Floating point

{admiral} recently ran into some trouble when dealing with floating point values, captured by this thread on GitHub. This post gives a brief overview on floating point values, recaps the discussion on GitHub, and explains how {admiral} deals with floating point values.

Floating point values

Floating point values are numeric objects representing numbers between integers, e.g. 0.5, 2.3, 3.1415, etc. However, floating point numbers are not stored like integers, and most floating point numbers are approximations to the number they represent. To see what value a floating point number is actually stored as, we can use the format() function where we can increase the number of digits shown:

format(1.4, digits = 22)

[1] "1.399999999999999911182"

These very small numerical differences impact the result of mathematical operations:

0.1 + 0.2 == 0.3

[1] FALSE

If we look at the actually stored values, this makes sense:

0.1 %>% format(digits = 22)

[1] "0.1000000000000000055511"

0.2 %>% format(digits = 22)

[1] "0.2000000000000000111022"

(0.1 + 0.2) %>% format(digits = 22)

[1] "0.3000000000000000444089"

0.3 %>% format(digits = 22)

[1] "0.2999999999999999888978"

The bottom line is: Avoid using exact comparators such as == and >= when comparing floating point values.

Exact floating point values

Floating point values are stored in binary format. While most floating point values are approximations, there are some exceptions which can be exactly represented, namely if they can be written down as \(\frac{x}{2^y}\), where x and y are integers. For example, 0.5 is stored as \(\frac{1}{2}\), 0.25 is stored as \(\frac{1}{4}\), 0.125 is stored as \(\frac{1}{8}\), etc.

# simple examples
0.5 %>% format(digits = 22)

[1] "0.5"

0.25 %>% format(digits = 22)

[1] "0.25"

0.125 %>% format(digits = 22)

[1] "0.125"

0.0625 %>% format(digits = 22)

[1] "0.0625"

# some weird values for x and y
(1121 / (2^9)) %>% format(digits = 22)

[1] "2.189453125"

All floating point values are stored as \(\frac{x}{2^y}\), where the outcome may be a very close approximation to the value they represent*.

https://en.wikipedia.org/wiki/Floating-point_arithmetic#Representable_numbers,_conversion_and_rounding If you would like to learn more about representable floating point values please read the wikipedia article on floating point values, especially section Representable numbers, conversion and rounding.

* Based on a recollection of the course associated with this GitHub Repository by Martin Mächler.

Issues arising

Gordon Miller came across this issue when he was creating DAIDS criteria for adverse events in cancer therapy when using case_when statements to implement the grade.

We can have a glimpse here:

atoxgr_criteria_daids %>%
  filter(TERM %in% c("Amylase, High", "Lipase, High")) %>%
  select(TERM, GRADE_CRITERIA_CODE) %>%
  reactable(defaultPageSize = 4, highlight = TRUE, bordered = TRUE, striped = TRUE, resizable = TRUE)

As you can see, the data-frame contains the column GRADE_CRITERIA_CODE which contains comparisons of floating point values. And there was a discrepancy of what Gordon expected to see, and how R actually computed the comparison initially:

The test is AVAL >= 1.1*ANRHI should give a value of “1” where AVAL = 110 and ANRHI = 100.

I tried it separately and I also got 1.1*ANRHI not equal to 110 where ANRHI = 100.

Where ANRHI is the analysis range upper limit and AVAL is an analysis value.

What happened here? Gordon Miller wanted to compute the analysis range upper limit plus 10% and compare it to the analysis value. He expected the comparison to yield TRUE (or 1 if converted to numeric) as AVAL (110) should be exactly 1.1 * 100. However, he multiplied an integer (100) with a floating point value (1.1). And the result was not exactly 110, as 1.1 is not exactly represented as a floating point value.

(1.1 * 100) %>% format(digits = 22)

[1] "110.0000000000000142109"

1.1 * 100 == 110

[1] FALSE

On my machine, the result is actually larger than 110, while on Gordon Miller’s machine the result was smaller than 110. In {admiral}, we strive towards removing platform specific and unexpected behavior, so we had to find a way to solve the floating point issue.

Potential solutions

A very crude option would be to round the result of the multiplication to the nearest integer.

round(1.1 * 100) %>% format(digits = 22)

[1] "110"

However, this does not work when the result is not an integer, i.e. the upper limit was 101 instead. We should then compare the analysis value to 101 * 1.1, which should be exactly 111.1. We could try to round to the nearest decimal place, but that value would again be stored as a floating point value:

(101 * 1.1) %>%
  round(digits = 1) %>%
  format(digits = 22)

[1] "111.0999999999999943157"

A workaround would be to multiply both sides of the equation with 10, and then round to the next integer:

(101 * 1.1 * 10) %>%
  round() %>%
  format(digits = 22)

[1] "1111"

(111.1 * 10) %>%
  round() %>%
  format(digits = 22)

[1] "1111"

This is very awkward, as you don’t know by how much you need to multiply each time, a very clunky solution.

Alternatively, we can compare the absolute value of the difference between the analysis value and the upper limit plus 10% to a very small number, e.g. 0.0000001:

AVAL <- 111.1
ANRHI <- 101

abs(AVAL - ANRHI * 1.1) < 0.0000001

[1] TRUE

Comparing to a very small value is also how the all.equal() function works, which compares two numeric values and returns TRUE if they are equal within a tolerance. By default the tolerance is around \(1.5 * 10^{-8}\) but you can set it yourself to a lower value, e.g. machine tolerance .Machine$double.eps - (one of**) the smallest positive floating-point number x such that 1 + x != 1.

1 + .Machine$double.eps == 1

[1] FALSE

# but:
1 + .Machine$double.eps / 2 == 1

[1] TRUE

# so we can use:
all.equal(AVAL, ANRHI * 1, 1, tolerance = .Machine$double.eps)

[1] "Mean absolute difference: 10.1"

This would still be a little clunky for greater than or equal to comparisons:

all.equal(AVAL, ANRHI * 1.1) | AVAL > ANRHI * 1.1

[1] TRUE

# unfortunately, the all.equal() function does not return a FALSE if they are not the same:
all.equal(AVAL, ANRHI * 1.1 + 1)

[1] "Mean relative difference: 0.0090009"

For some reason, the value it returns is also not correct.

There is also a dplyr function called near() which does essentially the same thing as all.equal():

ANRHI <- 100
AVAL <- 110
(ANRHI * 1.1) %>% format(digits = 22)

[1] "110.0000000000000142109"

AVAL > ANRHI * 1.1 | near(AVAL, ANRHI * 1.1)

[1] TRUE

Gordon Miller suggested to replace the standard comparators with the following functions across {admiral}

{base}	improved
A >= B	A > B | near(A, B)
A <= B	A < B | near(A, B)
A == B	near(A, B)
A != B	!near(A, B)
A > B	A > B & !near(A, B)
A < B	A < B & !near(A, B)

This would work perfectly fine, but especially for case_when() statements, it would add a lot of code-bloat.

Although a minor issue, it looks like the near() function tests for absolute differences, while the all.equal() function tests for relative differences, as discussed in this thread:

# Very large values:
# When checking for absolute differences
near(
  ANRHI * 1.1 * 10^6,
  AVAL * 10^6
)

[1] FALSE

# When checking for relative differences
all.equal(
  ANRHI * 1.1 * 10^6,
  AVAL * 10^6
)

[1] TRUE

# As:
(ANRHI * 1.1 * 10^6) %>% format(digits = 22)

[1] "110000000.0000000149012"

(AVAL * 10^6) %>% format(digits = 22)

[1] "1.1e+08"

{base}	{fpCompare}
A >= B	A %>=% B
A <= B	A %<=% B
A == B	A %==% B
A != B	A %!=% B
A > B	A %>>% B
A < B	A %<<% B

As an example to how this is implemented, we can have a look at the {fpCompare} source code for one of the operators:

`%<=%` <- function(x, y) {
  (x < y + getOption("fpCompare.tolerance"))
}

Even if y is ever so slightly smaller than x, adding the tolerance to y will make the result larger than x, and the comparison will return TRUE.

# we need to set the fpCompare.tolerance first, because we did not load the package:
options(fpCompare.tolerance = 1e-8)

(ANRHI * 1.1) %<=% AVAL

[1] TRUE

As long as {admiral} remains open source and free to use, using this package, or even reusing the code itself would be fine. Although this was my preferred option, we did not end up implementing it. Instead, we made use of the signif() function, which rounds a number to a specified number of significant digits. This way, we could use the regular infix operators and simply provide the number of significant digits we want to compare to:

signif_dig <- 15

signif(AVAL, signif_dig) == signif(ANRHI * 1.1, signif_dig)

[1] TRUE

# as:
(ANRHI * 1.1) %>%
  signif(signif_dig) %>%
  format(digits = 22)

[1] "110"

# and although when printed, the number still looks off:
ANRHI <- 101
((ANRHI * 1.1) %>% signif(signif_dig)) %>% format(digits = 22)

[1] "111.0999999999999943157"

# the comparison works now:
((ANRHI * 1.1) %>% signif(signif_dig)) == 111.1

[1] TRUE

This is now implemented throughout atoxgr_criteria_daids, atoxgr_criteria_ctcv4, and atoxgr_criteria_ctcv5, and we are working on an issue for the 1.0.0 release of {admiral} to implement this for derive_var_anrind as well.

atoxgr_criteria_daids %>%
  select(TERM, GRADE_CRITERIA_CODE) %>%
  reactable(defaultPageSize = 4, highlight = TRUE, bordered = TRUE, striped = TRUE, resizable = TRUE)

Conclusion

The recent challenges faced by {admiral} in dealing with floating point values shed light on the complexities and nuances of working with these numerical representations. Floating point values, as we’ve seen, are approximations of real numbers and can lead to unexpected issues in mathematical operations, especially when using exact comparators like == and >=. The differences between how these values are stored and computed can result in platform-specific discrepancies and unexpected behavior.

Several potential solutions were explored to address this issue, including rounding, using near() or all.equal() functions, or implementing custom infix operators as seen in the fpCompare package. However, the most elegant and practical solution adopted in {admiral} was to use the signif() function to round values to a specified number of significant digits. This approach allows for reliable and consistent comparisons without adding unnecessary complexity to the code base.

Readers and developers should be vigilant when working with floating point values in their own code or when utilizing {admiral} for their projects. Keep in mind that some floating point values can look like integers at first glance as in the above example of 1.1*100. The experience with floating point issues in {admiral} serves as a valuable reminder of the potential pitfalls associated with numerical precision in programming. It’s crucial to exercise caution when performing comparisons with floating point numbers as small discrepancies can have significant downstream implications. When writing your own comparisons consider the following best practices:

1. Avoid Exact Comparisons: As highlighted earlier, using exact comparators like == or >= when dealing with floating point values can lead to unexpected results. Instead, opt for methods that take into account a tolerance or margin of error, such as the near() function or the signif() approach discussed in this context.

2. Platform Independence: Be aware that floating point representations may differ across various platforms or environments. Always test your code on multiple platforms to ensure consistency in results.

3. Documentation and Comments: When writing code that potentially involves floating point comparisons, it’s advisable to include clear documentation and comments that explain the reasoning behind your approach. This will help others understand and maintain the code effectively.

4. Testing and Validation: Implement thorough testing and validation procedures to verify the correctness of your code, particularly when it relies on floating point comparisons. This should include specific tests that would flag floating point issues on any machine or platform.

By heeding these precautions and understanding the intricacies of floating point representations, you can mitigate the risk of encountering unexpected behavior in your code. Whether you’re working with {admiral} or any other software, a cautious and informed approach to handling floating point values is essential for maintaining code accuracy and reliability.

** This is a number of the smallest magnitude for which a difference is still detected. I.e. .Machine$double.eps / 1.8 is still detectable, while .Machine$double.eps / 2 is not detectable any longer (at least on my machine):

# eps / 1.8 is still detectable:
.Machine$double.eps / 1.8 + 1 == 1

[1] FALSE

.Machine$double.eps / 2 + 1 == 1

[1] TRUE

Last updated

2024-11-15 19:40:27.890212

Details

Source, Session info