plurimath 0.11.16 → 0.11.18
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- checksums.yaml +4 -4
- data/.github/workflows/dependent-tests.yml +2 -2
- data/.github/workflows/docs.yml +26 -0
- data/.github/workflows/gum-reference.yml +48 -0
- data/.github/workflows/performance-ci-legs.yml +54 -0
- data/.github/workflows/performance.yml +99 -0
- data/.github/workflows/roundtrip-cases.yml +33 -0
- data/.rubocop_todo.yml +0 -138
- data/Gemfile +13 -0
- data/Rakefile +2 -0
- data/docs/asciimath.adoc +578 -0
- data/docs/core-model-contract.adoc +748 -0
- data/docs/gum-evaluation.adoc +368 -0
- data/docs/latex.adoc +335 -0
- data/docs/mathml.adoc +638 -0
- data/docs/native-evaluation-request.md +28 -0
- data/docs/numeric-request.md +19 -0
- data/docs/omml.adoc +589 -0
- data/docs/performance-contract.adoc +279 -0
- data/docs/pma.adoc +949 -0
- data/docs/quantity-request.md +19 -0
- data/docs/uncertainty-asciimath-tree.md +2 -3
- data/docs/uncertainty.adoc +401 -425
- data/docs/unicodemath.adoc +453 -0
- data/docs/unitsml.adoc +414 -0
- data/lib/plurimath/asciimath/constants.rb +0 -98
- data/lib/plurimath/asciimath/mapper/rules.rb +1400 -0
- data/lib/plurimath/asciimath/mapper.rb +148 -0
- data/lib/plurimath/asciimath/parse.rb +26 -18
- data/lib/plurimath/asciimath/parser.rb +25 -11
- data/lib/plurimath/asciimath/slice_adapter.rb +1 -1
- data/lib/plurimath/asciimath.rb +1 -1
- data/lib/plurimath/catalog.rb +4 -5
- data/lib/plurimath/cli/convert.rb +3 -2
- data/lib/plurimath/cli.rb +2 -1
- data/lib/plurimath/configuration.rb +3 -0
- data/lib/plurimath/evaluation_request/quantity.rb +254 -0
- data/lib/plurimath/evaluation_request.rb +160 -0
- data/lib/plurimath/latex/constants.rb +0 -84
- data/lib/plurimath/latex/mapper.rb +418 -0
- data/lib/plurimath/latex/parse.rb +28 -288
- data/lib/plurimath/latex/parser.rb +49 -4
- data/lib/plurimath/latex.rb +1 -1
- data/lib/plurimath/math/core.rb +266 -1
- data/lib/plurimath/math/evaluation/accuracy.rb +62 -0
- data/lib/plurimath/math/evaluation/calibration.rb +85 -0
- data/lib/plurimath/math/evaluation/compiled_polynomial.rb +99 -0
- data/lib/plurimath/math/evaluation/compiled_rational.rb +131 -0
- data/lib/plurimath/math/evaluation/conformity.rb +133 -0
- data/lib/plurimath/math/evaluation/correlated_normal_sampler.rb +74 -0
- data/lib/plurimath/math/evaluation/covariance.rb +173 -0
- data/lib/plurimath/math/evaluation/distribution_sampler.rb +126 -0
- data/lib/plurimath/math/evaluation/evaluator.rb +236 -22
- data/lib/plurimath/math/evaluation/grouped_statistics.rb +56 -0
- data/lib/plurimath/math/evaluation/gum.rb +410 -0
- data/lib/plurimath/math/evaluation/implicit_model.rb +66 -0
- data/lib/plurimath/math/evaluation/monte_carlo.rb +198 -0
- data/lib/plurimath/math/evaluation/quantity_evaluator.rb +127 -0
- data/lib/plurimath/math/evaluation/result.rb +45 -0
- data/lib/plurimath/math/evaluation/statistics.rb +102 -0
- data/lib/plurimath/math/evaluation/taylor.rb +162 -0
- data/lib/plurimath/math/evaluation/wichmann_hill.rb +40 -0
- data/lib/plurimath/math/evaluation.rb +22 -0
- data/lib/plurimath/math/formula/mrow.rb +6 -1
- data/lib/plurimath/math/formula.rb +22 -4
- data/lib/plurimath/math/function/abs.rb +3 -1
- data/lib/plurimath/math/function/arccos.rb +1 -1
- data/lib/plurimath/math/function/arcsin.rb +1 -1
- data/lib/plurimath/math/function/arctan.rb +1 -1
- data/lib/plurimath/math/function/bar.rb +2 -1
- data/lib/plurimath/math/function/base.rb +2 -1
- data/lib/plurimath/math/function/binary_function.rb +1 -1
- data/lib/plurimath/math/function/ceil.rb +3 -1
- data/lib/plurimath/math/function/color.rb +2 -1
- data/lib/plurimath/math/function/cos.rb +1 -1
- data/lib/plurimath/math/function/cosh.rb +1 -1
- data/lib/plurimath/math/function/cot.rb +1 -1
- data/lib/plurimath/math/function/coth.rb +1 -1
- data/lib/plurimath/math/function/csc.rb +1 -1
- data/lib/plurimath/math/function/csch.rb +1 -1
- data/lib/plurimath/math/function/ddot.rb +2 -1
- data/lib/plurimath/math/function/dot.rb +2 -1
- data/lib/plurimath/math/function/exp.rb +1 -1
- data/lib/plurimath/math/function/fenced.rb +5 -3
- data/lib/plurimath/math/function/floor.rb +3 -1
- data/lib/plurimath/math/function/font_style/bold.rb +1 -1
- data/lib/plurimath/math/function/font_style/double_struck.rb +1 -1
- data/lib/plurimath/math/function/font_style/fraktur.rb +1 -1
- data/lib/plurimath/math/function/font_style/italic.rb +1 -1
- data/lib/plurimath/math/function/font_style/monospace.rb +1 -1
- data/lib/plurimath/math/function/font_style/normal.rb +22 -1
- data/lib/plurimath/math/function/font_style/sans-serif.rb +1 -1
- data/lib/plurimath/math/function/font_style/script.rb +1 -1
- data/lib/plurimath/math/function/frac.rb +2 -1
- data/lib/plurimath/math/function/hat.rb +2 -1
- data/lib/plurimath/math/function/int.rb +3 -2
- data/lib/plurimath/math/function/lg.rb +1 -1
- data/lib/plurimath/math/function/linebreak.rb +2 -1
- data/lib/plurimath/math/function/ln.rb +1 -1
- data/lib/plurimath/math/function/log.rb +4 -1
- data/lib/plurimath/math/function/longdiv.rb +3 -0
- data/lib/plurimath/math/function/mpadded.rb +2 -1
- data/lib/plurimath/math/function/msgroup.rb +3 -0
- data/lib/plurimath/math/function/multiscript.rb +4 -0
- data/lib/plurimath/math/function/nary.rb +9 -3
- data/lib/plurimath/math/function/norm.rb +3 -1
- data/lib/plurimath/math/function/obrace.rb +2 -1
- data/lib/plurimath/math/function/oint.rb +3 -2
- data/lib/plurimath/math/function/overleftrightarrow.rb +2 -1
- data/lib/plurimath/math/function/overset.rb +2 -1
- data/lib/plurimath/math/function/power.rb +1 -1
- data/lib/plurimath/math/function/power_base.rb +1 -1
- data/lib/plurimath/math/function/prod.rb +3 -2
- data/lib/plurimath/math/function/sec.rb +1 -1
- data/lib/plurimath/math/function/sech.rb +1 -1
- data/lib/plurimath/math/function/semantics.rb +35 -0
- data/lib/plurimath/math/function/sin.rb +1 -1
- data/lib/plurimath/math/function/sinh.rb +1 -1
- data/lib/plurimath/math/function/sqrt.rb +3 -2
- data/lib/plurimath/math/function/sum.rb +3 -2
- data/lib/plurimath/math/function/table/align.rb +1 -1
- data/lib/plurimath/math/function/table/bmatrix.rb +1 -1
- data/lib/plurimath/math/function/table/cases.rb +4 -0
- data/lib/plurimath/math/function/table/multline.rb +1 -1
- data/lib/plurimath/math/function/table/pmatrix.rb +1 -1
- data/lib/plurimath/math/function/table/split.rb +1 -1
- data/lib/plurimath/math/function/table/vmatrix.rb +1 -1
- data/lib/plurimath/math/function/table.rb +19 -2
- data/lib/plurimath/math/function/tan.rb +1 -1
- data/lib/plurimath/math/function/tanh.rb +1 -1
- data/lib/plurimath/math/function/td.rb +4 -0
- data/lib/plurimath/math/function/ternary_function.rb +2 -2
- data/lib/plurimath/math/function/text.rb +3 -0
- data/lib/plurimath/math/function/tilde.rb +2 -1
- data/lib/plurimath/math/function/tr.rb +3 -0
- data/lib/plurimath/math/function/ubrace.rb +2 -1
- data/lib/plurimath/math/function/ul.rb +2 -1
- data/lib/plurimath/math/function/unary_function.rb +1 -1
- data/lib/plurimath/math/function/uncertainty.rb +26 -11
- data/lib/plurimath/math/function/underset.rb +2 -1
- data/lib/plurimath/math/function/unitsml.rb +18 -6
- data/lib/plurimath/math/function/vec.rb +2 -1
- data/lib/plurimath/math/number.rb +30 -12
- data/lib/plurimath/math/symbol_index.rb +10254 -0
- data/lib/plurimath/math/symbol_table.rb +89 -0
- data/lib/plurimath/math/symbols/paren.rb +7 -0
- data/lib/plurimath/math/symbols/symbol.rb +30 -3
- data/lib/plurimath/math/symbols.rb +81 -5
- data/lib/plurimath/math.rb +8 -0
- data/lib/plurimath/mathml/parser.rb +2 -0
- data/lib/plurimath/mathml/translator.rb +17 -3
- data/lib/plurimath/native_kernel.rb +64 -0
- data/lib/plurimath/omml/parser.rb +1 -0
- data/lib/plurimath/omml/translator.rb +2 -1
- data/lib/plurimath/pma.rb +19 -0
- data/lib/plurimath/unicode_math/constants.rb +0 -258
- data/lib/plurimath/unicode_math/mapper/capture_order.rb +142 -0
- data/lib/plurimath/unicode_math/mapper/rules.rb +4086 -0
- data/lib/plurimath/unicode_math/mapper.rb +146 -0
- data/lib/plurimath/unicode_math/parse.rb +54 -294
- data/lib/plurimath/unicode_math/parser.rb +71 -2
- data/lib/plurimath/unicode_math.rb +1 -2
- data/lib/plurimath/unitsml.rb +0 -1
- data/lib/plurimath/utility.rb +34 -12
- data/lib/plurimath/version.rb +1 -1
- data/lib/plurimath.rb +42 -3
- data/plurimath.gemspec +3 -1
- data/scripts/check_gum_references.rb +69 -0
- data/scripts/check_suite_budget.rb +329 -0
- data/scripts/docs/asciimath_tables.rb +363 -0
- data/scripts/docs/core_inventory.rb +432 -0
- data/scripts/docs/format_specs.rb +672 -0
- data/scripts/generate_symbol_index.rb +118 -0
- data/scripts/measure_require.rb +362 -0
- data/scripts/replay_asciimath_corpus.rb +7 -7
- data/scripts/replay_stage_roundtrips.rb +242 -0
- data/scripts/require_baseline.json +34 -0
- data/scripts/roundtrip_cases.rb +462 -0
- metadata +97 -16
- data/lib/plurimath/asciimath/transform.rb +0 -1395
- data/lib/plurimath/latex/transform.rb +0 -1006
- data/lib/plurimath/unicode_math/parsing_rules/absence_rules.rb +0 -150
- data/lib/plurimath/unicode_math/parsing_rules/common_rules.rb +0 -136
- data/lib/plurimath/unicode_math/parsing_rules/constants_rules.rb +0 -136
- data/lib/plurimath/unicode_math/parsing_rules/helper.rb +0 -13
- data/lib/plurimath/unicode_math/parsing_rules/masked.rb +0 -84
- data/lib/plurimath/unicode_math/parsing_rules/sub_sup.rb +0 -281
- data/lib/plurimath/unicode_math/parsing_rules.rb +0 -14
- data/lib/plurimath/unicode_math/transform.rb +0 -4058
|
@@ -0,0 +1,368 @@
|
|
|
1
|
+
= Uncertainty propagation and measurement statistics
|
|
2
|
+
|
|
3
|
+
`Plurimath::Math::Evaluation::Gum` accepts explicit estimates, component
|
|
4
|
+
variances and analytic sensitivity coefficients in one coherent unit system.
|
|
5
|
+
`analyze` derives exact first, second and third derivatives of an existing
|
|
6
|
+
Formula with polynomial/rational arithmetic and integer powers. It reuses
|
|
7
|
+
the existing expression parser. Unsupported function derivatives are refused;
|
|
8
|
+
there is no finite-difference Hessian fallback. Neither entrypoint infers
|
|
9
|
+
units or independence.
|
|
10
|
+
|
|
11
|
+
[source,ruby]
|
|
12
|
+
----
|
|
13
|
+
gum = Plurimath::Math::Evaluation::Gum
|
|
14
|
+
component = gum.component(distribution: :rectangular, half_width: "0.3")
|
|
15
|
+
result = gum.propagate(estimate: "12.3", components: [component],
|
|
16
|
+
gradient: [1], independent: true, guard_digits: 30)
|
|
17
|
+
result[:variance] # => (3/100)
|
|
18
|
+
result[:standard_uncertainty] # BigDecimal, declared precision
|
|
19
|
+
----
|
|
20
|
+
|
|
21
|
+
Inputs accept integers, rationals, decimal strings and finite BigDecimal
|
|
22
|
+
values. Binary Float input is refused. `type_a(observations)` reports count,
|
|
23
|
+
mean, extrema, sample variance, variance of the mean and degrees of freedom.
|
|
24
|
+
A rectangular, triangular or arcsine component takes a half-width. Curvilinear trapezoid additionally takes
|
|
25
|
+
`boundary_half_width: d`, with `0 < d < half_width`, giving variance
|
|
26
|
+
`half_width^2/3 + d^2/9` (JCGM 101 6.4.3.4); a normal
|
|
27
|
+
component takes a standard uncertainty. Student-t takes its scale and degrees
|
|
28
|
+
of freedom, giving variance `scale^2 * nu/(nu-2)`; `nu <= 2` is refused.
|
|
29
|
+
|
|
30
|
+
First-order propagation gives the sum of squared sensitivities times input
|
|
31
|
+
variances. Optional positive component degrees of freedom give the
|
|
32
|
+
Welch-Satterthwaite effective degrees of freedom. An omitted value denotes
|
|
33
|
+
infinite degrees of freedom; a returned nil means the denominator is zero.
|
|
34
|
+
This does not choose a coverage factor or assess LPU adequacy.
|
|
35
|
+
|
|
36
|
+
For `order: 2`, all components must explicitly declare `distribution: :normal`.
|
|
37
|
+
Supply a complete symmetric `hessian` matrix and `third` rank-three tensor,
|
|
38
|
+
plus `derivatives_complete: true`. The estimate includes half the diagonal
|
|
39
|
+
Hessian contraction. The variance includes both the Hessian-square and
|
|
40
|
+
first-times-third derivative terms in JCGM 100:2008 5.1.2 note. A Hessian
|
|
41
|
+
alone is insufficient: `f(x)=x+x^3` at zero has a zero Hessian and a nonzero
|
|
42
|
+
order-four variance correction. Every derivative entry, including zeros,
|
|
43
|
+
must be supplied. Unsupported joint dependence, nonnormal GUF2, incomplete
|
|
44
|
+
derivatives and negative approximated variance raise a named error.
|
|
45
|
+
|
|
46
|
+
[source,ruby]
|
|
47
|
+
----
|
|
48
|
+
formula = Plurimath::Math.parse("x+x^3", :asciimath)
|
|
49
|
+
result = gum.analyze(formula, estimates: { x: 0 },
|
|
50
|
+
components: { x: { variance: "0.01", distribution: :normal } },
|
|
51
|
+
independent: true, order: 2, guard_digits: 40)
|
|
52
|
+
result[:variance] # => (53/5000), including the first-times-third term
|
|
53
|
+
----
|
|
54
|
+
|
|
55
|
+
The estimate and component hashes must have identical ordered variable keys.
|
|
56
|
+
Results identify the engine version and method revision. They retain exact
|
|
57
|
+
rational variances and BigDecimal standard uncertainties; this is a Ruby API,
|
|
58
|
+
not the complete serialized cross-runtime evaluator contract.
|
|
59
|
+
|
|
60
|
+
== H.1 conventions and verification
|
|
61
|
+
|
|
62
|
+
The fixture records four distinct source conventions: rounded GUM 100 H.1
|
|
63
|
+
standard uncertainties, unrounded H.1 source components, JCGM 101 Table 10
|
|
64
|
+
PDF population moments, and GUM 5:2026 Table 8.1 PDF population moments.
|
|
65
|
+
These are different inputs and produce different uncertainties. The tests
|
|
66
|
+
use explicitly synthetic independent normal inputs with each variance set
|
|
67
|
+
to check the Taylor contraction; they refuse the actual nonnormal H.1 PDFs.
|
|
68
|
+
No assertion forces the original 34 nm result to equal the amended 36 nm
|
|
69
|
+
result within 0.5 nm. A special multi-affine moment identity does not remove
|
|
70
|
+
the general normality restriction.
|
|
71
|
+
|
|
72
|
+
Run `python3 spec/support/gum_h1_reference.py` to regenerate the exact
|
|
73
|
+
rational derivative fixtures and independently factored variance references.
|
|
74
|
+
The focused spec also checks a polynomial with nonzero third derivatives,
|
|
75
|
+
a quadratic with a nonzero mean correction and the refusal boundaries.
|
|
76
|
+
|
|
77
|
+
Sources: https://www.bipm.org/documents/20126/2071204/JCGM_100_2008_E.pdf[JCGM 100:2008],
|
|
78
|
+
https://www.bipm.org/documents/20126/2071204/JCGM_101_2008_E.pdf[JCGM 101:2008],
|
|
79
|
+
and the BIPM JCGM publications for GUM 100:2008/Amd.1:2026 and GUM 5:2026.
|
|
80
|
+
|
|
81
|
+
This entrypoint supplies numerical propagation, not a complete JCGM engine.
|
|
82
|
+
Unsupported function derivatives, automatic unit-aware GUM propagation,
|
|
83
|
+
correlated inputs, global risks and Ruby/TS wire conformance remain outside
|
|
84
|
+
this capability. An LPU/GUF2 result retains `adequacy: :not_assessed`; producing
|
|
85
|
+
a coverage factor alone does not validate the approximation.
|
|
86
|
+
|
|
87
|
+
== Quantity arithmetic
|
|
88
|
+
|
|
89
|
+
`Formula#evaluate_quantity(bindings)` is an explicit arithmetic entrypoint.
|
|
90
|
+
It needs UnitsML Quantity/conversion support (available in 0.6.14). Older
|
|
91
|
+
UnitsML versions retain the ordinary numeric evaluator; a quantity request
|
|
92
|
+
receives a named unsupported-expression error when that engine is absent.
|
|
93
|
+
Construct bindings with `Unitsml::Quantity.new(value: ..., unit: ...)`.
|
|
94
|
+
Numeric bindings and literals are dimensionless; integer division retains
|
|
95
|
+
rational magnitudes. Unit parsing, dimensions, composition, conversion and
|
|
96
|
+
scale legality stay in UnitsML. Addition and subtraction convert the right
|
|
97
|
+
operand into the left operand's unit before delegating the operation.
|
|
98
|
+
|
|
99
|
+
[source,ruby]
|
|
100
|
+
----
|
|
101
|
+
require "unitsml"
|
|
102
|
+
formula = Plurimath::Math.parse("x+y", :asciimath)
|
|
103
|
+
outcome = formula.evaluate_quantity(
|
|
104
|
+
x: Unitsml::Quantity.new(value: 1, unit: "m"),
|
|
105
|
+
y: Unitsml::Quantity.new(value: 50, unit: "cm")
|
|
106
|
+
)
|
|
107
|
+
outcome.value.value # => (3/2)
|
|
108
|
+
outcome.value.norm_text # => "m"
|
|
109
|
+
----
|
|
110
|
+
|
|
111
|
+
The supported expression subset is arithmetic, rational fractions, parentheses
|
|
112
|
+
and integer powers. Other functions raise an evaluator unsupported-expression
|
|
113
|
+
error. Dimension mismatches, conversion failures and invalid scale operations
|
|
114
|
+
preserve UnitsML's named error types. This does not attach units to GUM
|
|
115
|
+
sensitivities automatically; callers of Gum still supply one coherent system.
|
|
116
|
+
|
|
117
|
+
Quantity results expose `accuracy`; limited Float or BigDecimal sources remain
|
|
118
|
+
limited even when UnitsML conversion returns a Rational magnitude.
|
|
119
|
+
|
|
120
|
+
== Series and coverage factors
|
|
121
|
+
|
|
122
|
+
`Gum.series(observations, axis:, window:, guard_digits:)` reports exact mean,
|
|
123
|
+
extrema, count and sample variance, plus decimal strings for sample standard
|
|
124
|
+
deviation and standard uncertainty of the mean. For a matrix, axis 0 treats
|
|
125
|
+
rows as observations and axis 1 treats columns as observations. A window is
|
|
126
|
+
nil (all observations) or zero-based `[start, length]` on the observation
|
|
127
|
+
axis; out-of-range, ragged or insufficient samples are refused.
|
|
128
|
+
|
|
129
|
+
`Gum.coverage(standard_uncertainty:, probability:, degrees_of_freedom:,
|
|
130
|
+
guard_digits:)` returns a coverage factor and expanded uncertainty as decimal
|
|
131
|
+
strings. Nil degrees of freedom selects a normal distribution; finite values
|
|
132
|
+
select Student-t. The verified domain is central probability 0.5 through 0.999,
|
|
133
|
+
Student-t degrees of freedom 1 through 10000, and at most 8 output significant
|
|
134
|
+
digits. Other requests receive a typed refusal. Student-t factors come from
|
|
135
|
+
integrating the normalized density and bracketing its CDF, independently
|
|
136
|
+
checked against 80-digit incomplete-beta references. They are not interpolated
|
|
137
|
+
from a table.
|
|
138
|
+
|
|
139
|
+
Explicit guard-digit calculations select half-up decimal rounding and restore
|
|
140
|
+
the caller's decimal context. Statistical wire rounding uses integer/rational
|
|
141
|
+
arithmetic and does not depend on ambient BigDecimal limits or modes.
|
|
142
|
+
|
|
143
|
+
== Deterministic adaptive Monte Carlo
|
|
144
|
+
|
|
145
|
+
[source,ruby]
|
|
146
|
+
----
|
|
147
|
+
result = gum.monte_carlo(formula,
|
|
148
|
+
inputs: { x: { distribution: :normal, location: "0", sd: "1" } },
|
|
149
|
+
independent: true, seeds: [12345,23456,34567,45678], probability: "0.95",
|
|
150
|
+
interval: :both, significant_digits: 2, guard_digits: 12,
|
|
151
|
+
max_samples: 200000
|
|
152
|
+
)
|
|
153
|
+
----
|
|
154
|
+
|
|
155
|
+
The sampler uses the enhanced four-state Wichmann-Hill recurrence from
|
|
156
|
+
JCGM 101:2008 C.3.3.3. Four nonzero seeds below their respective moduli are
|
|
157
|
+
required; they, the final state, draw count, method revision and variable
|
|
158
|
+
order are included in the result. No process-global random source is used.
|
|
159
|
+
Normal and Student-t sampling follow C.4 and C.6. Rectangular, symmetric
|
|
160
|
+
triangular, arcsine and fixed-midpoint curvilinear trapezoid PDFs are also
|
|
161
|
+
supported. Every input must declare its PDF and parameters; Student-t `scale`
|
|
162
|
+
is not its standard deviation.
|
|
163
|
+
|
|
164
|
+
The existing Formula parser compiles exact polynomials about the exact PDF
|
|
165
|
+
centers. Large offset cancellation is simplified before sampling. The current
|
|
166
|
+
sampling domain has at most 128 intermediate terms and total degree 32;
|
|
167
|
+
nonconstant divisors are refused because finite output moments are not
|
|
168
|
+
established. Finite output variance is checked before sampling: a Student-t input needs
|
|
169
|
+
degrees of freedom greater than twice its polynomial degree. A linear
|
|
170
|
+
Student-t input with three degrees of freedom is supported; its square is
|
|
171
|
+
refused because that output has infinite variance. The batch standard
|
|
172
|
+
deviation has a finite second moment whenever output variance is finite;
|
|
173
|
+
no unnecessary fourth-moment condition is imposed.
|
|
174
|
+
|
|
175
|
+
The batch size, significant-digit tolerance and stopping statistics follow
|
|
176
|
+
7.9.4 and equation 18. Both probabilistically symmetric and shortest intervals
|
|
177
|
+
use the one-based endpoint convention of 7.7.2. With `interval: :both`, every
|
|
178
|
+
endpoint is included in the stabilization test. Exhausting the explicit
|
|
179
|
+
sample cap raises `ConvergenceError`; it never reports an unstabilized
|
|
180
|
+
calculation as complete. This statistical stopping procedure is not a
|
|
181
|
+
mathematical guarantee of a particular realization's error.
|
|
182
|
+
|
|
183
|
+
Sampling and centered evaluation use declared binary64 arithmetic; the exact
|
|
184
|
+
polynomial coefficients and center expansion precede that computation.
|
|
185
|
+
Non-exact result fields are decimal strings with at most 12 significant
|
|
186
|
+
digits. `significant_digits` controls statistical stabilization independently
|
|
187
|
+
of the wire's `guard_digits`. Cross-runtime byte identity is not asserted.
|
|
188
|
+
|
|
189
|
+
The separate `GUM mathematical references` Actions job runs both actual
|
|
190
|
+
nonnormal H.1 PDF conventions at two meaningful uncertainty digits and both
|
|
191
|
+
interval kinds. Its artifact includes the actual sample counts, tolerance,
|
|
192
|
+
seed states and comparison with independently factored population moments.
|
|
193
|
+
The bounded local one-digit run is a coarser check, not a substitute for that
|
|
194
|
+
job. Reproduce the stronger check with:
|
|
195
|
+
|
|
196
|
+
[source,shell]
|
|
197
|
+
----
|
|
198
|
+
bundle exec ruby scripts/check_gum_references.rb --output /tmp/gum-reference.json \
|
|
199
|
+
--significant-digits 2 --max-samples 2000000 \
|
|
200
|
+
spec/fixtures/evaluation/gum_h1_mcm.json
|
|
201
|
+
----
|
|
202
|
+
|
|
203
|
+
== Normal conformity decisions
|
|
204
|
+
|
|
205
|
+
`Gum.conformity` implements the normal, measurement-only knowledge case of
|
|
206
|
+
JCGM 106:2012. Both `distribution: :normal` and
|
|
207
|
+
`knowledge: :measurement_only` are explicit. This identifies the posterior
|
|
208
|
+
estimate/uncertainty with the measurement estimate/uncertainty for the
|
|
209
|
+
capability index and guard geometry; informative priors are refused.
|
|
210
|
+
|
|
211
|
+
Supply `estimate:`, `standard_uncertainty:`, two `tolerance:` limits (nil for
|
|
212
|
+
an unbounded side), `rule:`, `coverage_factor:`, `guard_ratio:` and
|
|
213
|
+
`guard_digits:`. Rules are simple acceptance, guarded acceptance, guarded
|
|
214
|
+
rejection, and `iso14253_style` (the k=2,r=1 geometry described in JCGM 106).
|
|
215
|
+
The separately named `three_zone` policy reports conditional outcomes when
|
|
216
|
+
the expanded interval overlaps a tolerance boundary. It is not substituted
|
|
217
|
+
for a requested binary decision rule. Limits are closed, and boundary status
|
|
218
|
+
and signed residuals are explicit.
|
|
219
|
+
|
|
220
|
+
The returned specific consumer risk applies only to acceptance; specific
|
|
221
|
+
producer risk applies only to rejection. Conditional decisions report neither
|
|
222
|
+
as an actual decision risk. Global risks need a process/measurement joint PDF
|
|
223
|
+
and receive a typed refusal. Conformance probabilities use both normal tails;
|
|
224
|
+
probability fields declare 1 through 12 decimal places and their resolution.
|
|
225
|
+
|
|
226
|
+
A 2.3 percent one-sided tail ceiling must not be assumed for every two-sided
|
|
227
|
+
geometry. For example, with tolerance width 4u and w=2u, Cm=1 and the accepted
|
|
228
|
+
midpoint has nonconformance probability 2*Phi(-2), about 4.55 percent.
|
|
229
|
+
The engine returns the exact two-tail normal probability for the declared
|
|
230
|
+
geometry. A requested `consumer_risk_ceiling:` is refused if the guard band
|
|
231
|
+
does not establish it at the declared numerical resolution. The
|
|
232
|
+
`iso14253_style` name denotes the geometry documented in JCGM 106:2012 and does
|
|
233
|
+
not claim compliance with an unprovided later ISO edition.
|
|
234
|
+
|
|
235
|
+
Reference regeneration: `spec/support/gum_statistics_reference.py` uses
|
|
236
|
+
mpmath 1.3.0 at 80 decimal digits; `spec/support/wichmann_hill_reference.py` uses
|
|
237
|
+
the independent 32-bit recurrence and exact rational arithmetic. Runtime Ruby
|
|
238
|
+
uses different quadrature/recurrence paths. Normative publications remain at
|
|
239
|
+
their official source; no PDF is bundled with this implementation.
|
|
240
|
+
|
|
241
|
+
`Gum.validate_coverage` explicitly compares a proposed GUM interval against a
|
|
242
|
+
stabilized MCM interval at the same probability. It derives the target tolerance
|
|
243
|
+
from the proposed GUM standard uncertainty and meaningful digits, and requires
|
|
244
|
+
MCM to be at least five times tighter (JCGM 101 clauses 8.1/8.2). It returns
|
|
245
|
+
`validated` or `not_validated` for that specified probability and interval;
|
|
246
|
+
coarser, mismatched or incomplete MCM receives a typed refusal. This does not
|
|
247
|
+
turn `adequacy: :not_assessed` into an unconditional model-validity claim.
|
|
248
|
+
|
|
249
|
+
|
|
250
|
+
== Covariance, calibration and grouped observations
|
|
251
|
+
|
|
252
|
+
`Gum.type_a_covariance(paired_observations)` retains the sample covariance
|
|
253
|
+
and covariance of the mean of paired observations. `Gum.propagate_vector`
|
|
254
|
+
uses an explicit Jacobian and covariance matrix and returns the complete
|
|
255
|
+
output covariance, including cross-output terms. Covariance matrices must be
|
|
256
|
+
symmetric positive semidefinite; singular matrices are accepted without
|
|
257
|
+
adding diagonal noise.
|
|
258
|
+
|
|
259
|
+
`Gum.propagate_covariance` extends the scalar propagation interface with
|
|
260
|
+
`covariance:` and `joint_distribution:`. First-order propagation uses the
|
|
261
|
+
complete covariance matrix. Second-order propagation requires explicit
|
|
262
|
+
`:multivariate_normal` joint knowledge, normal component declarations, and
|
|
263
|
+
complete symmetric Hessian and third-derivative tensors. Normal marginal
|
|
264
|
+
PDFs and zero correlation alone do not establish joint normality. Correlated
|
|
265
|
+
Welch-Satterthwaite degrees of freedom are not inferred.
|
|
266
|
+
|
|
267
|
+
Student-t PDF shape and variance-estimate degrees of freedom are separate.
|
|
268
|
+
Use `pdf_degrees_of_freedom:` for the PDF and, when supported by the source,
|
|
269
|
+
`variance_degrees_of_freedom:` for the uncertainty estimate. The legacy
|
|
270
|
+
Student-t `degrees_of_freedom:` keyword remains a PDF-shape alias. Its
|
|
271
|
+
variance-estimate degrees default to `:unknown`, which propagates an unknown
|
|
272
|
+
effective-degrees status. For other distributions, nil explicitly denotes
|
|
273
|
+
infinite variance-estimate degrees; `:unknown` preserves missing information.
|
|
274
|
+
|
|
275
|
+
`Gum.linear_calibration` implements the unweighted GUM H.3 line fit. The
|
|
276
|
+
caller declares `abscissa_uncertainty: :negligible` and
|
|
277
|
+
`residual_model: :independent_common_variance`. Prediction distinguishes
|
|
278
|
+
`:mean_correction` from `:new_observation`; the latter includes residual
|
|
279
|
+
variance. Extrapolation is flagged. The engine does not infer linearity,
|
|
280
|
+
normality, or negligible temperature uncertainty.
|
|
281
|
+
|
|
282
|
+
`Gum.grouped_type_a` accepts balanced group means and sample variances
|
|
283
|
+
with `observations_per_group:`. It reports both mean squares, the F statistic
|
|
284
|
+
and its degrees of freedom. The caller chooses `:random_group_effect` or
|
|
285
|
+
`:no_group_effect`; no automatic F-test changes that choice. A negative
|
|
286
|
+
estimated random-group component requires an explicit constrained model and
|
|
287
|
+
is refused rather than silently clipped.
|
|
288
|
+
|
|
289
|
+
The independently regenerated Annex H fixture covers paired electrical
|
|
290
|
+
outputs (H.2), calibration (H.3), decay-corrected activity (H.4), balanced
|
|
291
|
+
observation summaries (H.5), and Rockwell hardness components (H.6). Exact input
|
|
292
|
+
profiles and numerical tolerances are in the focused specs. H.5 distinguishes
|
|
293
|
+
provided summaries from the publication's rounded intermediate values. H.2
|
|
294
|
+
checks both supplied raw observations and rounded uncertainty/correlation
|
|
295
|
+
summaries against independent calculations; neither reproduces every
|
|
296
|
+
published uncertainty to its last digit. Unrounded source observations or
|
|
297
|
+
the publication's calculation/rounding convention are needed to close that
|
|
298
|
+
separate boundary. Missing certificate or pooled-variance degrees of freedom
|
|
299
|
+
remain unknown. These examples do not establish general model adequacy.
|
|
300
|
+
|
|
301
|
+
|
|
302
|
+
== Implicit measurements and additional source profiles
|
|
303
|
+
|
|
304
|
+
`Gum.propagate_implicit` linearizes a caller-supplied root of `h(y,x)=0`.
|
|
305
|
+
Supply output estimates, `output_jacobian:`, `input_jacobian:`, input
|
|
306
|
+
`covariance:`, `residuals:`, `residual_tolerance:` and `guard_digits:`.
|
|
307
|
+
Exact pivoted elimination solves the sensitivity system, and the result
|
|
308
|
+
retains the complete output covariance. Singular output Jacobians and
|
|
309
|
+
residuals beyond the declared tolerance receive typed refusals. This is not
|
|
310
|
+
a root finder; the residual tolerance has the units of the supplied equations.
|
|
311
|
+
It does not establish the adequacy of the linearization.
|
|
312
|
+
|
|
313
|
+
The GUM5 profiles include all ten thermometer outputs in section13, with
|
|
314
|
+
shared calibration uncertainty; all six correlated Gaussian quadratic
|
|
315
|
+
comparison-loss moment cases in section16; and the linear-sum variances of
|
|
316
|
+
section6. The actual mass-calibration PDFs of section7 are used for LPU;
|
|
317
|
+
normal-only GUF2 is refused for that mixed-PDF profile. An independent exact
|
|
318
|
+
population-moment calculation uses reciprocal-uniform integrals. Its
|
|
319
|
+
reference-density half-width is 0.05 kg/m3, as specified in Table7.1.
|
|
320
|
+
MCM supports the mass model through exact rational compilation and an
|
|
321
|
+
explicit denominator-support certificate. Correlated normal MCM requires
|
|
322
|
+
`independent: false`, `joint_distribution: :multivariate_normal` and
|
|
323
|
+
`joint_covariance:` in the ordered input mapping. Marginal normal variances
|
|
324
|
+
must agree with the covariance diagonal. Exact LDL factorization supports
|
|
325
|
+
singular positive-semidefinite covariance without jitter; final factor roots
|
|
326
|
+
and samples use binary64. The result records the joint law, covariance and
|
|
327
|
+
factor policy.
|
|
328
|
+
Normal-equivalent local Taylor derivatives must not be interpreted as finite
|
|
329
|
+
population moments of reciprocals of unbounded Gaussian variables.
|
|
330
|
+
|
|
331
|
+
The GUM6:2020 section13.1.4 bromine profile uses the source's explicit normal
|
|
332
|
+
PDF example. It demonstrates vanishing first-order sensitivity, nonzero
|
|
333
|
+
quadratic variance, and the corrected mean. Neither unrequested truncation
|
|
334
|
+
nor a different physical PDF is inferred from the example. Source digests,
|
|
335
|
+
inputs, independent numerical reference generators and declared tolerances
|
|
336
|
+
are included in the fixtures and focused specs.
|
|
337
|
+
|
|
338
|
+
|
|
339
|
+
The rational sampler retains exact numerator/denominator polynomials and
|
|
340
|
+
centers the output before binary64 evaluation. Conservative exact interval
|
|
341
|
+
bounds must prove the denominator is separated from zero using explicitly
|
|
342
|
+
bounded rectangular input supports. Other denominator PDFs, zero-crossing
|
|
343
|
+
support and unsupported coefficient ranges receive typed refusals. A finite
|
|
344
|
+
normal-input reciprocal is not inferred from a small standard deviation.
|
|
345
|
+
Finite-output-variance checks use numerator polynomial degrees together with
|
|
346
|
+
the bounded reciprocal certificate. Result metadata identifies this backend
|
|
347
|
+
and records the exact denominator interval; it is distinct from the original
|
|
348
|
+
polynomial backend.
|
|
349
|
+
|
|
350
|
+
`spec/fixtures/evaluation/gum_mcm_profiles.json` supplies eleven explicit
|
|
351
|
+
source profiles: three linear sums, six correlated quadratic cases, the
|
|
352
|
+
actual mixed-PDF mass calibration, and the normal bromine example. The
|
|
353
|
+
mathematical-reference Actions job runs these at two meaningful uncertainty
|
|
354
|
+
digits and requires both interval kinds to stabilize. Independent population
|
|
355
|
+
means/uncertainties are checked within two predeclared adaptive tolerances.
|
|
356
|
+
This moment check does not establish population coverage of every returned
|
|
357
|
+
interval. The bounded local three-profile run uses one meaningful digit;
|
|
358
|
+
stronger full-profile execution is reported separately from Actions.
|
|
359
|
+
|
|
360
|
+
|
|
361
|
+
The extended reference job declares a maximum of eight million samples per
|
|
362
|
+
profile. Its first two-million-sample run did not stabilize the zero-mean,
|
|
363
|
+
correlation 0.9 quadratic case; that failure remains evidence of the resource
|
|
364
|
+
boundary. A convergence refusal includes the achieved sample count, numerical
|
|
365
|
+
tolerance, twice-standard-error diagnostics and PRNG state. The implementation
|
|
366
|
+
persists the original Welford recurrence in sample order, avoiding repeated
|
|
367
|
+
historical rescans while retaining the same binary64 prefix moments. Requested
|
|
368
|
+
stopping criteria and meaningful digits are unchanged.
|