The origin of this type is discussed in the above book by Burrow, but I'd explain it in Latin terms.
Many Proto-Indo-European nominals showed vowel alternations in the declension, in particular the nominative singular often possessed a lengthened grade: a pattern that survives (although modified) in the Latin pēs — pedis
With time, this alternation tended to disappear, with either form generalized across the paradigm: for example rēx — *reges
(from Proto-Indo-European *Hrēgʲs — *Hregʲ-
) was replaced with rēx — rēgis
. The relational adjective from this noun inherited the long vowel: rēgius
From pēs — pedis
, however, the attested adjective is -pedius
(e. g. acūpedius
Imagine that both variants of relational adjectives once existed, that is *regios/*pedios
. If by some chance the noun generalized the full grade (**rex, **pes
) while the adjectives still survived in both variants, we would get the situation described by Burrow for earlier Indic, where both full- and lengthened grade- derivatives are sometimes attested. At the next stage, the pattern **rex — rēgius
started to expand, which has resulted in the system so characteristic of Sanskrit.