Why is OCaml's pattern matching weaker than Erlang's?

First, you can always capture equality between pattern variables via a when clause in OCaml, e.g.:

let rec destutter = function
    | []              -> []
    | [hd]            -> [hd]
    | hd :: hd' :: tl
      when hd = hd'   -> destutter (hd :: tl)
    | hd :: hd' :: tl -> hd :: destutter (hd' :: tl)

There is a trade-off here. While Erlang is more expressive, OCaml pattern matching is simpler (which means a simpler language definition, compiler, etc.), and you still can do what you need (at the cost of writing more code).

Note that while you can rewrite a non-linear pattern as a linear pattern with a when clause, that is not the primary issue. More critically, pattern matching then needs to have a notion of equality for arbitrary types in order to support non-linear patterns. This is not an issue in Erlang, but OCaml not only already has = vs == built-in (structural equality vs. identity), but for any given type, it may not be the kind of equality that you need (think strings and case-sensitivity, for example). Then, as a consequence, checking for exhaustiveness or overlap becomes non-trivial. In the end, it's questionable whether providing a special case for one specific type of equality is worth it, given how many useful relations between parts of a pattern there can be. (I note that non-strict languages have additional issues.)

As an aside, Prolog's pattern matching is unification-based and strictly more powerful than either Erlang or OCaml (but also more difficult to implement).

Patterns in OCaml are compiled into a very efficient code with lots of sophisticated optimizations. Bjarne Stroustrup even bragged that they managed to write something comparable in certain cases in C++. But in general OCaml pattern matching is much faster. And it's charming to look at the assembly output. And maybe Erlang provides more flexibility, but it is expected from a dynamic language. Otherwise, why use them at all.

There is another issue also. Patterns are matched structurally. If you want to match [H,H|T] you are actually invoking the comparison of the first two elements. In most cases, the comparison function should be provided by a user, and it is not known in advance.

The Erlang pattern is more powerful because it can match against something determined at run time. The OCaml patterns match against things fixed at compile time. It follows that the OCaml patterns can probably be made to go faster. I also find OCaml-style patterns easier to reason about.