Character class subtraction is supported by the XML Schema, XPath, .NET (version 2.0 and later), and JGsoft regex flavors. It makes it easy to match any single character present in one list (the character class), but not present in another list (the subtracted class). The syntax for this is [class-[subtract]]. If the character after a hyphen is an opening bracket then these flavors interpret the hyphen as the subtraction operator rather than the range operator. You can use the full character class syntax within the subtracted character class.

ICU supports character class subtraction using a syntax of its own. ICU uses [class--[subtract]] or [class--subtract] with double hyphens. You only need the extra pair of square brackets if the subtracted class needs to be negated. The examples below all work with ICU if you add the extra hyphen.

The character class [a-z-[aeiuo]] matches a single letter that is not a vowel. In other words: it matches a single consonant. Without character class subtraction or intersection, the only way to do this is to list all consonants: [b-df-hj-np-tv-z].

The character class [\p{Nd}-[^\p{IsThai}]] matches any single Thai digit. The base class matches any Unicode digit. All non-Thai characters are subtracted from that class. [\p{Nd}-[\P{IsThai}]] does the same. [\p{IsThai}-[^\p{Nd}]] and [\p{IsThai}-[\P{Nd}]] also match a single Thai digit by subtracting all non-digits from the Thai characters.

Nested Character Class Subtraction

Since you can use the full character class syntax within the subtracted character class, you can subtract a class from the class being subtracted. [0-9-[0-6-[0-3]]] first subtracts 0-3 from 0-6, yielding [0-9-[4-6]], or [0-37-9], which matches any character in the string 0123789.

The subtraction works on the whole class. [\p{Ll}\p{Lu}-[\p{IsBasicLatin}]] matches all uppercase and lowercase Unicode letters, except any ASCII letters. The \p{IsBasicLatin} is subtracted from the combination of \p{Ll}\p{Lu} rather than from \p{Lu} alone. This regex will not match abc.

The class subtraction must always be the last element in the character class. [0-9-[4-6]a-f] is not a valid regular expression. It should be rewritten as [0-9a-f-[4-6]] if you meant to match a hexadecimal digit that is not between 4 and 6. ICU is different because it ignores the nested pair of brackets. Both [0-9--[4-6]a-f] and [0-9--[4-6a-f]] are equivalent to [0-9--4-6a-f] in ICU. You might just as well write this as [0-9--4-6] because subtracting a-f from 0-9 has no effect.

While you can use nested character class subtraction, you cannot subtract two classes sequentially. To subtract ASCII characters and Greek characters from a class with all Unicode letters, combine the ASCII and Greek characters into one class, and subtract that, as in [\p{L}-[\p{IsBasicLatin}\p{IsGreek}]]. For ICU you could write this as [\p{L}--[\p{InBasicLatin}\p{Greek}]] or [\p{L}--\p{InBasicLatin}\p{Greek}]. [\p{L}--[\p{InBasicLatin}]\p{Greek}] would also do the same but may be confusing to another developer reading your regex.

Subtraction in Negated Classes

The character class [^1234-[3456]] is both negated and subtracted from. In all flavors that support character class subtraction, except ICU, the base class is negated before it is subtracted from. This class should be read as "(not 1234) minus 3456". Thus this character class matches any character other than the digits 1, 2, 3, 4, 5, and 6. More practically, [^\p{IsThai}-[\p{Nd}]] matches any non-Thai character minus all digits. It matches A but not 1. [^\p{IsThai}-[^\p{Nd}]] matches any non-Thai character minus all non-digits. It matches 1 but not A. Neither of these regexes match the Thai letter ก or the Thai digit ๑.

ICU applies negation before subtraction. It reads [^1234-[3456]] as "not (1234 minus 3456)". Thus with ICU this character class matches any character other than the digits 1 or 2. [^\p{Thai}-[\p{Nd}]] matches any character that is not in the set of Thai characters excluding digits. It does not match the Thai letter ก but does match the Thai digit ๑. [^\p{Thai}-[^\p{Nd}]] matches any character that is not in the set of Thai characters excluding non-digits. Both regexes match the non-Thai characters A and 1 in ICU.

So be careful when combining negation and subtraction. The difference in priority between regex flavors is easily overlooked yet results in dramatically different matches from the same regexes. If this all seems quite confusing then it’s probably better not to subtract from a negated class at all even if you only work with a single regex flavor.

Notational Compatibility with Other Regex Flavors

Note that a regex like [a-z-[aeiuo]] does not cause any errors in most regex flavors that do not support character class subtraction. But it won’t match what you intended either. In most flavors, this regex consists of a character class followed by a literal ]. The character class matches a character that is either in the range a-z, or a hyphen, or an opening bracket, or a vowel. Since the a-z range and the vowels are redundant, you could write this regex as [a-z-[]] or [-[a-z]] in Perl or PCRE. A hyphen after a range is treated as a literal character, just like a hyphen immediately after the opening bracket. This is true in the XML, .NET, ICU and JGsoft flavors too. [a-z-_] matches a lowercase letter, a hyphen or an underscore in these flavors.

Strictly speaking, this means that the character class subtraction syntax is incompatible with Perl and the majority of other regex flavors. But in practice there’s no difference. Using non-alphanumeric characters in character class ranges is very bad practice because it relies on the order of characters in the ASCII character table. That makes the regular expression hard to understand for the programmer who inherits your work. While [A-[] would match any uppercase letter or an opening square bracket in Perl or PCRE, this regex is much clearer when written as [A-Z[]. The former regex would cause an error with the XML, .NET and JGsoft flavors, because they interpret -[] as an empty subtracted class, leaving an unbalanced [.