I will answer your question in what follows, but first it might be useful to say something about what derived functors are and what the point is.
The idea of derived functors is that if you have a functor which preserves (say)
left exactness when applied to a short exact sequence, but not right exactness
(e.g. $Hom(X, \text{--})$ for a module $X$ ), then its derived functors
do produce an exact sequence, but now a long exact sequence.
As you probably know, the long exact sequence is
$$0 \to F(A) \to F(B) \to F(C) \to R^1F(A) \to R^1F(B) \to R^1(C) \to R^2(A)
\to \cdots$$
Here I am writing $R^iF$ to denote the $i$th derived functor (a right derived
functor in this case, since the functor $F$ is left exact; so the letter $R$
is for "right").
Note that the derived functors are a so-called $\delta$-functor :
not only do we have the sequence of functors $R^iF$, but the above
long e.s.'s are themselves functorial in the s.e.s. $0 \to A \to B \to C \to 0.$
(The "$\delta$" in $\delta$-functor is a reference to the collection of connecting morphisms
$R^iF(C) \to R^{i+1}F(A)$, which are commonly denoted $\delta$.)
There are a few comments to make:
- The existence of these l.e.s.'s doesn't characterize the $R^iF$ as a $\delta$-functor.
E.g. if you had the collection $R^iF,$ I could define another functor
$\widetilde{R}^iF$ as follows: in all degrees $i$ except $i = 1$,
set $\widetilde{R}^iF =R^iF,$ but in degree $1$, define
$\widetilde{R}^iF(A) = R^iF(A) \oplus A$ (for any object $A$).
Define the connecting morphisms on the extra summand we put in to be just $0$.
So the new l.e.s. is the same as the old one, except that we have just direct
summed in a copy of the original s.e.s. into the degree $1$ part of the l.e.s.
of $R^iF$'s.
This is dealt with by the following point.
- The $R^iF$ are taken to be minimal, in some appropriate sense. Precisely,
they are a universal $\delta$-functor for $F$, i.e. they map canonically
to any other $\delta$-functor. (This is discussed in Hartshorne's algebraic
geometry book at the beginning of Ch. III, but in many other places too,
I presume.)
So, intuitively, they don't contain any extra garbage that we might have
added gratuitously in some degree.
- Although it's not obvious if you're learning this for the first time,
to construct the universal $\delta$-functor, you can use injective
resolutions of objects, and apply $F$ (the original, underived, functor)
to these resolutions. (This gives a cochain complex, whose cohomology
yields the $R^iF$'s.) To get the l.e.s.'s: first prove
that a s.e.s. of objects gives a s.e.s. of resolutions.
Then, since s.e.s.'s of injective objects are split, and since $F$ will take a split
s.e.s. to a (split) s.e.s., applying $F$ gives a s.e.s. of complexes.
Passing to cohomology, and using the fact recalled in your question,
gives the l.e.s. of derived functors.
So the connection between the fact you recalled in your question,
and derived functors, actually doesn't come until a fairly technical
part of the story, namely in the actual construction of the derived
functors via injective resolutions.
Note that although sometimes you use resolutions to compute the derived
functors, often you use more indirect means, and so the construction via
injective resolutions is often not a key point to focus on in applications
(even though it underlies the theory).
- What is the point of the whole story? Why do we go to all the trouble of constructing these derived functors, looking at the l.e.s.'s, and so on?
E.g. why don't people just do the naive thing of looking at the four term
exact sequence
$$0 \to F(A) \to F(B) \to F(C) \to \text{ cokernel } \to 0.$$
The point is that this cokernel depends not just on $A$, $B$, or $C$,
but on all three of them, and in fact on the whole data of the s.e.s.
$0 \to A \to B \to C \to 0.$
So basically, it's not a very flexible object, and there isn't a lot
of theory you can develop about it directly.
What happens in the theory of derived functors is that the information in
this cokernel is diffused into the theory in a much more subtle, flexible, and
useful way.
E.g. in the l.e.s. of derived functors, the terms only depend on one of
the objects $A$, $B$ ,or $C$. It is only the morphisms (especially the
connecting morphisms) that depend on the whole data of the s.e.s.
In particular, suppose in some particular case we want to show that
$ 0 \to F(A) \to F(B) \to F(C) \to 0$ is exact.
The general theory shows that what we have to do is prove that
the connecting morphism $F(C) \to R^1F(A)$ equals zero.
One way we can do this is just to show that $R^1F(A) = 0$ itself. This
depends only on $A$, and so doesn't depend on the original s.e.s. at all.
In particular, it is something we could try to check by making a computation
related to $A$, without having to think at all about $B$ or $C$ or the original
s.e.s. Possibly we could compute directly with injective resolutions, or perhaps we could try to use some other more indirect method: e.g. we could
try to put $A$ into some other s.e.s. we know more about, and use the derived
functor l.e.s. for that s.e.s. to get information about the values of the
$R^iF(A)$.
Even if we can't prove that $R^1F(A)$ vanishes, we can still hope to get
information (e.g. if we can bound its size, then we can bound the size of the
cokernel of $F(B) \to F(C)$). By diffusing the information through
the various derived functor values and the connecting morphisms, the theory of
derived functors provides a much more useful tool then you get from a more naive
approach to these questions.