Fitting XP by CR
Due to the non-uniform level pacing detailed here and in screamline's answer, it's more productive to fit a curve to XP rewards by CR (since this doesn't depend on the current level of the party and is therefore invariant to level pacing) rather than than XP needed per level. It turns out that a good fit using round numbers is
$$\text{XP} \approx 50 \left(\text{CR} + 1\right)^2$$
Here's a semilog plot:
This is then modulated by the desired level pacing to get the XP requirements per level.
(However, it seems not as simple as the actual values being rounded from this equation. For example, it would be strange to round from 3200 to 2900 for CR 7. Also, the actual XP curve suddenly starts increasing rapidly above CR 20; it's best to consider that a separate regime entirely.)
Why a quadratic?
I do not know of any direct designer statements that state that they intentionally chose a quadratic, let alone why they did so if they did. However, here's a possible explanation:
5e adopted the doctrine of bounded accuracy, where, according to designer Rodney Thompson:
The basic premise behind the bounded accuracy system is simple: we make no assumptions on the DM's side of the game that the player's attack and spell accuracy, or their defenses, increase as a result of gaining levels. Instead, we represent the difference in characters of various levels primarily through their hit points, the amount of damage they deal, and the various new abilities they have gained. Characters can fight tougher monsters not because they can finally hit them, but because their damage is sufficient to take a significant chunk out of the monster's hit points; likewise, the character can now stand up to a few hits from that monster without being killed easily, thanks to the character's increased hit points.
If damage and hit points both increase linearly with CR, their product increases quadratically. Granted, this does not take into account increases in attack bonus or AC, but if they play a lesser role then a quadratic may be a good-enough approximation.
In fact, under a Lanchester model of combat with a few additional assumptions, the estimated party resources consumed by an encounter is proportional to this product. It's a natural choice to award XP proportionally.