Ttyjfyk

Why does sed, if we use the negation expression[^ ] treat anchor characters like b or B as real characters? E. g. one would expect the following expressions to yield the same result, but they don't:

$ echo 'apple pear melon banana cherry papaya' | sed 's/[^b]a[^b]/u/g'
apple pu melon baua cherry uaya
$ echo 'apple pear melon banana cherry papaya' | sed 's/BaB/u/g'
apple peur melon bununa cherry pupuya

If there was no B, how could we negate b?

Neither of b or B is a character. Both are zero-width patterns that matches between characters.

The b pattern matches at a word boundary, i.e. between a character that is a "word character" and a character that is not a "word character".

The B pattern matches at a non-word boundary, i.e. between a two characters that are both either "word characters" or not.

The pattern [^b] matches one character. This is why pear is transformed into pu, you replace ear (the a and the surrounding characters).

With GNU sed, [^b] matches a character that is not an or a b.

There is no way to use a character class to replace the use of B that I'm aware of.

The b and B patterns are supported by GNU sed. Both GNU sed and BSD sed also has < and > for explicitly matching at the start and end of a word, and BSD sed additionally supports the POSIX patterns [[:<:]] and [[:>:]] (but GNU sed does not). The POSIX patterns can't be negated ([^[:>:]] does not work).

To get a similar effect without using B, you could use something like

$ echo 'apple pear melon banana cherry papaya' | sed 's/([[:alnum:]])a([[:alnum:]])/1u2/g'
apple peur melon bunana cherry pupaya

That is, match an alphanumeric character on either side of the a, and then include these two flanking characters in the replacement. Note that since the replacement only happens for non-overlapping matches, this would not properly substitute the a's in a string containing multiple consecutive a's (or a's in every second position). See how banana that does not come out as bununa due to this.

To sort that out, you could introduce a loop in the sed program:

sed -e :top -e 's/([[:alnum:]])a([[:alnum:]])/1u2/g' -e ttop

This performs the replacement over the input line as many times as needed until all overlapping pattern matches have been handled.