select regexp_matches('a a a', '([a-z]) a','g');

First: Please don't reply to an existing message to create a new thread.
Your mail client copies the replied-to message ID into the References:
header, and well-implemented mail clients will thread your message under
a now-unrelated thread.

Compose a new message instead.

Marc Mamin wrote:

> I have a string that contains a serie of chars, separated by single
> spaces.
>
> e.g 'a b x n r a b c b'
>
> Having such a string, I d'like to get a list of all predecessors of a
> given character.
> In the example, the predecessors of b are a,a,c.

OK, so wherever `b' occurs, you want the character at index `b -2'.

> select regexp_matches('a a a', '([a-z]) a','g');
> => {"a "} only

The issue is that regular expressions don't like to overlap matches. The
first match consumes _two_ leading `a' characters.

What you need is a zero-width lookahead assertion, available in
Perl-style extended regular expressions. Handily, recent PostgreSQL
versions support these, so you can write:

test=> select regexp_matches( 'a a a', '([a-z]) (?=a)', 'g');
regexp_matches
----------------
{a}
{a}
(2 rows)

--
Craig Ringer