Computational Complexity Theory

Lecture notes by Henrik Hald Nrgaard Lecture 14, 13/10 2006

Topics for the Second Part of the Course

There are 3 themes in modern complexity theory that we could consider:

Other suggestions or preferences for one of the above are welcome... The third theme will be the subject of this lecture. Constant depth computation is a classical topic that was studied in particular in the 1980ties. It is the study of tasks that can be solved immidiately using a few massively parallel computations. Examples of such tasks are rudimentary computational opertations such as multiplication and addition of integers. There are some negative result in this eld, e.g. stating that there are certain tasks that the human brain cannot solve immidiately.

ATM with Random Access

When studying constant depth computations the top of the hierarchy of complexity classes is L. This means that our previous reductions break down. In the rest of the lecture, we will formalize the meaning of constant depth computation and dene the complexity classes that we will use to study this kind of computation. We have previously dened a model for massive parallelism, namely alternating Turing machines. This model will not be suitable here. Instead we will study a related model for sublinear time massive parallelism. Denition 1 (Alternating TM with random access) An alternating Turing machine with random access is an alternating Turing machine with an output tape, that additionally diers from a the ATM from lecture 6 in the following way: 1

1. The TM only gets the length in binary of the input x on the input tape 2. The TM will write an integer i and a bit b on the output tape and then halt 3. The TM accepts i xi = b Running times for an ATM with random access are measured in terms of |x| and thus the classes AT S (, , ) can be dened as for ordinary ATMs. The time component should though be sublinear for ATMs with random access. We now have two denitions of AT S (, , ) - and old and a new one. They are related in the following way: AT Sold ((a), t(1) , (s)) = AT Snew ((a), t(1) , (s)) for t n and s log(n). We will not prove this fact. (1)

Logspace Uniform Curcuit Families

The curcuit computed in the desert by an ATM with random access only depends on the length of the input and thus denes a uniform curcuit family C1 , C2 , . . ., that will dene a language L. The ATM with random access and the corresponding curcuit family are related in the following way (looking at inputs of a xed length): Parallel time of machine = Depth of curcuit Parallel space (s) of machine = Size of curcuit (2(s) ) Alternations of machine = Depth of the corresponding unbounded fan-in curcuit (2) (3) (4) (5)

The fan-in of a gate in a curcuit is the number of inputs to the gate and the fan-in of a curcuit is the maximal fan-in of a gate in the curcuit. In the curcuit corresponding to an ATM with random access, the gates can be assumed to be either AND- or OR-gates corresponding to the two kinds of states in the ATM (-states can be eliminated). The gates of the curcuit are naturally organized in layers, where two gates of the same type are in the same layer, if the corresponding states are entered in the same alternation of the ATM. Thus, the number of alternations of the ATM corresponds to the number of layers in the curcuit. There may be wires between gates in the same layer, but each layer can be converted into a layer of gates whose wires are all to gates in dierent layers. This transition turns the curcuit into a curcuit of unbounded fan-in and for such a curcuit the number of layers equals the depth of the curcuit. Using the above correspondence we will use an alternate point of view as our model for constant depth computation: We will look at curcuit families with some uniformity condition. The leads us to the following denitions: 2

Denition 2 (Logspace uniformity) A curcuit family (Ci : {0, 1}i {0, 1})i1 is called logspace uniform if there exists a logspace TM with output tape that on input 1n outputs Ci . Denition 3 (NC i and AC i ) NC i = {L |L is decided by a logspace uniform family of bounded fan-in curcuits of depth O ((log n) ) and polynomial size} AC i = {L |L is decided by a logspace uniform family of unbounded fan-in curcuits of depth O ((log n)i ) and polynomial size} We also dene NC :=
i1 i

(6) (7) (8) (9)

NC i .

Example: AND = {1n |n 1} NC 1 . To see this, build a binary tree with n leafes = input gates and inner nodes consisting of AND-gates. Also, AND AC 0 . Here we can just use one (big) AND-gate of fan-in n. Lemma 1 NC i AC i NC i+1 P (10)

Proof : The rst inclusion holds by dention of NC i and AC i . To prove the second inclusion, notice that a gate in an AC i -curcuit can have at most p(n) inputs as the number of wires is polynomially bounded. Such a gate can be converted to a tree of binary gates of depth O (log(n)) and performing this conversion in the hole curcuit gives a logaritmic increase in the depth of the curcuit and a polynomial increase in the size. Thus the resulting curcuit is in NC i+1 . Finally, a curcuit of polynomial size can be evaluated in polynomial time and thus NC P . Remark: Notice that this above lemma implies that NC =
i1

NC i =
i1

AC i

(11)

(and that is why we didnt dene AC ). Remark: As a rule-of-thumb our denition of NC has less restrictive uniformity conditions that the curcuits resulting of our denition of ATM with random access. A technical result is that NC = AT Snew (, (log(n))O(1) , O (log(n))) = AT Sold ((log(n)O(1) , nO(1) , O (log(n))) 3 (12) (13)

Proposition 1 NC 1 L NL AC 1 (14)

Proof : NC 1 L: Given input x construct the curcuit C|x| in logspace. Then convert C|x| to a formula and evaluate it on input x. The last two steps can also be performed in logspace and thus we have a composition of 3 logspace computations, which is again a logspace computation. NL AC 1 : It is enough to prove that the NL-complete problem GAP is in AC 1 . Let A be the adjancy matrix for the the given graph G with 1s added on k the diagonal and compute A2 , A4 , . . . , A2 , where k = log(n), n = number of nodes in the graph. The resulting matrix is the adjancy matrix for the transitive closure of the graph. Using this matrix is it trivial to solve GAP . As the adjancy matrix of G has entries in {0, 1} the matrix multiplications can be modelled as a curcuit using (, )-matrixmultiplication:
n

cij =
k =1

aik bkj

(Depth 2 curcuit)

(15)

The complete curcuit has depth O (log(n)). Remark: The above algorithm for GAP results in the same curcuit as using Savitchs algorithm from lecture 2.

Oracles and reductions

Denition 4 (Curcuits using an oracle) If A denotes an oracle then NC i [A] = {L |L is decided by a logspace uniform family of bounded fan-in (16) A-curcuits of depth O ((log n)i ) and polynomial size. (17) An A-gate is charged log(t) in terms of depth and t in terms (18) of size, where t is the number of inputs to the gate} (19) An A-gate is a special gate in the curcuit where the inputs are interpreted as an input to the oracle and the output is likewise the answer from the oracle. AC i [A] = {L |L is decided by a logspace uniform family of unbounded fan-in A-curcuits of depth O ((log n) ) and polynomial size. An A-gate is charged 1 in terms of depth and t in terms of size, where t is the number of inputs to the gate}
i

(20) (21) (22) (23)

AC 0 will replace P in the notion of reduction that we will use for constant depth computations: Denition 5 (Reduction) Let L1 and L2 be langauges. Then L1 reduces to L2 under constant depth computations i L1 can be solved by an AC 0 -curcuit using an L2 -oracle: L1 cd L2 L1 AC 0 [L2 ] (24)

4.1

Complete problems for L

Dene P AP = Path Accessability Problem to be GAP , but where the in and out degree of each vertex of the graph is at most 1 - thus reducing the graph to a union of paths and cycles. P AP L as you can write a O (1)-variable program that follows the edges from a starting node until it 1. nds the wanted node 2. cant go any further in the graph 3. times out Another problem is F AP = Forest Accessability Problem. This problem is also the same as GAP , but where the graph is a forest, that is: the in and out degree are at most 1 and the graph is acyclic. Also F AP L using the same simple program as above. Now F AP P AP : Use the Euler-path of the tree containing the start node. Furthermore, F AP (and thus P AP ) is complete for L: To see this consider the conguration graph of a language in L. As the TM deciding to this language is deterministic, the in and out degree of each node is at most 1. The graph can also be assumed to be acyclic by adding a timer to the conguration nodes and letting the TM halt when it has computed for so long that it must be in a loop. Now acceptance of the TM can be decided by using F AP on the modied conguration graph. This also shows that every L-computation can be made reversible. A result from 2005 by Reingold is that UGAP = Undirected Graph Accessability Problem is in L and it thus also complete for L.

T C0
T C 0 := AC 0 [MAJORIT Y ] (25)

Denition 6 (T C 0 )

where MAJORIT Y is the following problem: MAJORIT Y (x1 , x2 , . . . , xn ) = 1 0

(26)

Thesis (Maas, 1980): Neuronal computation from one instant to the next is reasonably described by a T C 0 -computation. Here a neuronal computation means what a single neuron does. The thesis thus claims that a T C 0 -computation gives a model of the brain on a local level. Proposition 2 AC 0 T C 0 NC 1 L (27)

Remark: If we accept the above thesis and the hierarchy in the proposition does not collapse (T C 0 = NC 1 ), then there are computational tasks (e.g. complete problems for NC 1 ) that cant be solved instantly by the brain. Examples of such tasks are P UGAP = Planar UGAP and P P AP = Planar P AP , that are both complete for NC 1 . Start of proof : It is enough to prove that MAJORIT Y NC 1 since by denition this problem is complete for T C 0 . But this will have to wait for the next lecture... Exercise: ADDIT ION AC 0 . Here our school-methods for addition do not work as the carries requires a sequential scan of the given bitstrings. Remark: The following problems are complete for T C 0 : MULT IP LICAT ION , INT EGER DIV ISION and INT EGER SORT ING.