I suppose this is a terminology question. "A digital signature scheme which can sign many documents with one private key" means something like this:
- There are some sets $M$ (the "message space", often the set of bitstrings of any length, or some useful subset thereof), $K_{pub}$, $K_{priv}$ (the public and private "key spaces") and $S$ (the signature space).
- There is a tuple of (probabilistic and efficiently computable) functions $(g, s, v)$.
- $g : \varnothing \to K_{pub} \times K_{priv}$ ($g$ generates a random key pair).
- $s : M \times K_{priv} \to S$ ($s$ signs a message, i.e. generates a signature from the private key and message)
- $v : M \times K_{pub} \times S \to \{\mathsf{true}, \mathsf{false}\}$ ($v$ verifies a signature)
- If $(y, x) = g()$ and $m \in M$, then $v(m, y, s(m, x)) = \mathsf{true}$ (or at least with overwhelming probability).
Normally we also want that the scheme is secure, which then adds some more properties:
Given $y$ (the public key) and a message $m$, without knowledge of $x$ it is hard to create a $\sigma \in S$ such that $v(m, y, \sigma) = \mathsf{true}$, even if some (or even many) other $(m', \sigma')$ are known, or even signatures to other messages can be requested. (This is known as "universal forgery".)
It should even be hard to create any message $m \in M$ and $\sigma \in S$ such that $v(m, y, \sigma) = \mathsf{true}$ (other than the $m'$s which were given/queried as examples). (This is known as "existential forgery".)
(There are also variants of this definition where the number of documents which can be signed with each key before the security gets lost is limited, but these obviously are not "signature schemes which can sign many documents with one key".)
Usually there also is some formalization of "hard", often with a security parameter.
Similarly, we have the term of "hash function", something like that:
- There is a message space $M$ (usually the set of bitstrings of any length, or some useful subset thereof) and a hash space $H$ (usually the set of bitstrings of some fixed length).
- There is a function $h : M \to H$.
The "collision resistance" property then is something like
- It is hard to find $m_1, m_2 \in M$ with $h(m_1) = h(m_2)$.
The "there exists a digital signature scheme which ..." in the quoted lecture extract supposedly means as much as "there exists a secure signature scheme which ...".
We already know that there are signature schemes (RSA, DSA, ECDSA, ...), but it is not proven if they are actually secure. (We sure hope so, and nobody did break them publicly yet, but it might not actually be possible to prove anything, or it might be that there is no secure signature scheme at all.)
We also don't know if there exist any actually collision-resistant hash function. We have some candidates (like SHA-2), and some previous candidates which proved to actually be not collision resistant (MD4, MD5).
The argument then is:
If there is any secure signature scheme (for multiple documents with each key), then there has to be a collision-resistant hash function.
(The idea is that you actually can use the signature scheme as a hash function by using a fixed key, and the signature scheme is not secure if you can find two messages with the same signature (i.e. a hash collision). I think this doesn't actually follow from my security properties mentioned above, so maybe your lecture used some other security definition.)
Also there will be shown (in the following paragraphs/pages of your text) that from any collision-resistant hash function you can build a secure hash-based signature scheme.
So it follows: If there is any secure signature scheme (which can sign many messages) at all, then there is a secure hash based signature scheme.