Usage

Formatting

sciform provides two primary methods for formatting numbers into scientific formatted strings. The first is via the Formatter object and the second is using string formatting and the Format Specification Mini-Language (FSML) with the SciNum object.

Formatter

The Formatter object is initialized and configured using a number of formatting options described in Formatting Options. The Formatter object is then called with a number and returns a corresponding formatted string.

>>> from sciform import Formatter
>>> formatter = Formatter(
...     round_mode="dec_place", ndigits=6, upper_separator=" ", lower_separator=" "
... )
>>> print(formatter(51413.14159265359))
51 413.141 593
>>> formatter = Formatter(round_mode="sig_fig", ndigits=4, exp_mode="engineering")
>>> print(formatter(123456.78))
123.5e+03

It is not necessary to provide input for all options. At format time, any un-populated options will be populated with the corresponding options from the global options. See Global Options for details about how to view and modify the global options.

SciNum

The sciform FSML can be accessed via the SciNum object. Python numbers specified as string, int, float, or Decimal objects are cast to SciNum objects which can be formatted using the sciform FSML.

>>> from sciform import SciNum
>>> num = SciNum(123456)
>>> print(f"{num:!2f}")
120000

Value/Uncertainty Formatting

One of the most important use cases for scientific formatting is formatting a value together with its specified uncertainty, e.g. 84.3 ± 0.2. sciform provides the ability to format pairs of numbers into value/uncertainty strings. sciform attempts to follow BIPM or NIST recommendations for conventions when possible.

Value/uncertainty pairs can be formatted either by passing two numbers into a Formatter, configured with the corresponding Formatting Options and Value/Uncertainty Formatting Options, or by using the SciNum object.

>>> val = 84.3
>>> unc = 0.2
>>> formatter = Formatter(round_mode="sig_fig", ndigits=2)
>>> print(formatter(val, unc))
84.30 ± 0.20
>>> from sciform import SciNum
>>> num = SciNum(val, unc)
>>> print(f"{num:!2}")
84.30 ± 0.20

Value/uncertainty pairs can also be formatted using a parentheses notation in which the uncertainty is displayed in parentheses following the value. See BIPM Guide Section 7.2.2.

>>> print(f"{num:!2()}")
84.30(20)

Value/uncertainty pairs are formatted according to the following algorithm:

Rounding is always performed using significant figure rounding applied to the uncertainty. See Rounding for more details about possible rounding options.
The value is rounded to the decimal place corresponding to the least significant digit of the rounded uncertainty.
The value for the exponent is resolved by using exp_mode and exp_val with the larger of the value or uncertainty.
The value and uncertainty mantissas are determined according to the value of the exponent determined in the previous step.
The value and uncertainty mantissas are formatted together with the exponent according to other user-selected display options.

Formatted Input

Both the Formatter format function and SciNum constructor accept str, int, float, and Decimal input for the value and optionally accept the same types for the uncertainty.

>>> from decimal import Decimal
>>> from sciform import Formatter
>>>
>>> formatter = Formatter(round_mode="sig_fig", ndigits=4)
>>> print(formatter("32", "9"))
32.000 ± 9.000
>>> print(formatter(32, 9))
32.000 ± 9.000
>>> print(formatter(32.00, 9.00))
32.000 ± 9.000
>>> print(formatter(Decimal("32.000"), Decimal("9.000")))
32.000 ± 9.000

These example strings can be natively cast to numeric types like int, float, or Decimal. However, Formatter and SciNum also accept formatted strings which contain the numeric information, but with more sophisticated formatting.

>>> print(formatter("+  123_456,789 987 n"))
0.0001235

Here the string "+ 123_45,789 987 n" is re-interpreted by the Formatter as the number 12345.789987e-09. The e-09 arises as the reverse translation of the SI “nano” prefix n. When this is formatted in fixed point mode with 4 significant digits the result is "0.0001235".

Instead of passing in two inputs, one representing the value and one representing the uncertainty, it is possible to pass in one input which contains information about both the value and the uncertainty.

>>> print(formatter("(123.0 ± 0.4) m"))
0.1230000 ± 0.0004000
>>> print(formatter("(123 +/- 0.4) m"))
0.1230000 ± 0.0004000

Note that the first example input string is an example sciform output string that would have resulted from various options and inputs while the second example input string is not an example sciform output string. Any sciform output string can also be used as an input string. However, some input strings can be supplied which are not valid sciform outputs.

For inputs, it is not necessary that the value and uncertainty are expressed to the same lowest decimal place. By contrast, sciform outputs will always format the value and uncertainty to the same lowest decimal place.

For inputs, the ASCII +/- symbols may be used to separate the value from the uncertainty. By contrast, sciform will always use the Unicode ± symbol for direct outputs (noting that the ± symbol can be converted into +/- afterwards using FormattedNumber.as_ascii(), see Output Conversion).

Note that the formatting of the input strings has no bearing on the formatting of the resulting outputs.

Like sciform outputs, the input value/uncertainty pairs must always share the exponent string.

>>> print(formatter("(1.2 +/- 0.1)e+03"))
1200.0 ± 100.0
>>> print(formatter("1.2e+03 +/- 0.1e+03"))
Traceback (most recent call last):
  ...
ValueError: Input string "1.2e+03 +/- 0.1e+03" does not match any expected input format.

Inputs can use parentheses uncertainty notation and will be interpreted using the following rules.

The value string will always be directly converted to a float.

If the uncertainty string contains a decimal symbol then it will be directly converted to a float.

If the value string contains a decimal symbol but the uncertainty string does not then the string is parsed according to trimmed parentheses notation. In this case the uncertainty is left padded with a leading "0." and then a sufficient number of zeros so that the least significant digit of the uncertainty matches the least significant digit of the value.

If the uncertainty contains more digits than appear in the fractional part of the value then this padding is impossible and it is not clear what the value of the uncertainty is. In this case an exception is raised.

>>> print(formatter("1(100)"))
1.0 ± 100.0
>>> print(formatter("123.4(5.42)"))
123.400 ± 5.420
>>> print(formatter("123.4(5)"))
123.4000 ± 0.5000
>>> print(formatter("123.4(56)"))
Traceback (most recent call last):
  ...
ValueError: Invalid value/uncertainty pair for parentheses uncertainty: "123.4(56)". If a decimal symbol appears in the value but not in the uncertainty then the number of the digits in the uncertainty may not exceed the number of digits in the fractional part of the value.

If the input string contains information about the uncertainty then a second argument, which would otherwise specify the uncertainty, is not allowed

>>> print(formatter("123(4)", 4))
Traceback (most recent call last):
  ...
ValueError: Value input string "123(4)" already includes an uncertainty, (4). It is not possible to also pass in an uncertainty (4) directly.

As demonstrated above, the input string parser can parse translated exponents such as "n" -> "e-09". These translations are performed by first checking the default SI prefixes along with any global extra_si_prefixes, then checking the default parts-per forms along with any global extra_parts_per_forms. if no valid translations are discovered or more than one valid translation is discovered an exception is raised.

>>> from sciform import GlobalOptionsContext
>>> formatter = Formatter()
>>> with GlobalOptionsContext(add_c_prefix=True):
...     print(formatter("32 c"))
0.32
>>> print(formatter("32 c"))
Traceback (most recent call last):
  ...
ValueError: Unrecognized prefix: "c". Unable to parse input.
>>> with GlobalOptionsContext(extra_si_prefixes={-12: "ppb"}):
...     print(formatter("42 ppb"))
Traceback (most recent call last):
  ...
ValueError: Multiple translations found for "ppb": [-12, -9]. Unable to parse input.

In the final example, the SI translation for an exponent of -12 was re-mapped to "ppb" according to the “long scale” definition of billion. This collides with the “short scale” definition of billion that maps -9 to "ppb" as a default parts per form. This issue could be corrected as follows.

>>> with GlobalOptionsContext(
...     exp_mode="engineering",
...     extra_parts_per_forms={-9: None, -12: "ppb"}
... ):
...     print(formatter("42 ppb"))
42e-12

In most cases the parser can determine whether the decimal symbol is "." or ",".

If both "." and "," appear in a formatted value then whichever comes earlier must be the upper separator and whichever comes later must be the lower separator. E.g. 123.456,789 must have "," as the decimal separator.

If one separator appears more than once it must be the upper separator. E.g. 123,456,789 must have "." as the decimal separator.

If one separator appears once but it is preceded by more than 3 digits, or it is followed by any number of digits other than 3, then that separator must be the decimal separator. E.g. in both 1234.567 and 12.3 we must have that "." is the decimal separator.

In some cases the decimal separator cannot be determined, e.g. 123456, 123,456, 12.345.

sciform first tries to infer the decimal separator from the both the value and the uncertainty. If these inferences are successful and disagree then an exception is raised. If the inferences agree or only one succeeds then the succeeding result is selected as the decimal separator. If neither inference is successful then the decimal separator is selected either from the populated options on the local Formatter or from the global options.

>>> formatter = Formatter(decimal_separator=".")
>>> print(formatter("1234,567"))
1234.567
>>> print(formatter("123,45"))
123.45
>>> print(formatter("123,45 +/- 345.578"))
123 ± 345578
>>> print(formatter("12.45 +/- 2,34"))
Traceback (most recent call last):
  ...
ValueError: Value "12.45" and uncertainty "2,34" have different decimal separators.
>>> formatter = Formatter(decimal_separator=",")
>>> print(formatter("123,456"))
123,456
>>> formatter = Formatter()
>>> with GlobalOptionsContext(decimal_separator=","):
...     print(formatter("123,456"))
123,456

The same parsing rules apply to SciNum construction

>>> print(f'{SciNum("+  123_456,789 987 n"):!4f}')
0.0001235
>>> print(f'{SciNum("(123.0 ± 0.4) m"):!4f}')
0.1230000 ± 0.0004000
>>> print(f'{SciNum("(123 +/- 0.4) m"):!4f}')
0.1230000 ± 0.0004000
>>> print(f'{SciNum("123(4)")}')
123 ± 4

Output Conversion

Typically the output of the Formatter is used as a regular python string. However, the Formatter returns a FormattedNumber instance. The FormattedNumber class subclasses str and in many cases is used like a normal python string. However, the FormattedNumber class exposes methods to convert the standard string representation into LaTeX, HTML, or ASCII representations. The LaTeX and HTML representations may be useful when sciform outputs are being used in contexts outside of e.g. text terminals such as Matplotlib plots, Jupyter notebooks, or Quarto documents which support richer display functionality than Unicode text. The ASCII representation may be useful if sciform outputs are being used in contexts in which only ASCII, and not Unicode, text is supported or preferred.

These conversions can be accessed via the FormattedNumber.as_latex(), FormattedNumber.as_html(), and FormattedNumber.as_ascii() methods on the FormattedNumber class.

>>> formatter = Formatter(
...     exp_mode="scientific",
...     exp_val=-1,
...     upper_separator="_",
...     superscript=True,
... )
>>> formatted = formatter(12345)
>>> print(f"{formatted} -> {formatted.as_latex()}")
123_450×10⁻¹ -> $123\_450\times10^{-1}$
>>> print(f"{formatted} -> {formatted.as_html()}")
123_450×10⁻¹ -> 123_450×10<sup>-1</sup>
>>> print(f"{formatted} -> {formatted.as_ascii()}")
123_450×10⁻¹ -> 123_450e-01

>>> formatter = Formatter(
...     exp_mode="percent",
...     lower_separator="_",
... )
>>> formatted = formatter(0.12345678, 0.00000255)
>>> print(f"{formatted} -> {formatted.as_latex()}")
(12.345_678 ± 0.000_255)% -> $(12.345\_678\:\pm\:0.000\_255)\%$
>>> print(f"{formatted} -> {formatted.as_html()}")
(12.345_678 ± 0.000_255)% -> (12.345_678 ± 0.000_255)%
>>> print(f"{formatted} -> {formatted.as_ascii()}")
(12.345_678 ± 0.000_255)% -> (12.345_678 +/- 0.000_255)%

>>> formatter = Formatter(
...     exp_mode="engineering",
...     exp_format="prefix",
...     round_mode="sig_fig",
...     ndigits=4,
... )
>>> formatted = formatter(314.159e-6, 2.71828e-6)
>>> print(f"{formatted} -> {formatted.as_latex()}")
(314.159 ± 2.718) μ -> $(314.159\:\pm\:2.718)\:\text{\textmu}$
>>> print(f"{formatted} -> {formatted.as_html()}")
(314.159 ± 2.718) μ -> (314.159 ± 2.718) μ
>>> print(f"{formatted} -> {formatted.as_ascii()}")
(314.159 ± 2.718) μ -> (314.159 +/- 2.718) u

The LaTeX enclosing "$" math environment symbols can be optionally stripped:

>>> formatter = Formatter(
...     exp_mode="engineering",
...     exp_format="prefix",
...     round_mode="sig_fig",
...     ndigits=4,
... )
>>> formatted = formatter(314.159e-6, 2.71828e-6)
>>> print(f"{formatted} -> {formatted.as_latex(strip_math_mode=False)}")
(314.159 ± 2.718) μ -> $(314.159\:\pm\:2.718)\:\text{\textmu}$
>>> print(f"{formatted} -> {formatted.as_latex(strip_math_mode=True)}")
(314.159 ± 2.718) μ -> (314.159\:\pm\:2.718)\:\text{\textmu}

In addition to exposing FormattedNumber.as_latex() and FormattedNumber.as_html(), the FormattedNumber class defines the aliases FormattedNumber._repr_latex_() and FormattedNumber._repr_html_(). The IPython display functions looks for these methods, and, if available, will use them to display prettier representations of the class than the Unicode __repr__ representation. Here is example FormattedNumber usage in a Jupyter notebook.

Global Options

It is possible to modify the global options for sciform to avoid repetition of verbose configuration options or format specification strings. When the user creates a Formatter object or formats a string using the FSML, they typically do not specify settings for all available options. In these cases, the unpopulated options resolve their values from the global options at format time.

The sciform default global options can be viewed using get_default_global_options()

>>> from sciform import get_default_global_options
>>> print(get_default_global_options())
PopulatedOptions(
 'exp_mode': 'fixed_point',
 'exp_val': 'auto',
 'round_mode': 'all',
 'ndigits': 2,
 'upper_separator': '',
 'decimal_separator': '.',
 'lower_separator': '',
 'sign_mode': '-',
 'left_pad_char': ' ',
 'left_pad_dec_place': 0,
 'exp_format': 'standard',
 'extra_si_prefixes': {},
 'extra_parts_per_forms': {},
 'capitalize': False,
 'superscript': False,
 'nan_inf_exp': False,
 'paren_uncertainty': False,
 'left_pad_matching': False,
 'paren_uncertainty_trim': True,
 'pm_whitespace': True,
)

The global options can be modified using the set_global_options() function. Any options passed will overwrite the corresponding options in the current global options and any unfilled options will remain unchanged. The current global options can be viewed using get_global_options(). The global options can be reset to the sciform default global options using reset_global_options().

>>> from sciform import set_global_options, get_global_options, reset_global_options
>>> set_global_options(
...     left_pad_char="0",
...     exp_mode="engineering_shifted",
...     round_mode="sig_fig",
...     ndigits=4,
...     decimal_separator=",",
... )
>>> print(get_global_options())
PopulatedOptions(
 'exp_mode': 'engineering_shifted',
 'exp_val': 'auto',
 'round_mode': 'sig_fig',
 'ndigits': 4,
 'upper_separator': '',
 'decimal_separator': ',',
 'lower_separator': '',
 'sign_mode': '-',
 'left_pad_char': '0',
 'left_pad_dec_place': 0,
 'exp_format': 'standard',
 'extra_si_prefixes': {},
 'extra_parts_per_forms': {},
 'capitalize': False,
 'superscript': False,
 'nan_inf_exp': False,
 'paren_uncertainty': False,
 'left_pad_matching': False,
 'paren_uncertainty_trim': True,
 'pm_whitespace': True,
)
>>> reset_global_options()

The global options can be temporarily modified using the GlobalOptionsContext context manager. The context manager is configured using the same options as Formatter and set_global_options(). Within the context of GlobalOptionsContext manager, the global options take on the specified input settings, but when the context is exited, the global options revert to their previous values.

>>> from sciform import GlobalOptionsContext, SciNum
>>> num = SciNum(0.0123)
>>> print(f"{num:.2ep}")
1.23e-02
>>> with GlobalOptionsContext(add_c_prefix=True):
...     print(f"{num:.2ep}")
1.23 c
>>> print(f"{num:.2ep}")
1.23e-02

Note that the FSML does not provide complete control over all possible format options. For example, there is no code in the FSML for configuring the pdg_sig_figs option. If the user wishes to configure these options, but also use the FSML, then they must do so by modifying the global options.

Formatter Options

The Formatter options are configured by constructing a Formatter and passing in keyword arguments corresponding to the desired options. Only a subset of available options need be specified. The user input during Formatter construction can be viewed using the Formatter.input_options property.

>>> formatter = Formatter(
...     exp_mode="engineering",
...     round_mode="sig_fig",
...     ndigits=2,
...     superscript=True,
... )
>>> print(formatter.input_options)
InputOptions(
 'exp_mode': 'engineering',
 'round_mode': 'sig_fig',
 'ndigits': 2,
 'superscript': True,
)

The Formatter.input_options property is a InputOptions instance. The string representation of this object indicates the explicitly populated options. A dictionary of these populated options is available via the InputOptions.as_dict() method.

>>> print(formatter.input_options.as_dict())
{'exp_mode': 'engineering', 'round_mode': 'sig_fig', 'ndigits': 2, 'superscript': True}

Both populated and unpopulated options can be accessed by direct attribute access.

>>> print(formatter.input_options.round_mode)
sig_fig
>>> print(formatter.input_options.exp_format)
None

In addition to viewing the user input options, it is possible to preview the result of populating the unpopulated options with the corresponding global options by using the Formatter.populated_options property.

>>> print(formatter.populated_options)
PopulatedOptions(
 'exp_mode': 'engineering',
 'exp_val': 'auto',
 'round_mode': 'sig_fig',
 'ndigits': 2,
 'upper_separator': '',
 'decimal_separator': '.',
 'lower_separator': '',
 'sign_mode': '-',
 'left_pad_char': ' ',
 'left_pad_dec_place': 0,
 'exp_format': 'standard',
 'extra_si_prefixes': {},
 'extra_parts_per_forms': {},
 'capitalize': False,
 'superscript': True,
 'nan_inf_exp': False,
 'paren_uncertainty': False,
 'left_pad_matching': False,
 'paren_uncertainty_trim': True,
 'pm_whitespace': True,
)

The Formatter.populated_options property is a PopulatedOptions instance. It is recalculated each time the property is accessed so that the output always reflects the current global options. Like the InputOptions class, the PopulatedOptions class provides access to its options via direct attribute access and via a PopulatedOptions.as_dict() method.

The FormattedNumber class stores a record of the PopulatedOptions that were used to generate it.

>>> formatter = Formatter(
...     exp_mode="engineering",
...     round_mode="sig_fig",
...     ndigits=2,
...     superscript=True,
... )
>>> formatted = formatter(12345.678, 3.4)
>>> print(formatted)
(12.3457 ± 0.0034)×10³
>>> print(formatted.populated_options)
PopulatedOptions(
 'exp_mode': 'engineering',
 'exp_val': 'auto',
 'round_mode': 'sig_fig',
 'ndigits': 2,
 'upper_separator': '',
 'decimal_separator': '.',
 'lower_separator': '',
 'sign_mode': '-',
 'left_pad_char': ' ',
 'left_pad_dec_place': 0,
 'exp_format': 'standard',
 'extra_si_prefixes': {},
 'extra_parts_per_forms': {},
 'capitalize': False,
 'superscript': True,
 'nan_inf_exp': False,
 'paren_uncertainty': False,
 'left_pad_matching': False,
 'paren_uncertainty_trim': True,
 'pm_whitespace': True,
)

Formatter Options Edge Cases

In most cases, at format/option population time, any non-None options in the InputOptions will be exactly copied over to the PopulatedOptions and any None options will be exactly copied over from the global options at format time. However, a few options have slightly more complicated behavior.

>>> from sciform import set_global_options, get_global_options, reset_global_options
>>> set_global_options(extra_si_prefixes={-2: "cm"})
>>> print(get_global_options().extra_si_prefixes)
{-2: 'cm'}
>>> formatter = Formatter(add_c_prefix=True)
>>> print(formatter.input_options.extra_si_prefixes)
None
>>> print(formatter.populated_options.extra_si_prefixes)
{-2: 'c'}
>>> reset_global_options()

Somewhat surprisingly, even though extra_si_prefixes is unpopulated in the Formatter, it does not get populated with the corresponding global options extra_si_prefixes. This is because the Formatter has add_c_prefix=True. If extra_si_prefixes=None but add_c_prefix=True then the same population behavior as if extra_si_prefixes={-2: 'c'} is realized. If extra_si_prefixes is a dictionary then {-2: 'c'} is added to the dictionary if -2 is not already a key in the dictionary. If -2 already appears in the dictionary then its value is not overwritten.

>>> set_global_options(extra_si_prefixes={-2: "cm"})
>>> print(get_global_options().extra_si_prefixes)
{-2: 'cm'}
>>> formatter = Formatter(extra_si_prefixes={-15: 'fermi'}, add_c_prefix=True)
>>> print(formatter.input_options.extra_si_prefixes)
{-15: 'fermi'}
>>> print(formatter.populated_options.extra_si_prefixes)
{-15: 'fermi', -2: 'c'}
>>> reset_global_options()

>>> formatter = Formatter(extra_si_prefixes={-2: 'cm'}, add_c_prefix=True)
>>> print(formatter.input_options.extra_si_prefixes)
{-2: 'cm'}
>>> print(formatter.populated_options.extra_si_prefixes)
{-2: 'cm'}
>>> reset_global_options()

Analogous behavior occurs for the add_small_si_prefixes and add_ppth_form Formatter options. Note also that these three options, add_c_prefix, add_small_si_prefixes, and add_ppth_form appear in the InputOptions instance if they have been explicitly set, but they never appear in the PopulatedOptions instance. Rather, only the populated extra_si_prefixes or extra_parts_per_forms are appropriately populated.

Finally, if integer 0 is passed into left_pad_char then integer 0 will be stored in the InputOptions, but it will be converted to string "0" in the PopulatedOptions.

>>> formatter = Formatter(left_pad_char=0)
>>> print(type(formatter.input_options.left_pad_char))
<class 'int'>
>>> print(type(formatter.populated_options.left_pad_char))
<class 'str'>

Note on Decimals and Floats

Numerical data can be stored in Python float or Decimal objects. float instances represent numbers using binary which means they are often only approximations of the decimal numbers users have in mind when they use float. By contrast, Decimal objects store sequences of integers representing the decimal digits of the represented numbers so, Decimal instances are, therefore, exact representations of decimal numbers.

Both of these representations have finite precision which can cause unexpected issues when manipulating numerical data. However, in the Decimal class, the main issue is that numbers may be truncated if their precision exceeds the configured Decimal precision, but the rounding will be as expected. That said, the precision used for Decimal numbers can easily be modified if necessary. float instances, unfortunately, may exhibit more surprising behavior, as will be explained below. For these reasons, the sciform module uses Decimal representations in its internal formatting algorithms.

Note, however, that Decimal arithmetic operations are less performant that float operations. So, unless very high precision is needed at all steps of the calculation, the suggested workflow is to store and manipulate numerical data as float instances, and only convert to Decimal, or format using sciform, as the final step when numbers are being displayed for human readers.

Float Issues

Here we would like to highlight some important facts and possible issues with float objects that users should be aware of if they are concerned with the exact decimal representation of their numerical data.

Python uses double-precision floating-point format for its float. In this format, a float occupies 64 bits of memory: 52 bits for the mantissa, 11 bits for the exponent and 1 bit for the sign.
Any decimal with 15 digits between about ± 1.8e+308 can be uniquely represented by a float. However, two decimals with more than 15 digits may map to the same float. For example, float(8.000000000000001) == float(8.000000000000002) returns True. See “Decimal Precision of Binary Floating Point Numbers” for more details.
If any float is converted to a decimal with at least 17 digits then it will be converted back to the same float. See “The Shortest Decimal String that Round-Trips: Examples” for more details. However, many float instances can be “round-tripped” with far fewer digits. The __repr__() for the python float class converts the float to a string decimal representation with the minimum number of digits such that it round trips to the same float. For example we can see the exact decimal representation of the float to which 0.1 is mapped: print(Decimal(float(0.1))) gives "0.1000000000000000055511151231257827021181583404541015625". However print(float(0.1)) just gives "0.1". That is, 0.1000000000000000055511151231257827021181583404541015625 and 0.1 map to the same float but the float __repr__() algorithm presents us with the shorter (more readable) decimal representation.

The python documentation goes into some detail about possible issues one might encounter when working with float instances. Here we would like to highlight two specific issues.

Rounding. Python’s round() function uses a “round-to-even” or “banker’s rounding” strategy in which ties are rounded so the least significant digit after rounding is always even. This ensures data sets with uniformly distributed digits are not biased by rounding. Rounding of float instances may have surprising results. Consider the decimal numbers 0.0355 and 0.00355. If we round these to two significant figures using a “round-to-even” strategy, we expect the results 0.036 and 0.0036 respectively. However, if we try to perform this rounding for float we get an unexpected result. We see that round(0.00355, 4) gives 0.0036 as expected but round(0.0355, 3) gives 0.035. We can see the issue by looking at the decimal representations of the corresponding float instances. print(Decimal(0.0355)) gives "0.035499999999999996835864379818303859792649745941162109375" which indeed should round down to 0.035 while print(Decimal(0.00355)) gives "0.003550000000000000204003480774872514302842319011688232421875" which should round to 0.0036. So, we see that the rounding behavior for float may depend on digits of the decimal representation of the float which are beyond the minimum number of digits necessary for the float to round trip and, thus, beyond the number of digits that will be displayed by default.
Representation of numbers with high precision. Conservatively, float provides 15 digits of precision. That is, any two decimal numbers (within the float range) with 15 or fewer digits of precision are guaranteed to correspond to unique float instances. Decimal numbers with 16 digits or more of precision may not correspond to unique float instances. It is rare, in scientific applications, that we require more than 15 digits of precision, but in some cases we do. One example is precision frequency metrology, such as that involved in atomic clocks. The relative uncertainty of primary frequency standards is approaching one part in 10^-16. This means that measured quantities may require up to 16 digits to display. Indeed, consider Metrologia 55 (2018) 188–200. In Table 2 the ⁸⁷ Rb ground-state hyperfine splitting is cited as 6 834 682 610.904 312 6 Hz with 17 digits. Suppose the last digit was a 5 instead of a 6. Python float cannot tell the difference: float(6834682610.9043126) == float(6834682610.9043125) returns True.

How `sciform` Handles Decimals and Floats

To support predictable rounding and the representation of high precision numbers, sciform casts the numbers it is presenting to Decimal objects during its formatting algorithm. Numbers are input into sciform either as the input to a Formatter or when instantiating a SciNum object. In all cases the input will typically be a Decimal, float, str, or int. Decimal, str and int are unambiguously converted to Decimal objects. For float inputs, the values are first cast to str instances to get their shortest round-trippable decimal representations. These shortest round-trippable strings are then converted into Decimal instances. For high precision applications it is recommended that users provide input to sciform either as str or Decimal.