_h_'s blog

By _h_, history, 3 years ago, In English

Problem 1466I - The Riddle of the Sphinx, authored by gawry and Anadi, raised some interesting questions. The problem asks to find the maximum among $$$n$$$ secret $$$b$$$-bit $$$a_1 \ldots a_n$$$ numbers using $$$3n+3b$$$ questions of the form "is $$$a_i$$$ at most $$$x$$$"? ecnerwala described a solution using only $$$2n+2b-2$$$ questions. Here I will describe a solution using only $$$2n + 1.45b$$$ questions, using Fibonacci numbers instead of powers of two. In a sense this solution looks near-optimal, but there are better solutions when one of $$$n, b$$$ is much smaller than the other. In the other direction, a fairly simple argument shows that $$$n + b - 1$$$ questions are always needed in the worst case. I'll also sketch a proof that $$$n + b + 0.3 \min(n, b)$$$ (maybe minus some constant term) questions are needed. This shows that no solution can hope to achieve e.g $$$1.1 n + 1.1 b$$$, but still leaves a wide gap to $$$2n + 1.45b$$$.

I'll be sloppy with $$$+O(1)$$$ terms throughout, so while we are interested in the difference between $$$n$$$ and $$$2n$$$, I might forget about the difference between $$$n$$$ and $$$n + 10$$$. The ideas here arose in discussion with noncomlp. If you any ideas on how to improve the bounds in this post, please share them!

Terminology

Throughout this post, I will be using a straightforward reformulation of the original problem. In this reformulation, the state of the game is described by a multiset $$$P$$$ of integers all at least $$$2$$$. The game ends when $$$P$$$ is empty. In one move, we may name an integer $$$r$$$. Then one of two things happen: either (one occurrence of) $$$\max P$$$ is replaced by $$$r$$$, or $$$P$$$ is replaced by $$$P - r := [ p - r \mid p \in P, p - r \ge 2]$$$. The connection with the original problem comes from the observation that instead of remembering upper / lower bounds $$$u_i$$$ / $$$l_i$$$ on $$$a_i$$$ for each $$$1 \le i \le n$$$, it suffices to remember upper bounds $$$u_i$$$ and the greatest lower bound $$$L$$$ we've seen so far. We then consider the multiset $$$[ u_i - L + 1 \mid u_i > L]$$$, and observe that asking if $$$a_i \stackrel{?} \le x$$$ is like naming $$$r = x - L + 1$$$ in the reformulation. (This relies on the observation that we can always pick the $$$i$$$ which maximises $$$u_i$$$.) Thus the number of questions needed for $$$n, b$$$ is the number of moves needed to win the game starting from the multiset of size $$$n$$$ containing just the number $$$2^b$$$.

Fibonacci strategy

The problem hints to use powers of two, but let's instead use Fibonacci numbers. Basically this is motivated by the observation that if $$$\min P = F_i$$$ and we name $$$F_{i-1}$$$, then if we end up subtracting $$$F_{i-1}$$$, $$$F_i$$$ becomes $$$F_{i-2}$$$. Instead of starting from $$$P = [2^b, \ldots, 2^b]$$$, let's suppose we start from $$$P = [F_c, \ldots, F_{c+n-1}]$$$. I'll describe a solution which uses only $$$2n + c$$$ questions. This solves the original problem in $$$2n + 1.45b$$$ questions, where $$$1.45 = \log_2 \phi$$$, since $$$F_{n+1} \ge \phi^n$$$.

It is necessary to consider a more general class of multisets than the one above, since we might end subtracting Fibonacci numbers from much larger Fibonacci numbers. We'll assume we have some $$$c \ge 3$$$ and a string $$$S$$$ containing characters 1 and x, where no x is followed by another x. The multiset $$$P = P(c, S)$$$ is then read off as $$$P(c, \emptyset) = \emptyset$$$, $$$P(c, 1S) = [F_c] \cup P(c+1, S)$$$, $$$P(c, xS) = P(c+1, S) - F_c$$$. Thus the multiset above corresponds to the string $$$S$$$ of $$$n$$$ 1s. We claim that any such state can be done in $$$2n + c$$$ questions, where $$$n$$$ is the number of 1s in $$$S$$$.

The strategy in this class is fairly simple. We always name $$$F_{c-1}$$$. In the case where $$$\max P$$$ is replaced by $$$F_{c-1}$$$, we simply move the last 1 of $$$S$$$ to the start, and subtract one from $$$c$$$. In the case where $$$F_{c-1}$$$ is subtracted from all elements of $$$P$$$, we prepend an x to $$$S$$$ and subtract one from $$$c$$$. Now if the string looks like xx1T, then we reduce it to xT, and increase c by 2. This doesn't change $$$P(c, S)$$$ (since $$$F_c + F_{c+1} = F_{c+2}$$$), nor our monovariant $$$2n + c$$$, and after repeatedly doing this operation, no x in $$$S$$$ is followed by another x.

We should be careful to note that $$$c$$$ need never be negative, since we don't want negative Fibonacci numbers, and we also don't want our monovariant to go negative. Indeed already if $$$c = 2$$$, we can increase $$$c$$$ and remove the first one or two characters of $$$S$$$, since they give some number which is at most $$$1$$$.

Here is my implementation of this strategy: 103038134.

Lower bounds

Here is a nice fact: if some strategy never finishes in less than $$$z$$$ moves starting from $$$P$$$, then no strategy can guarantee to finish in less than $$$z$$$ moves starting from $$$P$$$. The proof is simple. Say, starting from $$$P$$$, the first strategy names $$$r$$$, and the second strategy names $$$t$$$. If $$$r \le t$$$, consider what happens when $$$\max$$$ is replaced by $$$r$$$ or $$$t$$$: the first strategy is better off. On the other hand, if $$$r \ge t$$$, then when $$$r$$$ or $$$t$$$ is subtracted from all numbers, then the first strategy is better off.

Thus the problem of showing that no strategy always does well reduces to finding a strategy which never does well. Here is one such strategy, starting from $$$P = [2^b, \ldots, 2^b]$$$ (size $$$n$$$). If $$$b = 1$$$, then we name $$$r = 0$$$. If $$$b \ge 1$$$, then we name $$$2^{b-1}$$$. In the best case, $$$b$$$ decreases by $$$1$$$ each step until $$$b = 1$$$, at which point $$$n$$$ decreases by $$$1$$$ each step until $$$n = 0$$$. Thus $$$n + b - 1$$$ questions are needed.

The Fibonacci strategy looks like a good candidate for "strategy that never does well", since the monovariant should decrease by exactly 1 in every step. Unfortunately, it can start doing well in special cases, like when $$$c$$$ or $$$n$$$ is too small. There is a good reason for this, explained in the next section. For the rest of this section, I'll sketch a modified Fibonacci strategy which really never does too well.

Where we previously considered the alphabet with two letters 1 and x, let's now expand the alphabet to allow for arbitrary non-negative integers. Where (c, "1") corresponds to $$$[F_c]$$$, we'll let (c,"m") correspond to $$$[F_c, \ldots, F_c]$$$ (size $$$m$$$). The states we consider are given by $$$c\ge 3$$$ and strings $$$S$$$ of the form above followed by a string 0m with $$$m \ge 1$$$. Thus our initial state looks like (c, 0n). Our strategy is the same as before (instead of moving 1s from end to beginning, we replace n with (n-1) and put a 1 at the beginning), with the added reduction rule (c, x0m) -> (c, 0m) (corresponding to $$$F_{c+2} - F_c = F_{c+1}$$$). Eventually $$$c$$$ or $$$m$$$ gets small enough that we can't stay in this class of states, and our strategy might start doing well. But for the first $$$M := \min(c, n) - 5$$$ moves, we'll certainly stay in this class. Thus for the first $$$M$$$ turns, the monovariant decreases by at most 1. After these first $$$M$$$ turns, we consider the lower bound $$$|P| + \log_2 \min P$$$ from before. The most this could decrease after $$$M$$$ turns is $$$M \log_2 \phi$$$.

Taken together, this means that in the original problem with $$$n, b$$$, $$$n > M + 5$$$, $$$2^b > F_{M+5}$$$, at least $$$M + n + b - M \log_2 \phi > n + b + 0.3M$$$ questions are needed.

Good strategies when either $$$n$$$ or $$$b$$$ is small

A silly strategy is to always name $$$1$$$. This always wins in at most $$$|P| + \max P - 2$$$ moves. So we get a lower bound on the original problem of $$$n + 2^b - 2$$$. Note that this is better than $$$2n$$$ when $$$b$$$ is very small.

At the other extreme, consider $$$n$$$ fixed. I claim that the answer is $$$(1 + o(1)) b$$$ as $$$b \to \infty$$$. The idea is to pick some positive integer parameter $$$D$$$, and consider multisets of the form $$$[2^k | k \in Q]$$$, where $$$\max Q - \min Q \le D$$$. Given such a state, I claim that we can finish in $$$\min Q + (D+1)|Q| + \sum Q / D$$$ moves. Indeed if $$$\max Q - \min Q < D$$$, then we can name $$$2^{\min Q-1}$$$, and $$$\min Q$$$ decreases by 1. If $$$\max Q = \min Q + D$$$, then we name $$$2^{\min Q}$$$. Now either $$$|Q|$$$ decreases by $$$1$$$ and $$$\min Q$$$ increases by at most $$$D$$$, or $$$\sum Q$$$ decreases by $$$1$$$.

In the original problem, we start with $$$\min Q = \max Q = b$$$, $$$|Q| = n$$$, $$$\sum Q = bn$$$. The optimal value of $$$D$$$ is about $$$\sqrt b$$$. This shows that we need at most about $$$2 n\sqrt b + b$$$ questions. This is good when $$$n \ll \sqrt b$$$.

  • Vote: I like it
  • +195
  • Vote: I do not like it

»
3 years ago, # |
  Vote: I like it +56 Vote: I do not like it

Awesome writeup! The lower bound proof strategy in particular is very elegant.

The motivation for Fibonacci base makes sense since in the binary solution each "No" amortizes to 1 query cost per bit while each "Yes" takes up to 2 -- essentially Fibonacci base fine-tunes the costs so both "No" and "Yes" cost 1.45 queries.

»
3 years ago, # |
  Vote: I like it +32 Vote: I do not like it

Wow, thanks for the blog! I'm particularly fond of interactive problems, and really enjoyed reading about this.