If you're interested in learning more about CPU hardware details and how they affect program performance, read my blog post here on how memory latency impacts binary search and causes an $$$O(\sqrt[3]{n})$$$ runtime.

Enjoy!

I recently wrote an article about Schonhage-Strassen on my personal blog here. It is an algorithm to multiply two N bit integers in time. If you want additional background on the signal processing stuff mentioned in the article, read the first part of the post here.

I doubt this will be useful for competitive programming because Schonhage-Strassen has a pretty high constant factor and is pretty tedious to implement--in general floating point FFT-based solutions are fine for competitive programming size inputs and NTT-based stuff like this is overkill. Hopefully some of you will still find it interesting though!

Here's a technique to do C — New Year and Curling in time with segment tree. Brute force n² is of course fast enough, but I think speeding it up was interesting.

Suppose we are about to place disk i at x_i. The relevant x range we should look for disks is [x_i - 2r, x_i + 2r] because things outside of this range are too far away to intersect with this disc. Say we find j such that j < i, x_j is in the needed range, and y_j is maximized. We can find such j with a basic segment tree in time. The intuition behind the approach is that is a good lower bound for y_i. Specifically, y_i < y_i' + 2r as shown by the diagram:

In this worst case scenario we find that disk j corresponds to the left disk. However, another disk is centered at (x_i, y_j - ε). So y_i' in this case is y_j but the true answer is y_j + 2r - ε. If we could find the second-highest circle with our segment tree, we'd get the answer right in this case. We can do this by first querying [x_i - 2r, x_i + 2r] and getting back a result x_q1. Assume x_q1 < x_i (the other case is symmetric); we next query (x_q1, x_i + 2r]. Say this second query result is x_q2 with x_q2 > x_i (again other case is symmetric); we next query (x_q, x_q2) to get the third-highest circle. If we do this just a few times we're guaranteed to get the right answer.

The reasoning behind this is that a constant number of circles fit in the "relevant area," regardless of what n or r are. The relevant area is a 4r x 2r rectangle horizontally centered at x_i and with the top y value y_q1. It's the black box in the diagram above. If a disk's center is horizontally outside this range, the disk won't intersect the query disk. If the disk is below the relevant area, it's too low to matter (y_i' is higher). And no disk can be above this area because of how y_q1 is defined. The relevant box has area 8r², and any circle centered in it must occupy at least π r² / 4 area, so at most 10 circles inside of the box. So, if we do our procedure 10 times (so 10 segment tree queries), we're guaranteed to get the answer. 10 is a pretty loose bound and I got AC with 4.

See AC code here