acoustics standards (ISO 1996-2, ISO 12999-1 and -2, ISO 9612) are
  implemented on their own pages, listed above, and are not derived here.

## See also

- [Data qualification](https://docs.jmrp.io/phonometry/signals/metrology/data-qualification/): the stationarity checks every averaged input silently assumes.
- [Calibration](https://docs.jmrp.io/phonometry/signals/metrology/calibration/): the tolerance and drift inputs of every calibrated chain.
- API reference: [`metrology.uncertainty`](https://docs.jmrp.io/phonometry/reference/api/metrology/uncertainty/).
- Theory: [Measurement uncertainty (GUM)](https://docs.jmrp.io/phonometry/reference/theory/signal-analysis/#measurement-uncertainty-isoiec-guide-98-3-gum-and-supplement-1): the law of propagation of uncertainty, the coverage factor and where the Monte Carlo supplement takes over.

## References

- **International Organization for Standardization. (2020).** [*Acoustics — Determination and application of measurement uncertainties in building acoustics — Part 1: Sound insulation*](https://www.iso.org/standard/73930.html) (ISO 12999-1:2020).
  The domain-specific reproducibility budget for building-acoustics single-number ratings, a separate companion to the general GUM machinery.
- **Joint Committee for Guides in Metrology. (2008).** [*Evaluation of measurement data — Guide to the expression of uncertainty in measurement*](https://www.bipm.org/documents/20126/2071204/JCGM_100_2008_E.pdf) (JCGM 100:2008, the GUM). BIPM. https://doi.org/10.59161/JCGM100-2008E
  The law of propagation of uncertainty (clause 5), Type B evaluation (clause 4.3), the expanded uncertainty and coverage factor (clause 6, Annex G), the Welch–Satterthwaite effective degrees of freedom (Annex G.4) and the Annex H worked examples that section 1 implements and reproduces. Also published as ISO/IEC Guide 98-3:2008, Uncertainty of measurement — Part 3: Guide to the expression of uncertainty in measurement (GUM:1995). The linked PDF is the free download.
- **Joint Committee for Guides in Metrology. (2008).** [*Evaluation of measurement data — Supplement 1 to the "Guide to the expression of uncertainty in measurement" — Propagation of distributions using a Monte Carlo method*](https://www.bipm.org/documents/20126/2071204/JCGM_101_2008_E.pdf) (JCGM 101:2008). BIPM. https://doi.org/10.59161/JCGM101-2008
  The Monte Carlo propagation of section 2, its probabilistically symmetric coverage interval (clause 7) and the clause 8 validation of the GUM framework against the Monte Carlo result. Also published as ISO/IEC Guide 98-3-1:2008. The linked PDF is the free download.
