Randomized Algorithms (Spring 2010)/Complexity classes and lower bounds

Computational Models

Upper bounds, lower bounds

Bounds are just inequalities. An inequality

[math]\displaystyle{ A\le B }[/math]

is read "[math]\displaystyle{ B }[/math] is a lower bound of [math]\displaystyle{ A }[/math]" or equivalently "[math]\displaystyle{ A }[/math] is an upper bound of [math]\displaystyle{ B }[/math]".

In Computer Science, when talking about upper or lower bounds, people really mean the upper or lower bounds for complexities.

In this lecture, we are focused on the time complexity, yet there are several other complexity measures in various computational models (e.g. space complexity, communication complexity, query complexity). The complexity is represented as a function of the length [math]\displaystyle{ n }[/math] of the input.

There are two fundamental ways of measuring complexities

Complexity of algorithms

For an algorithm [math]\displaystyle{ A }[/math], the (worst-case) time complexity of [math]\displaystyle{ A }[/math] is the maximum running time for all inputs [math]\displaystyle{ x }[/math] of length [math]\displaystyle{ n }[/math].

Complexity of problems

For a computational problem [math]\displaystyle{ f }[/math], its time complexity is the time complexity of the best possible algorithm solving the problem.

The complexity of an algorithm tells how good a solution is, yet the complexity of a problem tells how hard a problem is. While the former is what we care mostly about in practice, the later is more about the fundamental truths of computation.

In Theoretical Computer Science, when talking about upper or lower bounds, people usually refer to the bounds of the complexity of a problem, rather than that of an algorithm. Therefore, an algorithm or a solution is called an upper bound; and bad news such as impossibility results are called lower bounds.

Today's lecture is devoted to lower bounds, i.e. the necessary prices which have to paid by any algorithms which solve the given problems. Speaking of necessary prices, we have to be specific about the model, the mathematical rules which describe what problems are and what algorithms are.

Decision problems

Computational problems are formalized as mappings from inputs to outputs, also called instances and solutions. Decision problems are the computational problems with yes-or-no answers.

For a decision problem [math]\displaystyle{ f }[/math], its positive instances are the inputs with "yes" answers. Sometimes, a decision problem is equivalently represented as the set of all positive instances, denoted [math]\displaystyle{ L }[/math]. We call [math]\displaystyle{ L }[/math] a formal language (not entirely the same thing as the c language). The task of computing [math]\displaystyle{ f(x) }[/math] is equivalent to that of determining whether [math]\displaystyle{ x\in L }[/math]. Therefore the two formulations of "decision problems" and "formal languages" can be interchangeably used.

Turing Machine

To talk anything about the complexity, the limit of computations, we have to formulate what computation is! This work was done by Alan Turing in 1937. Today, the model is referred by the name Turing machine.

Complexity Classes

Problems are organized into classes according to their complexity in respective computational models. There are nearly 500 classes collected by the Complexity Zoo.

P, NP

A deterministic algorithm [math]\displaystyle{ A }[/math] is polynomial time if its running time is within a polynomial of [math]\displaystyle{ n }[/math] on any input of length [math]\displaystyle{ n }[/math].

Definition 1 The class P is the class of decision problems that can be computed by polynomial time algorithms.

We introduce the infamous NP class.

Definition 2 The class NP consists of all decision problems [math]\displaystyle{ f }[/math] that have a polynomial time algorithm [math]\displaystyle{ A }[/math] such that for any input [math]\displaystyle{ x }[/math],

[math]\displaystyle{ f(x)=1 }[/math] if and only if [math]\displaystyle{ \exists y }[/math], [math]\displaystyle{ A(x,y)=1 }[/math], where the size of [math]\displaystyle{ y }[/math] is within polynomial of the size of [math]\displaystyle{ x }[/math]

Informally, NP is the class of decision problems that the "yes" answers can be verified in polynomial time. The string [math]\displaystyle{ y }[/math] in the definition is called a certificate or a witness. The algorithm [math]\displaystyle{ A }[/math] verifies (instead of computing) the positive result of [math]\displaystyle{ f(x) }[/math] providing a polynomial size certificate.

This definition one of the equivalent definitions of the NP class. Another definition (also a classic one) is that NP is the class of decision problems that can be computed by polynomial time nondeterministic algorithms.

Common misuses of terminology:

• "This algorithm is NP." --- NP is a class of decision problems, not algorithms.
• "This problem is a NP problem, so it's quite hard." --- By definition, a problem is in NP if it's positive instances are poly-time verifiable, which implies nothing about the hardness. You probably means the problem is NP-hard.
• "NP problems are the hardest problems." --- There are infinitely many harder problems outside NP. Actually, according to a widely believed conjecture, there are infinitely many classes of problems which are harder than NP (see [1]).

Note that unlike P, the definition of NP is asymmetric. It only requires the positive instances (the [math]\displaystyle{ x }[/math] that [math]\displaystyle{ f(x)=1 }[/math]) to be poly-time verifiable, but does not say anything abou the negative answers [math]\displaystyle{ f(x)=0 }[/math]. The class for that case is co-NP.

Definition 3 The class co-NP consists of all decision problems [math]\displaystyle{ f }[/math] that have a polynomial time algorithm [math]\displaystyle{ A }[/math] such that for any input [math]\displaystyle{ x }[/math],

[math]\displaystyle{ f(x)=0 }[/math] if and only if [math]\displaystyle{ \exists y }[/math], [math]\displaystyle{ A(x,y)=0 }[/math], where the size of [math]\displaystyle{ y }[/math] is within polynomial of the size of [math]\displaystyle{ x }[/math]

Clearly, P [math]\displaystyle{ \subseteq }[/math] NP [math]\displaystyle{ \cap }[/math] co-NP. Does P = NP [math]\displaystyle{ \cap }[/math] co-NP? We don't know.

RP, BPP, ZPP

Now we proceeds to define complexity classes of the problems that can be solved efficiently by randomized algorithms. There are three types of randomized algorithms: Monte Carlo algorithms with one-sided errors, Monte Carlo algorithms with two-sided errors, and Las Vegas algorithms, which define the following three complexity classes: RP (for Randomized Polynomial time), BPP (for Bounded-error Probabilistic Polynomial time), and ZPP (for Zero-error Probabilistic Polynomial time).

For Monte Carlo algorithms, the running time is fixed but the correctness is random. If a Monte Carlo algorithm [math]\displaystyle{ A }[/math] errs only when [math]\displaystyle{ f(x)=1 }[/math], i.e. [math]\displaystyle{ A }[/math] only has fals negatives, then it is called with one-sided error; otherwise if a Monte Carlo algorithm has both false positives and false negatives, then it is called with two-sided error.

We first introduce the class RP of the problems which can be solved by polynomial time Monte Carlo algorithms with one-sided error.

Definition 4 The class RP consists of all decision problems [math]\displaystyle{ f }[/math] that have a randomized algorithm [math]\displaystyle{ A }[/math] running in worst-case polynomial time such that for any input [math]\displaystyle{ x }[/math],

• if [math]\displaystyle{ f(x)=1 }[/math], then [math]\displaystyle{ \Pr[A(x)=1]\ge 1/2 }[/math];
• if [math]\displaystyle{ f(x)=0 }[/math], then [math]\displaystyle{ \Pr[A(x)=1]=0 }[/math].

Remark

The choice of the error probability bound is arbitrary. In fact, replacing the 1/2 with any constant [math]\displaystyle{ 0\lt p\lt 1 }[/math] will not change the definition of RP. With polynomially many independent repetitions (the total running time is still polynomial), the error probability can be reduced to exponentially small from any constant [math]\displaystyle{ 0\lt p\lt 1 }[/math].

Example

Define the decision version of the minimum cut problem as follows. For a graph [math]\displaystyle{ G }[/math], [math]\displaystyle{ f(G)=1 }[/math] if and only if there exists any cut of size smaller than [math]\displaystyle{ k }[/math], where [math]\displaystyle{ k }[/math] is an arbitrary parameter. Obviously, [math]\displaystyle{ f }[/math] can be answered probabilistically by Karger's min-cut algorithm in polynomial time. The error is one-sided, because if there does not exists any cut of size smaller than [math]\displaystyle{ k }[/math], then obviously the algorithm cannot find any. Therefore [math]\displaystyle{ f\in }[/math]RP.

Like NP, the class RP is also asymmetrically defined, which hints us to define its co-class.

Definition 5 The class co-RP consists of all decision problems [math]\displaystyle{ f }[/math] that have a randomized algorithm [math]\displaystyle{ A }[/math] running in worst-case polynomial time such that for any input [math]\displaystyle{ x }[/math],

• if [math]\displaystyle{ f(x)=1 }[/math], then [math]\displaystyle{ \Pr[A(x)=1]= 1 }[/math];
• if [math]\displaystyle{ f(x)=0 }[/math], then [math]\displaystyle{ \Pr[A(x)=1]\le 1/2 }[/math].

We define the class BPP of the problems which can be solved by polynomial time Monte Carlo algorithms with two-sided error.

Definition 6 The class BPP consists of all decision problems [math]\displaystyle{ f }[/math] that have a randomized algorithm [math]\displaystyle{ A }[/math] running in worst-case polynomial time such that for any input [math]\displaystyle{ x }[/math],

• if [math]\displaystyle{ f(x)=1 }[/math], then [math]\displaystyle{ \Pr[A(x)=1]\ge 3/4 }[/math];
• if [math]\displaystyle{ f(x)=0 }[/math], then [math]\displaystyle{ \Pr[A(x)=1]\le 1/4 }[/math].

Remark

Replacing the error probability from [math]\displaystyle{ \frac{1}{4} }[/math] to [math]\displaystyle{ \frac{1}{3} }[/math], or any constant [math]\displaystyle{ 0\lt p\lt \frac{1}{2} }[/math] will not change the definition of the BPP class.

And finally the class ZPP of the problems which can be solved by polynomial time Las Vegas algorithms. For Las Vegas algorithms, the running time is not fixed, but a random variable, therefore we actually refer to the Las Vegas algorithms whose expected running time is a polynomial of the size of the input, where the expectation is taken over the internal randomness (random coin flippings) of the algorithm.

Definition 7 The class ZPP consists of all decision problems [math]\displaystyle{ f }[/math] that have a randomized algorithm [math]\displaystyle{ A }[/math] running in expected polynomial time such that for any input [math]\displaystyle{ x }[/math], [math]\displaystyle{ A(x)=f(x) }[/math].

---

What is known (you can even prove by yourself):

• RP, BPP, ZPP all contain P
• ZPP = RP[math]\displaystyle{ \cap }[/math]co-RP;
• RP [math]\displaystyle{ \subseteq }[/math] BPP;
• co-RP [math]\displaystyle{ \subseteq }[/math] BPP.

Open problem:

• BPP vs P (the second most important open problem in the complexity theory to the P vs NP).