(Analysis by Benjamin Qi)

Let $f(x,y)=x^2+y^2$ be the function we want to maximize. It turns out that for any convex function $f$, we can apply the same solution.

Claim: We only need to consider evaluating $f$ at the vertices of the convex hull of the set of all possible Minkowski sums.

Proof: Any point $p$ within the convex hull that is not a vertex can be written as $p=\lambda_1v_1+\lambda_2v_2+\cdots+\lambda_kv_k$ for some $\lambda_i>0, \sum \lambda_i=1$, $k>1$, and $v_1,\ldots,v_k$ distinct vertices of the convex hull. Then assuming $f$ is convex,

$$f(p)\le f\left(\sum \lambda_iv_i\right)\le \sum \lambda_i f(v_i)\implies \max_i f(v_i)\ge f(p).$$

It remains to describe how to efficiently compute the convex hull of the Minkowski sum.

For a slow solution, we may directly compute the convex hull of the Minkowski sum of two sets of sizes $|A|$ and $|B|$, by summing every pair of points and then computing the convex hull in $\mathcal O(|A|\cdot |B|\log (|A|\cdot |B|))$ time. It turns out the convex hull of the Minkowski sum always has at most $|A|+|B|$ vertices, leading to an $\mathcal O(T^2\log T)$ time solution if we apply this reasoning for every group in the input, where $T$ is the total number of vectors.

My code:

#include <bits/stdc++.h>
using namespace std;

using P = pair<int, int>;
#define f first
#define s second

using ll = long long;

P operator+(P a, P b) { return {a.f + b.f, a.s + b.s}; }
P operator-(P a, P b) { return {a.f - b.f, a.s - b.s}; }
ll cross(P a, P b) { return (ll)a.f * b.s - (ll)a.s * b.f; }
ll cross(P a, P b, P c) { return cross(b - a, c - a); }
ll sq(ll x) { return x * x; }
ll norm(P p) { return sq(p.f) + sq(p.s); }

// Graham Scan, assumes the hull has size > 1
vector<P> convex_hull(vector<P> v) {
nth_element(begin(v), begin(v), end(v));
const P leftmost = v;
for (P &p : v) p = p - leftmost;
sort(begin(v), end(v), [](P a, P b) { // sort points by argument
if (cross(a, b) == 0) return norm(a) < norm(b);
return cross(a, b) > 0;
});
vector<P> hull;
for (P p : v) {
while (hull.size() >= 2 && cross(end(hull)[-2], end(hull)[-1], p) <= 0)
hull.pop_back();
hull.push_back(p);
}
for (P &p : hull) p = p + leftmost;
return hull;
}

int main() {
ios::sync_with_stdio(false);
cin.tie(nullptr);
int N;
cin >> N;
vector<P> prefix_hull{{}};
while (N--) {
int G;
cin >> G;
vector<P> group(G);
for (P &p : group) cin >> p.f >> p.s;
vector<P> next_prefix;
for (P a : prefix_hull)
for (P b : group) next_prefix.push_back(a + b);
prefix_hull = convex_hull(next_prefix);
}
ll ans = 0;
for (P p : prefix_hull) ans = max(ans, norm(p));
cout << ans << "\n";
}


For a more efficient solution, let's start by computing the convex hull of each group. Then consider the following question: How can we more efficiently compute the convex hull of the Minkowski sum of two convex polygons $A$ and $B$ (and why is it true that it has size at most $|A|+|B|$)?

Define the offset sequence of a convex polygon to be the sequence of vectors traversed when touring the hull starting at the lexicographically least vertex of the hull and going in counterclockwise order. It turns out that the offset sequence of the Minkowski sum is always a permutation of the combined offset sequences of the summands. Consider the following example:

A = [(0, 0), (2, -1), (3, 0), (2, 2)]
offsets_A = [(2, -1), (1, 1), (-1, 2), (-2, -2)]

B = [(0, 0), (4, -3), (0, 5)]
offsets_B = [(4, -3), (-4, 8), (0, -5)]

sum(A,B) = C = [(0, 0), (4, -3), (6, -4), (7, -3), (6, -1), (2, 7), (0, 5)]
offsets_C = [(4, -3), (2, -1), (1, 1), (-1, 2), (-4, 8), (-2, -2), (0, -5)]


Observe that $\texttt{offsets}_C$ may be computed by merging the sequences $\texttt{offsets}_A$ and $\texttt{offsets}_B$ such that the final sequence is sorted in order of argument (as described here). For a more complete explanation of Minkowski sums, you can check the editorial for this problem.

If we want to compute the offsets of the Minkowski sum of more than two convex hulls, we can place all of the offsets into a single sequence and sort the entire sequence by argument. The overall time complexity is $\mathcal O(T\log T)$.

My code:

#include <bits/stdc++.h>
using namespace std;

using P = pair<int, int>;
#define f first
#define s second

using ll = long long;

P operator+(P a, P b) { return {a.f + b.f, a.s + b.s}; }
P operator-(P a, P b) { return {a.f - b.f, a.s - b.s}; }
ll cross(P a, P b) { return (ll)a.f * b.s - (ll)a.s * b.f; }
ll cross(P a, P b, P c) { return cross(b - a, c - a); }
ll sq(ll x) { return x * x; }
ll norm(P p) { return sq(p.f) + sq(p.s); }
int half(P p) {
if (p.f > 0 || (p.f == 0 && p.s > 0)) return 0;
return 1;
}

// Graham Scan, assumes the hull has size > 1
vector<P> convex_hull(vector<P> v) {
nth_element(begin(v), begin(v), end(v));
const P leftmost = v;
for (P &p : v) p = p - leftmost;
sort(begin(v), end(v), [](P a, P b) { // sort points by argument
if (cross(a, b) == 0) return norm(a) < norm(b);
return cross(a, b) > 0;
});
vector<P> hull;
for (P p : v) {
while (hull.size() >= 2 && cross(end(hull)[-2], end(hull)[-1], p) <= 0)
hull.pop_back();
hull.push_back(p);
}
for (P &p : hull) p = p + leftmost;
return hull;
}

int main() {
ios::sync_with_stdio(false);
cin.tie(nullptr);
int N;
cin >> N;
P start{}; // lexicographically minimum reachable point
vector<P> offsets;
while (N--) {
int G;
cin >> G;
vector<P> group(G);
for (P &p : group) cin >> p.f >> p.s;
vector<P> hull = convex_hull(group);
start = start + hull;
for (int i = 1; i <= hull.size(); ++i)
offsets.push_back(hull[i % hull.size()] - hull[i - 1]);
}
// sort offsets by CCW angle such that (0, -1) comes last
sort(begin(offsets), end(offsets), [](P a, P b) {
if (half(a) != half(b)) return half(a) < half(b);
return cross(a, b) > 0;
});
ll ans = 0;
// tour the hull in counterclockwise order
for (int i = 0; i < offsets.size(); ++i) {
ans = max(ans, norm(start));
start = start + offsets[i];
}
cout << ans << "\n";
}


Danny Mittal's code:

import java.io.BufferedReader;
import java.io.IOException;
import java.util.*;

public class MultipleChoiceTest {

public static void main(String[] args) throws IOException {
List<Vector> exchanges = new ArrayList<>();
Vector curr = new Vector(0, 0);
for (; n > 0; n--) {
Vector[] group = new Vector[g];
for (int j = 0; j < g; j++) {
long x = Long.parseLong(tokenizer.nextToken());
long y = Long.parseLong(tokenizer.nextToken());
group[j] = new Vector(x, y);
}
Arrays.sort(group, (u, v) -> {
if (v.x == u.x) {
return Long.compare(v.y, u.y);
} else {
return Long.compare(v.x, u.x);
}
});
List<Vector> upperHull = new ArrayList<>();
for (Vector v : group) {
while (upperHull.size() >= 2 && sideOfLine(upperHull.get(upperHull.size() - 2), upperHull.get(upperHull.size() - 1), v) >= 0L) {
upperHull.remove(upperHull.size() - 1);
}
}
List<Vector> lowerHull = new ArrayList<>();
for (Vector v : group) {
while (lowerHull.size() >= 2 && sideOfLine(lowerHull.get(lowerHull.size() - 2), lowerHull.get(lowerHull.size() - 1), v) <= 0L) {
lowerHull.remove(lowerHull.size() - 1);
}
}
for (int j = upperHull.size() - 1; j > 0; j--) {
Vector before = upperHull.get(j);
Vector after = upperHull.get(j - 1);
}
for (int j = 0; j < lowerHull.size() - 1; j++) {
Vector before = lowerHull.get(j);
Vector after = lowerHull.get(j + 1);
}
Vector lowest = group;
for (Vector v : group) {
if (v.y < lowest.y || (v.y == lowest.y && v.x < lowest.x)) {
lowest = v;
}
}
curr = curr.plus(lowest);
}
exchanges.sort((u, v) -> {
int uHalf = u.y > 0L || (u.y == 0L && u.x > 0L) ? 0 : 1;
int vHalf = v.y > 0L || (v.y == 0L && v.x > 0L) ? 0 : 1;
if (vHalf != uHalf) {
return uHalf - vHalf;
}
int uWhetherZero = u.y == 0L ? 0 : 1;
int vWhetherZero = v.y == 0L ? 0 : 1;
if (uWhetherZero == 0 || vWhetherZero == 0) {
return uWhetherZero - vWhetherZero;
}
return Long.compare(v.x * u.y, u.x * v.y);
});
for (Vector exchange : exchanges) {
curr = curr.plus(exchange);
}
}

static class Vector {
final long x;
final long y;

Vector(long x, long y) {
this.x = x;
this.y = y;
}

Vector plus(Vector other) {
return new Vector(x + other.x, y + other.y);
}

Vector minus(Vector other) {
return new Vector(x - other.x, y - other.y);
}

long magnitude() {
return (x * x) + (y * y);
}

@Override
public String toString() {
return "Vector{" +
"x=" + x +
", y=" + y +
'}';
}
}

static long sideOfLine(Vector a, Vector b, Vector c) {
long left = (b.x - a.x) * (c.y - a.y);
long right = (b.y - a.y) * (c.x - a.x);
return left - right;
}
}