When has discrete understanding preceded continuous? From my limited perspective, it appears that the understanding
of a mathematical phenomenon has usually been achieved,
historically, in a continuous setting
before it was fully explored in a discrete setting.
An example I have in mind is the relatively recent activity in
discrete differential geometry, and discrete minimal surfaces in particular:

      


      

Image:
A discrete Schwarz-P surface. Bobenko, Hoffmann, Springborn: 2004 arXiv abstract.


I would be interested in examples where the reverse has happened:
a topic was first substantively explored in a discrete setting,
and only later extended to a continuous setting.
One example—perhaps not the best example—is the contrast
between $n$-body dynamics and galactic dynamics. Of course the former
is hardly "discrete," but rather is more discrete than the study of
galaxy dynamics leading to the conclusion that there must exist vast
quantities of dark matter.
Perhaps there are clearer examples where discrete understanding preceded
continuous exploration?
 A: I would say that a lot of topology was discrete before it was continuous. 
The Euler characteristic was first observed (in 1752) as an invariant of
polyhedra. Around 1900 Poincaré first calculated Betti numbers, and 
generalized the Euler characteristic, in a polyhedral setting. The first
general treatment of topology, by Dehn and Heegaard in their Enzyklopädie
article of 1907, was also in a polyhedral setting.
It was only after 1910, with simplicial approximation methods introduced by
Brouwer, that the topological invariance of various combinatorial invariants
was proved. For example, Alexander proved the topological invariance of the
Betti numbers in 1915.
A: I can think of two good examples. The first is rather straight forward. The hyper-operators. Namely,
$$a \uparrow^n b : \mathbb{N}^3 \to \mathbb{N}$$
$$a \uparrow^0 b = a \cdot b$$
$$a \uparrow^n 1 = a$$
$$a \uparrow^{n} (a \uparrow^{n+1} b) = a \uparrow^{n+1} (b+1)$$
$$a \uparrow^{n+1} b = a \uparrow^n a \uparrow^n \cdots (b\text{ times}) \cdots \uparrow^n a$$
It is a rather interesting open problem to construct the same object in analysis. Namely, replace the $\mathbb{N}$'s with some domain in $\mathbb{C}$. Pivotally, tetration, or $e \uparrow^2 z$, is somewhat solved (though controversy exists as to which solution is the right solution). You can perceive the problem better with tetration. It's very easy to get
$$e \uparrow^2 k = e^{e^{\cdots k\text{ times}\cdots^e}}$$
which satisfies
$$e^{e \uparrow^2 k} = e \uparrow^2 (k+1)$$
but how do we get holomorphic $e \uparrow^2 z$? Surely in this scenario the discrete instance inspired the continuous instance.
Another one would be indefinite summation. Everyone knows that
$$\sum_{j=1}^n f(j) = a(n)$$
is well defined, and satisfies $a(n) + f(n+1) = a(n+1)$, but how do we get holomorphic
$$\sum_{j=1}^z f(j) = a(z)$$
where $a(z) + f(z+1) = a(z+1)$. Surely the discrete instance inspired the continuous instance, again.
In fact, the whole field of study dedicating itself to extending recursive relationships defined on the naturals to the complex plane fits the bill of discrete before continuous very well.
A: Investigations of harmonic motion and heat flow (by Huygens, the Bernoullis, Euler, Fourier, et al.) as discussed in Sections 1.2, 3, and 5 of "The acoustic origins of harmonic analysis" by Olivier Darrigol involved using discrete models as a stepping stone to the continuous.
(This research motivated much work devoted to establishing rigourous foundations of mathematical analysis, such as the definition of a function.)
A: To add to the theoretical computer science angle: the study of graph algorithms has traditionally been done in a combinatorial setting, because that is way more natural at first of course. But recently our understanding has been greatly aided by methods from linear algebra and continuous optimization, especially in the context of the central to the field maximum flow problem. These insights have in particular lead to the fastest known graph algorithms for certain problems. For some overview of this, see these recent slides by Madry.
A: The main line of inquiry in information theory concerns itself with theoretical limits on data compression for communication and storage, as well as theoretical limits on transmission rates for noise-resilient communication (or fault-tolerant computing, etc.). Claude Shannon's famous 1948 paper, which lays the foundation for the field rather comprehensively, concerns itself entirely with discrete information sources (that is discrete in both time and state space), modeled as stochastic processes. Later treatments (even until now) remain much more concerned with discrete information sources. As far as I know, there is substantial treatment of continuous sources (especially continuous state-space), but it was not pursued first and it remains secondary.
A: Fredholm integral equations of the first kind can be viewed as continuous analogues of matrix multiplication of a vector. The use of the terminology for matrices, such as the trace and determinant, for integral kernels are vestiges of the development from the discrete to the continuous as explained in "On the origin and early history of functional analysis" by Lindstrom (see in particular Section 4).
A: Quantum field theory is a very cogent example of contemporary interest: Path integrals and quantum fields on a lattice are easy to define (and lattice quantum chromodynamics is a big industry nowadays), but establishing even a good definition for the continuum analogs has proved hitherto elusive.  You can win a million dollars for proving that such a continuum field theory exists.
A: Probability and stochastic processes
A: Group theory might be a good example. The first examples of groups were in a discrete setting, namely Galois groups of number fields, which were first understood as permutation groups of the finite set of roots of a polynomial. The emergence of Lie's theory of continuous groups and Klein's Erlangen program came somewhat later (although, in writing this question, I discovered that it was closer than I expected).
A: 
In general you seem to be right. Since the invention of Calculus,
  "continuous" became easier, and usually is investigated before the
  discrete. One example of the opposite is the differential equations vs
  difference equations. As I understand the first difference equation
  studied was the one that defined Fibonacci sequence. But this was long
  before Calculus.

On the contrary, I would argue this is a wholly 'modern' 20th century perspective. Pretty much all the great mathematical-physicists that introduced major classes of new mathematical tools, from Newton to Maxwell to Feynman, were very much constructive mathematicians, in their thinking (Einstein comes to mind as a counterexample; sure there are more).
Newton pretty much hated infinitesimal calculus, and wrote all his original proof in terms of finite geometrical constructions; that the differential algebraic method was a more convenient method of obtaining the same results is definitely the opinion of people that came after him, not of the guy that came up with the original logic of it.
Maxwell preferred to think about vector calculus in terms of finite elements / mechanical gears; and if I had to teach someone electrodynamics today, I would most certainly prefer to do it using discrete exterior calculus rather than using vector calculus (and similarly, einstein notwithstanding, I think Regge calculus is a lot more instructive than differential tensor equations).
The same applies to Feynman. His methods were developed using discrete logic; and he was highly critical of the logical acrobatics required to make sense of the UV divergences as you take the limit to continuous space.
A: Surely summation of finite discrete series was well-understood, conceptually, long before integration. And I would not be surprised to see similarly for many other ideas of the differential/integral calculus (though not all…).
A: Pretty much every Analytic Continuation from the integer, or other discrete, into real (or other continuous) domain is an example of this.
The Gamma function extending the factorial into real domain is a classic example of analytic continuation, but there are many more.
$\Gamma (n) = (n-1)!$
${\displaystyle \Gamma (z)=\int _{0}^{\infty }x^{z-1}e^{-x}\,dx}  $
A: Benjamini-Schramm convergence was originally defined for graphs, later extended to metric spaces with applications especially in the setting of locally symmetric spaces. See http://annals.math.princeton.edu/2017/185-3/p01
A: The Riemann zeta function is a prime example.
Far into antiquity are investigations of partial sums of integral powers of the natural numbers, leading eventually to the Bernoulli polynomials and the discovery of their exponential generating function by Euler. It was natural to then look at partial sums of the reciprocals of the natural integers.  
From Wikipedia:
The Basel problem is a problem in mathematical analysis with relevance to number theory, first posed by Pietro Mengoli in 1650 and solved by Leonhard Euler in 1734, and read on 5 December 1735 in The Saint Petersburg Academy of Sciences. ...
The Basel problem asks for the precise summation of the reciprocals of the squares of the natural numbers, i.e. the precise sum of the infinite series:
$$ \sum_{n>0} \frac{1}{n^2}$$
Excerpt from "On some historical aspects of the theory of the Riemann zeta function" by Giuseppe Iurato:
Euler's application of infinite series to different number theoretical problems was of principal importance. The study
of the series $ \frac{1}{n^{2k}}$ leads to series of the form $ \frac{1}{n^s}$. ....
Euler (1707-1783) was probably the first to see that these series can be applied to number theory. He was in correspondence with C. Goldbach and J.L Lagrange just on number theory questions. His proof of the existence of infinitely many primes uses the divergence of the harmonic series $\sum \frac{1}{n}$ and using the above fundamental theorem of arithmetic which says that every natural number can uniquely be written as a product of powers of primes. Afterwards, P.L.G. Dirichlet (1805-1859) systematically introduced analytical methods in number theory. Among other things, he investigated the series $ \frac{1}{n^s}$ for real $s$, while B. Riemann (1826-1866) allowed complex $s$. 
There is also Euler's formula for reflection symmetry of the nascent Riemann $\xi$ function. 
It is precisely the e.g.f. for the Bernoulli numbers that appears in Riemann's integral formula for $(s-1)! \zeta(s)$, as he was well aware of.
A: Integers to rationals to reals.
A: Theoretical computer science offers lots of examples of what you want, since people almost always think first about the discrete setting, even if the continuous setting also turns out to be important.  Two examples off the top of my head:
Grover's search algorithm, one of the central quantum computing algorithms, was first understood in the discrete-time setting and only later in the continuous-time one.  (In an amusing turnabout, the Farhi-Goldstone-Gutmann algorithm, which generalizes Grover's algorithm to games of alternation like chess and Go, was first understood in a continuous-time setting and only later in a discrete-time one.)
The number of samples needed to learn a hypothesis was first understood in the discrete case (through the concept of VC-dimension), and only later generalized to the continuous case (through the concept of fat-shattering dimension).
A: From SKETCHES OF KDV by E. Arbarello:
Studying a dynamical system consisting of a finite number of particles of unitary mass distributed along a line segment with forces acting on adjacent pairs, Fermi, Pasta and Ulam introduced a natural non-linear perturbation which gave rise to an unexpected, discrete version of the KdV equation. Starting from this model, ten years later, Zabusky and Kruskal [ZK65] undertook a numerical study of the KdV which exhibited, for the first time, solutions of KdV having the appearence of packets of N localized waves which collide and resurrect in the same way as Scott Russel’s waves do in a shallow canal. Because of the particle-like behaviour of these collisions, they named these solutions: N-solitons.
A: A recent example would be the generalization of graphs to continuous objects graphons, symmetric measurable functions on a square. Topological properties of spaces of these objects (e.g., compactness) have been shown to yield known properties in the discrete setting (existence of Szemeredi partitions).  Also the continuous setting allows graphons to be interpreted as probability distributions which form a very general model for random graphs.  Such things are detailed in the book of Lovasz, "Large networks and graph limits".
A: Discrete (more specifically, finite field version) Kakeya conjecture is solved by Zeev Dvir using polynomial method, while original continuous problem is wide open. 
If you let me speculate on the reasons, I would say that though we have some ideas what are 'continuous polynomials' (say, Fourier integrals are analogues of trigonometric polynomials), we do not understand what is a right substitute for degree of a polynomial. 
A: Another example is coin tossing and other probabilistic models. Coin tossing was studied long before Brownian motion. Discrete probability precedes continuous probability.
A: In general you seem to be right. Since the invention of Calculus, "continuous" became easier, and usually is investigated before the discrete. One example of the opposite is the differential equations vs difference equations. As I understand the first difference equation studied was the one that defined Fibonacci sequence.
But this was long before Calculus.
EDIT. As I recently learned, Fibonacci discussed this sequence but he did not derive the formula for the general term, did not "solve" the difference equation. It was solved by D. Bernoulli AFTER the invention of calculus. The solution is called Binet formula. 
A: One of Ramanujan's favorite insights on how the discrete informs the continuous is described in my reply to the MO-Q "Ramanujan's Master Formula: A proof and relation to umbral calculus."
Surmising the Mellin transform of $f(t)$ from the coefficients of the Taylor series expansion about the origin of $f(-t)$ is a beautiful heuristic, which I've yet to see incorporated into tables of Mellin transforms. (Hardy gave a formal presentation of this heuristic after noting R's use of it.)
