Make decorated interval the default (#590)

* Make decorated interval the default * Remove warning * Fix decoration * Fix tests * Fix tests * Fix NaI * Minor cleanup * Add docstring for `_unsafe_bareinterval` * Remove decoration `bad` in favour of a new field * Minor cleanup * Remove `signbit` * Fix `pow` and cleanup * Fix tests * Special display for unguaranteed interval * Cleanup * Update docs * Fix doctest * Do not error for missing docstring * Rename `guarantee` to `isguaranteed` * Delete empty file decorations.md * Fix missing field infos in docstring * Update `isguaranteed` docstring * Update docs * Ensures `NaN` is converted to a NaI * Update docs * Minor cleanup * Fix typo * Fix error for boolean functions from Base * Add docstrings * Fix bug * Fix `StackOverflow` * Fix `StackOverFlow` * Set correct cuurent version
JuliaIntervals · Dec 1, 2023 · 742a26f · 742a26f
1 parent c07cb3f
commit 742a26f
Show file tree

Hide file tree

Showing 78 changed files with 12,514 additions and 12,154 deletions.
diff --git a/Project.toml b/Project.toml
@@ -1,7 +1,7 @@
 name = "IntervalArithmetic"
 uuid = "d1acc4aa-44c8-5952-acd4-ba5d80a2a253"
 repo = "https://github.com/JuliaIntervals/IntervalArithmetic.jl.git"
-version = "0.21.1"
+version = "0.21.2"
 
 [deps]
 CRlibm = "96374032-68de-5a5b-8d9e-752f78720389"

diff --git a/docs/make.jl b/docs/make.jl
@@ -22,10 +22,10 @@ makedocs(;
         "Manual" => [
             "Constructing intervals" => "manual/construction.md",
             "Usage" => "manual/usage.md",
-            "Decorations" => "manual/decorations.md",
             "API" => "manual/api.md"
         ]
-    ]
+    ],
+    warnonly = true
 )
 
 deploydocs(;

diff --git a/docs/src/index.md b/docs/src/index.md
@@ -25,18 +25,3 @@ If you use the IntervalArithmetic library in your publication, research, teachin
 
 - [MPFI.jl](https://github.com/andrioni/MPFI.jl), a Julia wrapper around the [MPFI C library](http://perso.ens-lyon.fr/nathalie.revol/software.html), a multiple-precision interval arithmetic library based on MPFR
 - [Intervals.jl](https://github.com/invenia/Intervals.jl), an alternative implementation of basic interval functions by Invenia Technical Computing
-
-
-
-## History
-
-This project began in 2014 during a masters' course in the postgraduate programs of Mathematics and Physics at the Universidad Nacional Autónoma de México. It was initially written in Python, then reinitiated in 2015 and rewritten in Julia. We thank the participants of the courses for their contribution, energy and ideas.
-
-
-## Support
-
-Financial support is acknowledged from DGAPA-UNAM PAPIME grants PE-105911 and PE-107114, and DGAPA-UNAM PAPIIT grant IN-117214.
-
-Luis Benet acknowledges support from *Cátedra Marcos Moshinsky* (2013).
-
-David P. Sanders acknowledges a sabbatical fellowship from CONACYT and thanks Alan Edelman and the Julia group at MIT for hosting his sabbatical visit.
diff --git a/docs/src/manual/construction.md b/docs/src/manual/construction.md
@@ -1,46 +1,188 @@
-# Constructing intervals
-
-Constructing an interval is the most basic operation in the library. The [`interval`](@ref) constructor is the standard way to create an interval. It accepts one or two values, and an optional bound type.
+The library provides two interval types. The first one is [`BareInterval`](@ref), corresponding to a basic implementation of intervals, stored by their infimum and supremum. The second type is [`Interval`](@ref) and builds on top of bare intervals, with the additional fields `decoration` and `isguaranteed`. See the sections below.
 
 ```@repl construction
 using IntervalArithmetic
-setformat(:full) # print the interval in full
-interval(0.1) # interval(Float64, 0.1)
-interval(0.1, 0.2) # interval(Float64, 0.1, 0.2)
-interval(3.1f0) # interval(Float32, 3.1f0)
-interval(π) # interval(Float64, π)
-interval(BigFloat, π)
-interval(Inf) # not valid since infinity is not part of an interval
-interval(3, 2) # not valid since the lower bound is strictly greater than the upper bound
+setdisplay(:full) # print the interval in full
+bareinterval(1, π) # `bareinterval(Float64, 1, π)`
+interval(1, π) # `interval(Float64, 1, π)`, interval decorated with `com` (common)
 ```
 
-The submodule `IntervalArithmetic.Symbols` exports the infix operator `..` as an alias for `interval`.
+Therefore, we **strongly recommend the use of [`Interval`](@ref) over [`BareInterval`](@ref)** to better track the effect of functions according to the IEEE Standard 1788-2015 specifications. For instance, taking the square root of an interval discards the negative part of the interval, without any notice for bare intervals:
 
 ```@repl construction
-using IntervalArithmetic.Symbols
-0.1..0.2 # interval(0.1, 0.2)
+sqrt(bareinterval(-1, 1)) # `sqrt(bareinterval(0, 1))`
+sqrt(interval(-1, 1)) # interval decorated with `trv` (trivial)
+```
+
+
+
+### Decorations
+
+A *decoration* is a label that indicates the status of a given interval. Decorated intervals provide valuable information on the result of evaluating a function on an initial interval.
+
+Upon the application of a function ``f`` on an interval ``x``, the resulting interval ``f(x)`` has either one of the following decorations:
+
+- `com` (common): ``x`` is a closed, bounded, non-empty subset of the domain of ``f``, ``f`` is continuous on the interval ``x``, and ``f(x)`` is bounded.
+
+- `dac` (defined and continuous): ``x`` is a non-empty subset of the domain of ``f``, and ``f`` is continuous on ``x``.
+
+- `def` (defined): ``x`` is a non-empty subset of the domain of ``f``; in other words, ``f`` is defined at each point of ``x``.
+
+- `trv` (trivial): ``f(x)`` carries no meaningful information.
+
+- `ill` (ill-formed): ``f(x)`` is Not an Interval (NaI).
+
+Each decoration is paired with an integer as follows: `ill = 0`, `trv = 1`, `def = 2`, `dac = 3` and `com = 4`. Then, decorations degrade according to the propagation order `com > dac > def > trv > ill`.
+
+One can specify a decoration when constructing intervals. Otherwise, the interval is initialised with a decoration according to the underlying bare interval:
+
+- `com`: non-empty and bounded.
+
+- `dac`: unbounded.
+
+- `trv`: empty.
+
+- `ill`: NaI.
+
+
+#### Examples
+
+##### Common
+
+```@repl construction
+x = interval(0.5, 3)
+sqrt(x)
+```
+
+Both input `x` and output `sqrt(x)` are common intervals since they are closed, bounded, non-empty and that ``\sqrt`` is continuous over ``[1/2, 3]``.
+
+Observe that these decorations, together with the fact that any element of the interval `sqrt(x)` is also in the interval `x`, imply that the [Schauder Fixed-Point Theorem](https://en.wikipedia.org/wiki/Schauder_fixed-point_theorem) is satisfied. More precisely, this computation proves the existence of a fixed-point of ``\sqrt`` in ``[1/2, 3]`` (in this simple example, ``\sqrt(1) = 1``).
+
+##### Defined and continuous
+
+```@repl construction
+x = interval(3, Inf)
+sqrt(x)
+```
+
+Both the intervals are unbounded, hence the maximum possible decoration is `dac`.
+
+Note that overflows can also produce the decoration `dac`:
+
+```@repl construction
+x = interval(floatmax(Float64))
+x + interval(1)
 ```
 
-The [`±`](@ref) (`\pm<tab>`) infix operator creates the interval from the midpoint and the radius.
+##### Defined and continuous
 
 ```@repl construction
-0 ± 1
+x = interval(-3, 4)
+sign(x)
 ```
 
-The various string formats are as follows:
-- No string parameter or `Empty` string ("[Empty]") returns an empty interval.
-- `entire` ("[entire]") and "[,]" string returns entireinterval
-- "[nai]" returns `Nai{Type}`
-- "[m]" returns `Interval(m,m)`
-- "[l, r]" returns `Interval(l, r)`
-- "m?r" returns `Interval(m-r, m+r)`
-- "m?ren" returns `Interval((m-r)en, (m+r)en)`
-- "m?ru" or "m?rd" returns `Interval(m, m+r)` or `Interval(m-r, m)` respectively
-- "m?" returns `Interval(m + 5 precision units, m - 5 precision units)`
-- "m??" returns `Interval(-Inf, +Inf)`
-- "m??u" or "m??d" returns `Interval(m, +Inf)` or `Interval(-Inf, m) respectively`
-
-!!! warning
+The ``\sign`` function is discontinuous at ``0``, but is defined everywhere on the input interval, so the decoration of the result is `def`.
+
+##### Trivial
+
+```@repl construction
+x = interval(-3.5, 4)
+sqrt(x)
+```
+
+The negative part of `x` is discarded before evaluating the ``\sqrt`` function since its domain is ``[0, \infty)``. The process of discarding parts of an interval that are not in the domain of a function is called *loose evaluation*. This event has been recorded by degrading the decoration of the resulting interval to `trv`, indicating that nothing is known about the relationship between `x` and `sqrt(x)`.
+
+In this case, we know why the decoration was reduced to `trv`. Generally, if this were just a single step in a longer calculation, a resulting decoration `trv` shows only that something like this occured at some step.
+
+For instance,
+
+```@repl construction
+f = asin ∘ sqrt
+x = interval(-3, 3)
+f(x)
+y = interval(0, 3)
+f(y)
+```
+
+In both cases, `asin(sqrt(X))` gives a result with the decoration `trv`; to find out where things went wrong, the function must be analyzed.
+
+```@repl construction
+sqrt(x) # `f(x)` has the decoration is `trv` since `x` contains negative values
+sqrt(y) # the decoration is `com`
+asin(sqrt(y)) # `f(x)` has the decoration is `trv` since `sqrt(y)` contains values stricly greater than `1`
+```
+
+This shows that loose evaluation occurred in different parts of `f` for `x` and `y`.
+
+!!! danger
+    The decoration `trv` is an indicator of information loss. Often this also reveals that something unexpected occured. Therefore, any interval marked by this decoration may not be trusted and the code may need to be revised.
+
+##### Ill-formed
+
+```@repl construction
+interval(2, 1)
+interval(NaN)
+```
+
+These are all examples of ill-formed intervals, resulting in the decoration `ill`.
+
+!!! danger
+    The decoration `ill` is an indicator that an error has occured. Therefore, any interval marked by this decoration cannot be trusted and the code needs to be debugged.
+
+
+
+### Guarantee
+
+A *guarantee* is yet another label, independent of decorations, and not described by the IEEE Standard 1788-2015 specifications. Its purpose is to accomodate for Julia's extensive conversion and promotion system, while retaining reliability in computations. Specifically, an interval `x` constructed via [`interval`](@ref) satisfies `isguaranteed(x) == true`. However, if a call to `convert(::Type{<:Interval}, ::Real)` occurs, then the resulting interval `x` satisfies `isguaranteed(x) == false`.
+
+In contrast, a [`BareInterval`](@ref) can only be constructed via [`bareinterval`](@ref), it is not a subtype of `Real`, and there are no allowed conversion with `Number`. Thus, this interval type is always guaranteed.
+
+!!! danger
+    A user interested in validated numerics should **always** have a resulting interval for which [`isguaranteed`](@ref) is `true`.
+
+
+
+## More constructors
+
+The submodule `IntervalArithmetic.Symbols` exports the infix operator `..` and `±` as an alias for `interval`; this submodule must be explicitly imported.
+
+```@repl construction
+using IntervalArithmetic.Symbols
+0.1 .. 0.2 # interval(0.1, 0.2; format = :standard)
+0.1 ± 0.2 # interval(0.1, 0.2; format = :midpoint)
+```
+
+Moreover, one can parse strings into intervals. The various string formats are the following:
+
+- `"[m]"` is equivalent to `interval(m, m)`.
+
+- `"[l, r]"` is equivalent to `interval(l, r)`.
+
+- `"m?r"` is equivalent to `interval(m-r, m+r)`.
+
+- `"m?ren"` is equivalent to `interval((m-r)*1en, (m+r)*1en)`.
+
+- `"m?rd"` is equivalent to `interval(m-r, m)`.
+
+- `"m?ru"` is equivalent to `interval(m, m+r)`.
+
+- `"m?"` is equivalent to `interval(m + 5 precision units, m - 5 precision units)`.
+
+- `"m??"` is equivalent to `interval(-Inf, +Inf)`.
+
+- `"m??d"` is equivalent to `interval(-Inf, m)`.
+
+- `"m??u"` is equivalent to `interval(m, +Inf)`.
+
+- `"[Entire]"`, `"[entire]"` and `"[,]"` are equivalent to `entireinterval()`.
+
+- `"[Empty]"`, `"[empty]"` and `"[]"` are equivalent to `emptyinterval()`.
+
+- `"[nai]"` and any other unsupported string formats are equivalent to `nai()`.
+
+To add a specific decoration, add `"_com"`, `"_dac"`, `"_dec"`, `"_trv"` and `"_ill"` at the end of the string.
+
+!!! danger
     Most real numbers cannot be exactly represented by floating-points. In such cases, the literal expression is rounded at parse time. To construct an interval enclosing the true real number, one must rely on the string constructor mentioned above.
 
     For instance, consider
@@ -49,15 +191,15 @@ The various string formats are as follows:
     x = 0.1
     ```
 
-    This appears to store the real number ``1/10`` in a variable `x` of type `Float64`. Yet,
+    This appears to store the real number ``1/10``  in a variable `x` of type `Float64`. Yet,
 
     ```@repl construction
     x > 1//10
     ```
 
-    Hence, the floating-point `0.1` is (slightly) greater than the real number ``1/10`` since ``1/10`` *cannot be represented exactly in binary floating-point arithmetic, at any precision*. The true value must be approximated by a floating-point number with fixed precision -- this procedure is called rounding.
+    Hence, the floating-point `0.1` is (slightly) greater than the real number ``1/10`` since ``1/10`` **cannot be represented exactly in binary floating-point arithmetic, at any precision**. The true value must be approximated by a floating-point number with fixed precision -- this procedure is called rounding.
 
-    In particular, this implies that `interval(0.1)` *does not* contain the real number ``1/10``. A valid interval containing the real number ``1/10`` can be constructed by
+    In particular, this implies that `interval(0.1)` **does not** contain the real number ``1/10``. A valid interval containing the real number ``1/10`` can be constructed by
 
     ```@repl construction
     I"0.1"