Worst number of questions needed to learn a monotonic predicate over a poset

Question

Consider $(X, \leq)$ a finite poset over $n$ items, and $P$ an unknown monotonic predicate over $X$ (i.e., for any $x$, $y \in X$, if $P(x)$ and $x \leq y$ then $P(y)$). I can evaluate $P$ by providing one node $x \in X$ and finding out if $P(x)$ holds or not. My goal is to determine exactly the set of nodes $x \in X$ such that $P(x)$ holds, using as few evaluations of $P$ as possible. (I can choose my queries depending on the answer of all previous queries, I am not required to plan all queries in advance.)

A strategy $S$ over $(X, \leq)$ is a function which tells me, as a function of the queries that I ran so far and their answers, which node to query, and which ensures that on any predicate $P$, by following the strategy, I will reach a state in which I know the value of $P$ on all nodes. The running time $r(S, P)$ of $S$ on a predicate $P$ is the number of queries required to know the value of $P$ on all nodes. The worst running time of $S$ is $wr(S) = \max_P r(S, P)$. An optimal strategy $S'$ is such that $wr(S') = \min_S wr(S)$.

My question is the following: given as input the poset $(X, \leq)$, how can I determine the worst running time of the optimal strategies?

[It is clear that for an empty poset $n$ queries will be needed (we need to ask about each single node), and that for a total order around $\lceil \log_2 n \rceil$ queries will be needed (doing a binary search to find the frontier). A more general result is the following information-theoretic lower bound: the number of possible choices for the predicate $P$ is the number $N_X$ of antichains of $(X, \leq)$ (because there is a one-to-one mapping between monotonic predicates and antichains interpreted as the maximal elements of $P$), so, since each query gives us one bit of information, we will need at least $\lceil \log_2 N_X \rceil$ queries, subsuming the two previous cases. Is this bound tight, or are them some posets whose structure is such that learning can require asymptotically more queries than the number of antichains?]

Answer 1

This is not a complete answer, but it's too long to be a comment.

I think I found an example for which the bound $\lceil \log_2 N_X \rceil$ is not tight.

Consider the following poset. The ground set is $X=\{a_1, a_2, b_1, b_2\}$, and $a_i$ is smaller than $b_j$ for all $i,j\in\{1,2\}$. The other pairs are incomparable. (The Hasse diagram is a $4$-cycle).

Let me identify the monotone properties with the upsets of the poset. This poset has seven upsets: $\emptyset$, $\{b_1\}$, $\{b_2\}$, $\{b_1,b_2\}$, $\{a_1,b_1,b_2\}$, $\{a_2,b_1,b_2\}$, $\{a_1,a_2,b_1,b_2\}$, and this poset has seven antichains since the antichains are in one-to-one correspondence with the upsets. So, $\lceil \log_2 N_X \rceil=\lceil \log_2 7 \rceil = 3$ for this poset.

Now, by adversary argument I'll show that any strategy needs at least four queries (so needs to query all elements). Let's fix an arbitrary strategy.

If the strategy first queries $a_1$, then the adversary answers "$P(a_1)$ doesn't hold." Then, we are left with five possibilities: $\emptyset$, $\{b_1\}$, $\{b_2\}$, $\{b_1,b_2\}$, $\{a_2,b_1,b_2\}$. Thus, to determine which is the case, we need at least $\lceil \log_2 5\rceil = 3$ more queries. In total, we need four queries. The same argument applies if the first query is $a_2$.

If the strategy first queries $b_1$, then the adversary answers "$P(b_1)$ holds." Then, we are left with five possibilities: $\{b_1\}$, $\{b_1,b_2\}$, $\{a_1,b_1,b_2\}$, $\{a_2,b_1,b_2\}$, $\{a_1,a_2,b_1,b_2\}$. Therefore, we need at least three more queries as before. In total, we need four queries. The same argument applies when the first query is $b_2$.

If we take $k$ parallel copies of this poset, then it has $7^k$ antichains, and thus the proposed bound is $\lceil \log_2 7^k \rceil = 3k$. But, since each of the copies needs four queries, we need at least $4k$ queries.

Probably, there is a larger poset with larger gap. But this argument can only improve the coefficient.

Here, the problem looks to be a situation where no query partitions the search space evenly. In such a case, the adversary can force the larger half to remain.

Answer 2

In their paper Every Poset Has a Central Element, Linial and Saks show (Theorem 1) that the number of queries required to solve the ideal identification problem in a poset $X$ is at most $K_0 \log_2 i(X)$, where $K_0 = 1/(2 - \log_2(1 + \log_2 5))$ and $i(X)$ is the number of ideals of $X$. What they call an "ideal" is actually a lower set and there is an obvious one to one correspondance between monotonic predicates and the lower set of the points at which they don't hold, besides their "identification problem" is to identify by querying nodes just like in my setting, so I think they are dealing with the problem I'm interested in and that $i(X) = N_X$.

So, according to their result, the information-theoretic lower bound is tight up to a relatively small multiplicative constant. So this basically settles the question of the number of questions required, as a function of $N_X$ and up to a multiplicative constant: it is between $\log_2 N_X$ and $K_0 \log_2 N_X$.

Linial and Saks quote a personal communication by Shearer to say that there are known orders for which we can prove a lower bound of $K_1 \log_2 N_X$ for some $K_1$ which is just slightly less than $K_0$ (this is in the spirit of Yoshio Okamoto's answer who tried this approach for a smaller value of $K_1$).

This does not fully answer my question of computing the number of questions required from $X$, however, since computing $N_X$ from $X$ is #P-complete, I have a feeling that there is little hope. (Comments about this point are welcome.) Still, this result by Linial and Saks is enlightening.

Answer 3

For the Boolean n-cube $(\{0, 1\}^n, \leq)$ (or, equivalently, for the poset $(2^S, \subseteq)$ of all subsets of an n-element set), the answer is given by Korobkov and Hansel's theorems (from 1963 and 1966, respectively). Hansel's theorem [1] states that an unknown monotone Boolean function (i.e., an unknown monotone predicate on this poset) can be learned by a deterministic algorithm making at most $\phi(n) = \binom{n}{\lfloor n/2 \rfloor} + \binom{n}{\lfloor n/2 \rfloor + 1}$ queries (that is, asking $\phi(n)$ questions in the worst case). This algorithm matches the lower bound of Korobkov's theorem [2], which says that $\phi(n) - 1$ queries do not suffice. (So Hansel's algorithm is optimal in the worst-case setting.) An algorithm in both statements is understood as a deterministic decision tree.

The logarithm of the number of antichains in $(\{0, 1\}^n, \le)$ is asymptotically equal to $\binom{n}{\lfloor n/2 \rfloor} \sim 2^n / \sqrt{\pi n / 2}$, so there is a constant-factor gap between $\log N_X$ and the optimal algorithm performance $\phi(n) \sim 2 \binom{n}{\lfloor n/2 \rfloor}$ for this poset.

Unfortunately, I have not been able to find a good treatment of Hansel's algorithm in English available on the web. It is based on a lemma that partitions the n-cube into $\phi(n)$ chains with special properties. Some description can be found in [3]. For the lower bound, I don't know any reference to a description in English.

Since I am familiar with these results, I can post a description on arXiv, if the treatment in Kovalerchuk's paper does not suffice.

If am not much mistaken, there have been attempts to generalize Hansel's approach, at least to the poset $(E_k^n, \le)$, where $(E_k, \le)$ is a chain $0 < 1 < \ldots < k - 1$, although I cannot give any reference straight away. For the Boolean case, people have also investigated notions of complexity other than worst-case for this problem.

[1] G. Hansel, Sur le nombre des fonctions booléennes monotones de n variables. C. R. Acad. Sci. Paris, 262(20), 1088-1090 (1966)

[2] V. K. Korobkov. Estimation of the number of monotonic functions of the algebra of logic and of the complexity of the algorithm for finding the resolvent set for an arbitrary monotonic function of the algebra of logic. Soviet Math. Doklady 4, 753-756 (1963) (translation from Russian)

[3] B. Kovalerchuk, E. Triantaphyllou, A. S. Deshpande, E. Vityaev. Interactive learning of monotone Boolean functions. Information Sciences 94(1), 87-118 (1996) (link)

Answer 4

[NOTE: The following argument doesn't seem to work, but I'm leaving it here so others don't make the same mistake / in case someone can fix it. The issue is that an exponential lower bound on learning/identifying a monotone function, as below, does not necessarily contradict an incrementally polynomial algorithm for the problem. And it is the latter which is equivalent to checking the mutual duality of two monotone functions in poly time.]

I believe your conjecture on $\log N_X$ is false in general. If it is indeed the case that $\log N_X$ queries are needed, that implies quite a strong lower bound on learning monotone functions using membership queries. In particular, let the poset $X$ be the Boolean cube with the usual ordering (if you like, $X$ is the powerset of $\{1,...,n\}$ with $\subseteq$ as its partial order). The number $M$ of maximal antichains in $X$ satisfies $\log M = (1 + o(1))\binom{n-1}{\lfloor n/2 \rfloor}$ [1]. If your idea on $\log N_X$ is correct, then there is some monotone predicate on $X$ that requires essentially $\binom{n-1}{n/2} \approx 2^n$ queries. In particular, this implies a lower bound of essentially $2^n$ for the complexity of any algorithm solving this problem.

However, if I've understood correctly [which I now know I hadn't], your problem is equivalent to checking the mutual duality of two monotone functions, which can be done in quasi-polynomial time (see the intro of this paper by Bioch and Ibaraki, which cites Fredman and Khachiyan), contradicting anything close to a $2^n$ lower bound.

[1] Liviu Ilinca and Jeff Kahn. Counting maximal antichains and independent sets. arXiv:1202.4427