Computational Depth and Reducibility
Computational Depth and Reducibility
Abstract
This paper reviews and investigates Bennett's notions of strong
and weak computational depth (also called logical depth) for innite
binary sequences. Roughly, an innite binary sequence x is dened to
be weakly useful if every element of a non-negligible set of decidable
sequences is reducible to x in recursively bounded time. It is shown
that every weakly useful sequence is strongly deep. This result (which
generalizes Bennett's observation that the halting problem is strongly
deep) implies that every high Turing degree contains strongly deep
sequences. It is also shown that, in the sense of Baire category, almost
every innite binary sequence is weakly deep, but not strongly deep.
Contents
1 Introduction 2
2 Preliminaries 5
3 Measure and Category 7
4 Algorithmic Information and Randomness 14
5 Strong Computational Depth 22
6 Weak Computational Depth 37
7 Conclusion 40
This research was supported in part by National Science Foundation Grant CCR-
9157382, with matching funds from Rockwell International and Microware Systems
Corporation.
1
1 Introduction
Algorithmic information theory, as developed by Solomono 51],
Kolmogorov 21, 22, 23], Chaitin 9, 10, 11, 12], Martin-Lof 39, 40], Levin
26, 27, 28, 29, 30, 31, 55], Schnorr 47], Gacs 15], Shen 48, 49], and others,
0
oracle x, computes y n], the nth bit of y , in at most s(l) steps, where l is the
number of bits in the binary representation of n. For example, consider the
diagonal halting problem K 2 f0 1g , whose nth bit K n] is 1 if and only
1
time.
An interesting feature of this example is that K has relatively low in-
formation content. In fact, an n-bit prex of K , denoted K 0::n ; 1],
contains only O(log n) bits of algorithmic information 3]. Intuitively, this
is because K 0::n ; 1] is completely specied by the number of indices
2
i 2 f0 : : : n ; 1g such that the ith Turing machine Mi halts on input i. Once
this O(log n)-bit number is known, direct simulation of M0 M1 Mn 1 ;
dom in the equivalent senses of Martin-Lof 39], Levin 26], Schnorr 47],
Chaitin 11], Solovay 52], and Shen 48, 49]. (See section 4 below for a
0
decides y n], the nth bit of y , with error probability less than n1 , using time
that is at most polynomial in the number of bits in the binary representation
of n.) Since BPP contains only the simplest recursive sequences, this means
that, for the purpose of eciently deciding recursive sequences, K is much
more useful than an algorithmically random sequence z .
Bennett has argued that the computational usefulness of K derives not
from its algorithmic information content (which is relatively low), but rather
from its computational depth. In support of this thesis, Bennett 5] has
proven that K is strongly deep, while no algorithmically random sequence
can even be weakly deep. (Precise denitions of these terms appear in sec-
tions 5 and 6 below.)
This paper furthers Bennett's investigation of the computational depth
of innite binary sequences. We pay particular, quantitative attention to
interactions between computational depth and time-bounded Turing reduc-
tions.
In order to further investigate the above-discussed notion of the compu-
tational usefulness of a sequence x 2 f0 1g , we quantify the size of the set
1
3
this, we use a special case of the resource-bounded measure theory of Lutz
37, 36]. (A detailed description of the relevant special case appears in sec-
tion 3 below.) Intuitively, this theory, a generalization of classical Lebesgue
measure theory, denes a set X of innite binary sequences to have measure
0 in REC if X \ REC is a negligibly small subset of REC.
In this paper, we dene a sequence x 2 f0 1g to be weakly useful
1
4
section 3.) Intuitively, this means that weakly deep sequences are \topolog-
ically abundant." (They \cannot be avoided" by one player in a two-person
game described in section 3.) In contrast, weakly deep sequences are \prob-
abilistically scarce," in the sense that, with respect to Lebesgue measure,
almost every sequence x 2 f0 1g is algorithmically random 39], hence not
1
weakly deep.
In order to provide a basis for further investigation of Bennett's funda-
mental ideas, this paper also includes a self-contained mathematical treat-
ment of the weak and strong computational depth of innite sequences. In
section 2 we introduce our basic terminology and notation. In section 3
we review fundamental ideas of Baire category and measure that are used
in our work. In section 4 we give a similar review of algorithmic informa-
tion and randomness. Section 5 is the main section of the paper. In this
section, we present the strong computational depth of innite binary se-
quences in a unied, self-contained framework using a convenient family of
parametrized depth classes, Dtg . This framework is used to prove our main
result (Theorem 5.11), that every weakly useful sequence is strongly deep.
In the course of our development, we prove several results, some of which
were already proven by Bennett 5], giving precise, quantitative relationships
among depth, randomness, and recursiveness. We also prove (Theorem 5.16)
that strongly deep sequences are extremely rare, in that they form a meager,
measure 0 subset of f0 1g . In section 6 we give a brief discussion of weak
1
2 Preliminaries
We work primarily in the set f0 1g of all (innite, binary) sequences. We
1
also use the set f0 1g of all (nite, binary) strings. We write jxj for the
length of a string x, and for the empty string. The standard enumeration
of f0 1g is the sequence s0 s1 : : :, in which shorter strings precede longer
xm::n] for the string consisting of the mth through nth bits of x. In par-
ticular, x0::n ; 1] is the string consisting of the rst n bits of x. We write
xn] for xn::n], the nth bit of x.
5
We write '] for the Boolean value of a condition ', i.e.,
(
'] = 10 ifif '' isis true
false
The characteristic sequence of a set A N is then the sequence A 2 f0 1g 1
n o
DTIMEy (s) = x 2 f0 1g x DTIME(s) y :
1
T
(Note that the time bound here is \sharp" there is no \big-O.") The1 un-
relativized complexity class DTIME(s) is then dened to be DTIME0 (s),
where 0 is the sequence consisting entirely of 0's.
1
6
standard textbook denitions of truth-table reducibility 44, 50].) Given a
set Y f0 1g , we write
1
RECtt (Y ) = fx
2 f0 1g j (9y 2 Y ) x tt y g
1
= DTIMEy (s):
recursive s y Y
2
We write REC for the set of all recursive (i.e., decidable) sequences x 2
f0 1g . Note that REC Y RECtt (Y ) for all sets Y f0 1g . A
1 1
we write x
T y , if x T y and y T x. A Turing degree is an equivalence
class of f0 1g under the equivalence relation
T .
1
is small \in the sense of Baire category." This section reviews the basic ideas
from Lebesgue measure, resource-bounded measure, and Baire category that
are involved in our use of these three notions of \smallness." The interested
reader may consult 6, 18, 36, 37, 43, 45] for further discussion of these
notions, but the material in the present section is sucient for following the
arguments of this paper.
Resource-bounded measure 36, 37] is a generalization of classical
Lebesgue measure. As such it has classical Lebesgue measure and measure
in REC as special cases. We use this fact to present the notions \measure
0" and \measure 0 in REC" more or less simultaneously.
Consider the random experiment in which a binary sequence x 2 f0 1g 1
7
element of X . In this case, we write (X ) = 0 (\X has measure 0"). We
now develop the necessary denitions.
A string w 2 f0 1g is a prex of a string or sequence x 2 f0 1g
Cw = fx 2 f0 1g j w v xg
1
i.e., the set of all innite binary sequences beginning with the string w.
Denition 37]. A density function is a function d : f0 1g ! 0 1)
satisfying
d(w) = d(w0) +2 d(w1) (3:1)
for all w 2 f0 1g . The global value of a density function d is d(). The set
such that, for all ~k 2 Nn , the function d~k is a density function. (Recall that
d~k (w) = d(~k w) for all ~k 2 Nn and w 2 f0 1g .)
Taken together, parts (3.1) and (3.2) of the above denition imply that
Prx 2 S d]] d()
in our coin-tossing random experiment. We thus intuitively regard d as a
\detailed verication" that Prx 2 X ] d() for all X S d]. With this
intuition in mind, we present the central idea of resource-bounded measure
0 sets.
Denition 37]. A null cover of a set X f0 1g is a 1-DS d that satises
1
8
(ii) dk () 2 k .
;
of dyadic rationals. These are real numbers whose standard binary repre-
sentations are nite.
Denition 37]. An n-DS d is computable if there is a total recursive
function db : Nn+1 f0 1g ! D such that, for all ~k 2 Nn , r 2 N, and
w 2 f0 1g ,
db~kr (w) ; d~k (w) 2 r :
;
Note that the above denition is uniform, in the sense that it requires
a single total recursive function db to compute approximations for all the
density functions d~k (given ~k, a precision parameter r, and the input to d~k
as inputs to db).
Denition 37]. A recursive null cover of a set X f0 1g is a null cover
1
write (X j REC) = 1, if (X c j REC) = 0. In this latter case, we say that
X contains almost every recursive sequence x 2 REC.
Note that the implications
9
µrec(X)= 0 µrec (X)= 1
and
µ (X ) = 0 µ (X | REC ) = 0 µ (X ) = 1 µ (X | REC ) = 1
all follow immediately from the above denitions. It is easy to see that
every subset of a recursive measure 0 set has recursive measure 0, that every
nite subset of REC has recursive measure 0, and that every nite union
of recursive measure 0 sets has recursive measure 0. In fact, the recursive
measure 0 sets enjoy a stronger closure property, which we now dene.
Denition 37]. Let Z Z Z : : : f0 1g . Then Z is a recursive union
1
S
0 1
j =0
computable 2-DS d such that, for all j 2 N, dj is a recursive null cover of
Zj \ REC.
Theorem 3.1 (Lutz 37]). If Z f0 1g is a recursive union of sets of
1
k=0
dened by the following recursion.
(i) w0 = .
(ii) For all m 2 N, w2m+1 = m (w2m).
(iii) For all m 2 N, w2m+2 = m (w2m+1).
Intuitively, Player I uses strategy , Player II uses strategy , and wk is
the prex of R( ) that has been constructed when the two players have
moved a total of k times. For example, if and are dened by
m (w) = w0m+1 m (w) = w1
then
w0 = w1 = 0 w2 = 01 w3 = 0100 : : :
so
R( ) = 01001000100001000001 :
We write GX ] for the Banach-Mazur game with payo set X f0 1g . 1
A winning strategy for Player I in GX ] is a strategy such that, for all
strategies , R( ) 2 X . A winning strategy for Player II in GX ] is a
strategy such that, for all strategies , R( ) 2= X .
Denition. A set X f0 1g is meager if there exists a winning strat-
1
11
Then the strategy dened by m (w) = w1 is a winning strategy for Player
II in GFIN], so FIN is meager.
The proof that the above denition is equivalent to the \standard text-
book" denition of the meager sets is due to Banach and may be found in
42] or 43]. It is clear that every subset of a meager set is meager and that
every countable set X f0 1g is meager. In fact, it is well-known that
1
every countable union of meager sets is meager 43]. On the other hand, for
every w 2 f0 1g , the strategy
(
u <= w
m (u) = wu0 ifotherwise
6
is 0k+1 if X c is
0k+1 . (The \boldface" classes
01 01
02 02 : : : are col-
lectively known as the nite Borel hierarchy. This hierarchy is closely analo-
gous to the \lightface" arithmetical hierarchy 01 01 02 02 : : : of recursion
theory 42].)
A nite variation of a sequence x 2 f0 1g is a sequence y 2 f0 1g
1 1
such that y n] = xn] for all but nitely many n 2 N. A set X f0 1g is 1
12
Fact 3.3.
1. Let X and Y be disjoint subsets of f0 1g . If X is
02, Y 6=
, and
1
meager.
Proof. To prove part 1, assume the hypothesis and x a sequence z 2 Y .
Since X is
, there exist closed sets X X : : : f0 1g such that X =
0 1
k
k=0
for Player II in the Banach-Mazur game GX ]. Player II's strategy uses
z as a source of bits. To specify this strategy, let wk 2 f0 1g be the
string constructed by the game play prior to move k of Player II, where
k 2 N. Let wk ==z be the sequence obtained from z by putting wk in place
of the rst jwk j bits of z . Since z 2 Y and wk ==z is a nite variation of
z , it must be the case that wk ==z 2 Y . In particular, this implies that
wk ==z 2= Xk . Since Xk is closed, it follows that there exists n > jwk j such
that C(wk ==z)0::n 1] \ Xk =
. Player II's strategy in move k is to extend
;
k=0
Hausdor space f0 1g , so each Yk is compact. Since f is continuous, it
1
S
follows that each f (Yk ) is compact, hence closed. Since f (X ) = f (Yk ),
1
k=0
this implies that f (X ) is
02 . 2
We have described three notions of smallness in this section. It should
be noted that no two of them coincide. Although some sets (e.g. nite
sets) are small in all three senses, it is possible for a set to be small in any
one of these senses without being small in the other two. For example, in
section 4 below, we dene the set RAND, consisting of all algorithmically
13
random sequences. Consider also the set REC of all recursive sequences. It
is well-known 39] that REC \ RAND =
, that RAND is meager, and that
RAND has measure 1. (See also Theorems 4.7 and 6.2 below.) Also, since
REC is countable, REC is meager and has measure 0. The following three
things follow easily from these observations.
(a) RAND REC is meager, but has measure 1 and measure 1 in REC.
(b) RECc has measure 0 in REC but is comeager and has measure 1.
(c) RANDc has measure 0, but is comeager and has measure 1 in REC.
As Oxtoby 43] has noted, \There is of course nothing paradoxical in the
fact that a set that is small in one sense may be large in some other sense."
14
A Turing machine M starts in the initial state with a program 2 f0 1g
on its program tape, the output tape blank, and the worktapes blank. The
leftmost cell of the program tape is blank, with the program tape head
initially scanning this cell. The program lies immediately to the right of
this cell. The rest of the program tape is blank. The output tape head
initially scans the leftmost cell of the output tape.
If, after nitely many steps, M halts with the program tape head scan-
ning the last bit of , then the computation is deemed to be a success, we
write M ( )#, and the output of the computation is the string M ( ) 2 f0 1g
that has been written on the output tape. Otherwise, the computation is a
failure, we write M ( )", and there is no output (i.e., we disregard the con-
tents of the output tape). If M ( )#, then timeM ( ) denotes the number of
steps executed in this computation. If M ( )", then we write timeM ( ) = 1.
It should be emphasized that a successful computation must end with
the program tape head scanning the last bit of the program. Since the
program tape head is read-only and cannot move left, this implies that, for
every Turing machine M , the set
PROGM = f 2 f0 1g j M ( )#g
2PROGM
It is well-known that there are Turing machines U that are universal, in
the sense that, for every Turing machine M , there exists a program prex
M 2 f0 1g such that, for all 2 f0 1g ,
U (M ) = M ( ):
(This condition means that M ( ) # if and only if U (M ) #, in which case
U (M ) = M ().) Furthermore, there are universal Turing machines U
that are ecient, in the sense that, for each Turing machine M there is a
constant c 2 N (which depends on M ) such that, for all 2 f0 1g ,
15
Notational Convention. Throughout this paper, U is a xed, ecient,
universal Turing machine.
M is
PROGM (x) = f 2 f0 1g j M ( ) = xg :
X
Pr(I ) = 2 w : ;j j
w I
2
independent toss of aSfair coin to decide each bit of x, then Pr(I ) is the
probability that x 2 Cw , i.e., the probability that some element of I is
w I
a prex of x.
2
a Turing machine.
1. The (self-delimiting) Kolmogorov complexity of x relative to M is
n o
KM (x) = min jj 2 PROGM (x) :
(Here we use the convention that min
= 1.) The (self-delimiting)
Kolmogorov complexity of x is
K (x) = KU (x)
The quantity K (x) is also called the algorithmic entropy, or algorithmic
information content, of x.
16
2. The t-time-bounded (self-delimiting) Kolmogorov complexity of x rela-
tive to M is
n o
KMt (x) = min j j 2 PROGtM (x) :
The t-time-bounded (self-delimiting) Kolmogorov complexity, or t-time-
bounded algorithmic entropy, of x is
K t (x) = KUt (x):
3. The algorithmic probability of x relative to M is
mM (x) = Pr(PROGM (x)):
The algorithmic probability of x is
m(x) = mU (x):
4. The t-time-bounded algorithmic probability of x relative to M is
mtM (x) = Pr(PROGtM (x)):
The t-time-bounded algorithmic probability of x is
mt(x) = mtU (x):
In general, we omit the adjective \self-delimiting", since this is the only
type of Kolmogorov complexity in this paper.
We now present some basic properties of Kolmogorov complexity and
algorithmic probability that are used in this paper. The rst is obvious,
well-known, and useful.
Lemma 4.1. There is a constant c 2 N such that, for all x 2 f0 1g and
0
all 2 PROG(x),
K (x) K ( ) + c0:
The next two important theorems express the fundamental relationship
between Kolmogorov complexity and algorithmic probability.
Theorem 4.2 (Levin 26, 27], Chaitin 11]). There is a constant ec 2 N such
that, for all x 2 f0 1g ,
17
A straightforward modication of the proof of Theorem 4.2 yields the
following time-bounded version. (This result also follows immediately from
Lemma 3 of 34].)
Theorem 4.3. Let t : N ! N be recursive.
1. For all x 2 f0 1g ,
fashion.
Lemma 4.4 (Bennett 5]). There is a constant c 2 N such that, for all
0
(Note that ;jxj log Pr(I ) 0, so that the bound becomes tighter as Pr(I )
becomes smaller.)
Proof. Let M be a Turing machine that performs as indicated in Figure 1
with program b , where computes a nite instantaneous code and b 2
f0 1g . (If the program for M is not of this form, then the computation is
c = jM j + 1:
0
18
begin
simulate U ( ) to obtain I (on a worktape) in the form
I = fx0 : : : xn 1g, ;
:=
0
for 0 i < n do
begin
if i = 0 then w := 0ki else w := next(w ki),
where ki = jxi j ; b; log Pr(I )c and next(w ki) is the
immediate lexicographic successor of the string w1ki w
;j j
while v w do
0
19
and note that
nX1
;
X X
rn = 2 ki = 2
; b; I
log Pr( )c;jx Pr(I )
j ;1
2 x = 1:
;j j
i=0 x I
2 x I 2
Dene strings b : : : bn 1 2 f0 1g by
;
induction hypothesis, this means that bi does not consist entirely of 1's, so
bi+1 = next(bi ki+1) contains only ki+1 bits.) Moreover, it is easily checked
that, for all 0 i < n, bi is the value assigned to w by M during iteration
i of the for-loop, and that
U (M bi ) = M (bi ) = xi
whence
K (xi ) jM bi j = ki + j j + c ; 1 0
2
Corollary 4.5. For every recursive function t : N ! N there exists a
constant c 2 N such that, for all y 2 f0 1g and all 2 PROGt (y ),
K () jj + log mt(y ) + K (y ) + c :
20
Since U is a universal Turing machine, there is a program prex M 2
f0 1g such that, for all 2 f0 1g , U (M ) = M ( ). Let
c = jM j + c
0
K () jj + log Pr(PROGt (y )) + jM y j + c 0
= j j + log mt(y ) + K (y ) + c :
2
In this paper we are especially interested in the Kolmogorov complexi-
ties of initial segments of innite binary sequences. In this regard, given a
function g : N ! 0 1) and a recursive time bound t : N ! N, we dene
the classes
K : :< g(n)] = fx 2 f0 1g j K (x0::n ; 1]) < g(n) i.o.g
io
1
and
n o
Kt: :< g(n)] = x 2 f0 1g j K t(x0::n ; 1]) < g(n) i.o. :
io
1
Thus we are using g (n) as a \threshold value" for the Kolmogorov complexity
of the n-bit prex of a sequence x 2 f0 1g . These classes contain those
1
sequences for which this Kolmogorov complexity is below the threshold value
for innitely many prexes.
The following theorem, which is used in proving our main result, says
that almost every recursive sequence has very high time-bounded
Kolmogorov complexity almost everywhere.
Theorem 4.6 (Lutz 37]). For every recursive bound t : N ! N and every
real number 0 <
< 1,
(Kti:o: <
n] REC) = 0:
(In fact, Corollary 4.9 of 37] is stronger than this in several respects.)
21
We conclude this section with a brief discussion of the algorithmic ran-
domness of innite binary sequences. Algorithmic randomness was origi-
nally dened by Martin-Lof 39], using constructive versions of ideas from
measure theory. Subsequently, Levin 26, 27], Schnorr 47], and Chaitin
11] showed that algorithmic randomness could be characterized in terms
of self-delimiting Kolmogorov complexity. (Indeed, this was an important
motivation for developing the self-delimiting formulation.) For the purposes
of the present paper, it is convenient to use this characterization as the
denition.
Denition. A sequence x 2 f0 1g is algorithmically random, and we write
1
and
\ t
1 1
Dtg = Dg (n)
m=0 n=m
= fx 2 f0 1g x 2 Dtg (n) a:e:g:
1
22
A sequence x 2 f0 1g is strongly deep, and we write x 2 strDEEP, if for
1
gram for x0::n ; 1] can be compressed by at least g (n) bits. Note that, if
t(n) tb(n) and g (n) gb(n), then Dbbtg (n) Dtg (n). Thus, if t(n) bt(n) a:e:
and g (n) gb(n) a:e:, then Dbbtg Dtg . In particular, if g (n) = c and gb(n) = cb
are constant, then we have the situation depicted in Figure 2.
{0,1}∞
t
Dc
t
D
c
strDEEP
Figure 2: The classes Dtc , Dbbtc , in the case where t(n) bt(n) a.e. and c cb.
23
numbers and a sequence 0 1 : : : of programs satisfying the following three
conditions for all i 2 N.
(1) For all n ki , K (x0::ki ; 1]) ; ki K (x0::n ; 1]) ; n.
(2) U (i) = x0::ki ; 1] and jij = K (x0::ki ; 1]).
(3) ki+1 > ki + timeU (i ).
Proof. Let x 2 RAND. Dene f : N ! Z by f (n) = K (x0::n ; 1]) ; n. For
each i 2 N, x the least argument ni i such that f (ni ) f (n) for all n i.
(Since x 2 RAND, f is bounded below, so ni exists.) Dene the sequences
k0 k1 : : : and 0 1 : : : recursively as follows. Let k0 = n0 and let 0 be a
minimal program for x0::k0 ; 1]. Given ki and i , let ki+1 = nki +timeU (i )+1
and let i+1 be a minimal program for x0::ki+1 ; 1]. It is easily veried
that the sequences k0 k1 : : : and 0 1 : : : satisfy conditions (1), (2), and
(3). 2
Bennett 5] has noted that no algorithmically random sequence is
strongly deep. We now prove this fact. Moreover, we show that it holds
in a very strong way. Intuitively, we show that every algorithmically ran-
dom sequence lies \very near the top" of the diagram in Figure 2.
Theorem 5.2 (Bennett 5]). RAND \ strDEEP =
. In fact, there exist
a recursive function t(n) = O(n log n) and a constant c 2 N such that
RAND \ Dtc =
.
Proof. Let M be a Turing machine that, with program y, does the follow-
ing. The machine M simulates U ( ), recording timeU ( ) while doing so.
If the simulated computation succeeds, M then reads and outputs the rst
timeU () bits of y (appended to the string U ( ) already produced as out-
put) and halts. Note that if jy j = timeU ( ), then the computation of M (y )
succeeds, with M (y ) = U ( )y . Otherwise, the computation of M (y ) is a
failure.
On successful computations, the Turing machine M takes O(jy j) steps
to produce U ( )y . Thus there exist a program prex M and a recursive,
nondecreasing time bound t(n) = O(n log n) such that, for all successful
computations U ( ) and all strings y with jy j = timeU ( ), the following two
conditions hold.
(i) U (M y ) = U ( )y .
24
(ii) timeU (M y ) t(jy j).
Let c = jM j + c0, where c0 is the constant from Lemma 4.1. We prove that
RAND \ Dtc =
.
Let x 2 RAND. Fix sequences k0 k1 : : : and 0 1 : : : as in Lemma 5.1.
For each i 2 N, let ni = ki + timeU (i ). Note that the sequence n0 n1 : : :
is strictly increasing. We prove that x 62 Dtc by showing that, for all i 2 N,
x 62 Dtc(ni).
Conditions (i) and (ii) above imply that the following conditions hold
for all i 2 N.
(iii) U (M i xki::ni ; 1]) = x0::ni ; 1].
(iv) timeU (M i xki::ni ; 1]) t(ni ; ki ) t(ni ).
Then, for all i 2 N
M ixki::ni ; 1] 2 PROGt (x0::ni ; 1])
and Lemma 5.1 tells us that
K (x0::ki ; 1]) K (x0::ni ; 1]) ; ni + ki
= K (x0::ni ; 1]) ; timeU (i )
whence
K (M i xki ::ni ; 1]) K (x0::ni ; 1]) ; c0
K (x0::ki ; 1]) + timeU (i ) ; c0
= ji j + ni ; ki ; c0
= ji xki::ni ; 1]j ; c0
= jM i xki::ni ; 1]j ; c:
Thus x 62 Dtc (ni ) for all i 2 N, so x 62 Dtc . 2
We next show that strong computational depth can be characterized in
several equivalent ways. For this, we need some notation and a lemma. We
rst recall Bennett's denition of the computational depth of nite strings.
Denition 5]. Let w 2 f0 1g and c 2 N. Then the computational depth
of w at signicance level c is
depthc (w) = minft 2 N j (9 2 PROGt (w)) j j < K ( ) + cg:
25
That is, the depth of a nite string at signicance level c is the minimum
time required to compute w from a program that is not compressible by c
or more bits.
Our alternate characterizations of strong depth also use the following
classes.
Denition. For t g : N ! N and n 2 N, we dene the sets
\
Db tg = Db tg (n)
1 1
m=0 n=m
\ et
De tg =
1 1
Dg (n):
m=0 n=m
The following lemma shows that the classes Db tg and De tg are, in a quanti-
tative sense, \minor variants" of the classes Dtg . This result was proven in
a slightly dierent form in 5].
Lemma 5.3 (Bennett 5]). If t : N ! N is recursive, then there exist
constants c c c 2 N and a recursive function t : N ! N such that the
0 1 2 1
following six conditions hold for all g : N ! N and all n 2 N.
1. Dtg+c0 (n) Db tg (n) 4. Dtg+c0 Db tg
2. Db tg1+c1 (n) De tg (n) 5. Db tg1+c1 De tg
3. De tg+c2 (n) Dtg (n) 6. De tg+c2 Dtg
Proof. It suces to prove 1, 2, and 3, since 4, 5, and 6 then follow imme-
diately.
1. Let c0 be as in Lemma 4.1 and assume that x 2 Dtg+c0 (n). Let
be a shortest element of PROGt (x0::n ; 1]). Since x 2 Dtg+c0 (n), we have
K () jj ; g(n) ; c0 . It follows that
K (x0::n ; 1]) K () + c0
j j ; g (n)
= K t (x0::n ; 1]) ; g (n)
26
whence x 2 Db tg (n).
2. Choose c1 and t1 for t as in Theorem 4.3 and assume that x 2
bDtg1+c1 (n). Then K (x0::n ; 1]) K t1 (x0::n ; 1]) ; g (n) ; c1. It follows by
Theorems 4.2 and 4.3 that
m(x0::n ; 1]) 2 K (x0::n 1])
; ;
t
2g(n)+c1 K 1 (x0::n 1])
; ;
j j ; g (n)
whence x 2 Dtg (n). 2
We now prove the equivalence of several characterizations of strong com-
putational depth.
Theorem 5.4 (Bennett 5]). For x 2 f0 1g , the following four conditions
1
are equivalent.
(1) x is strongly deep.
(2) For every recursive time bound t : N ! N and every constant c 2 N,
depthc (x0::n ; 1]) > t(n) a:e:
(3) For every recursive time bound t : N ! N and every constant c 2 N,
x 2 Db tc.
(4) For every recursive time bound t : N ! N and every constant c 2 N,
x 2 De tc.
27
Proof. The equivalence of (1) and (2) follows immediately from the deni-
tions. The equivalence of (1), (3), and (4) follows immediately from Lemma
5.3. 2
In 5], Bennett uses condition (2) of Theorem 5.4 above as the denition
of strong computational depth. As noted above, this is trivially equivalent
to condition (1), i.e., to our denition in terms of the classes Dtc . Bennett
5] also considers denitions in terms similar to those used in dening the
classes Db tc and De tc and implicitly proves the equivalence of conditions (1),
(3), and (4). The discussions of depth by Li and Vitanyi in the Handbook of
Theoretical Computer Science 33] and their recent book 35] essentially use
condition (4) as the denition. In any case, a sequence x is strongly deep if,
for every recursive t and constant c, almost every prex x0::n ; 1] is \more
than t deep at signicance level c," in the sense that more than t(n) time is
required to derive x0::n ; 1] from any description whose length is within c
bits of the minimum possible length.
We next prove a technical lemma on the quantitative relationship be-
tween computational depth and time-bounded Turing reducibility. This can
be regarded as a quantitative, innitary version of Bennett's deterministic
slow-growth law 5]. We need two special notations for this lemma. First,
for any function s : N ! N, we dene the function s : N ! N by
bt = cM (1 + (n)dlog (n)e)
gb = g(s (n + 1)) + cM
U (M 0 ) = M ( ) 0
and
timeU (M 0 ) cM 0 (1 + timeM 0 () log timeM 0 ()):
Let M be a Turing machine that, with program 2 f0 1g , simulates
00
U (M 00 ) = M ( ). Let
00
cM = max fcM 0 jM 0 j + jM 00 jg :
Fix m0 2 N such that (s ) 1(m) > 0 for all m m0 .
;
Now dene , bt, and gb as in the statement of the lemma and assume that
x y 2 f0 1g satisfy y DTIME( s) x via M and y 2 Dbt . Fix n 2 N such
bg
1
T 0
b
that y 2 Dbtg (n) for all n n0 and let
m1 = max fm0 s (n0) + 1g :
The following two claims are veried at the end of this proof.
Claim 1. For all m m and 2 f0 1g , if 2 PROGt(x0::m ; 1]),
0
K () j j ; gb(n) + cM
29
begin
u := U ()
n := (s ) 1(juj)
;
for 0 i < n do
append the bit M u01 (i) to the output
halt
end M ().
0
where n = (s ) 1 (m).
;
j j ; g (m):
Thus x 2 Dtg (m). Since this holds for all m m1 , it follows that x 2 Dtg ,
arming the lemma. All that remains, then, is to prove the two claims.
To prove Claim 1, assume that m m0 and 2 PROGt (x0::m;1]). Let
u = x0::m ; 1] and n = (s ) 1(m). Since m m0, we
;
must have s (n) < m.
all 0 i < n. (All queries in these computations must be made to bits xj ]
for j < juj.) Thus
U (M 0 ) = M () = y 0::n ; 1]:
0
timeU (M 0 ) bt(n):
Thus M 0 2 PROGbt (y 0::n ; 1]). This proves Claim 1.
30
Finally, to prove Claim 2, let m m1, let 2 PROGt (x0::m ; 1]),
and let n = (s ) 1 (m). Since m > s (n0 ), it must be the case that n =
;
U (M 00 ) = M ( ) =
00
so
K () jM 00 j
= K (M 0 ) + jM 00 j
j j ; gb(n) + cM :
This proves Claim 2 and completes the proof of Lemma 5.5
2
Using Lemma 5.5, we prove that a strongly deep sequence cannot be
truth-table reducible (equivalently, reducible in recursively bounded time)
to a sequence that is not also strongly deep. This implies the fact, noted by
Bennett 5], that strong depth is invariant under truth-table equivalence.
Theorem 5.6. Let x y 2 f0 1g . If y x and y is strongly deep, then x
1
tt
is strongly deep.
Proof. Assume the hypothesis. To see that x is strongly deep, x a recursive
function t : N ! N and a constant c 2 N. It suces to prove that x 2 Dtc .
Since y tt x, there exist a strictly increasing time-constructible function
s : N ! N and an s-time-bounded oracle Turing machine M such that
y DTIME( s) x via M . Choose a constant c for M as in Lemma 5.5 and
M
dene g : N ! N by g (n) = c for all n 2 N. Then, in the notation of
T
time bound s : N ! N such that DTIMEx (s) does not have measure 0 in
REC.
That is, x is weakly useful if it can be used to \eciently" (i.e., in some
recursive time s) solve all the problems in a non-negligible subset of REC.
If x 2 REC, then for every recursive time bound s, there is a recur-
sive time bound t such that DTIMEx (s) DTIME(t): Since every such
set DTIME(t) has measure 0 in REC by Theorem 4.6, this shows that no
recursive sequence is weakly useful.
The following result, which is the main theorem of this paper, shows
that much more is true.
Theorem 5.11. Every weakly useful sequence is strongly deep.
Proof. Let x 2 f0 1g be weakly useful. To see that x is strongly deep,
1
33
function is bounded above by a strictly increasing, time-constructible func-
tion, we can assume without loss of generality that s is strictly increasing
and time-constructible.
Let et(n) = n (1 + (n)dlog (n)e), where is dened from t and s as
in Lemma 5.5, and let = 12 . Since te is recursive, Corollary 5.9 tells us
that Detn has measure 1 in REC. Since DTIMEx (s) does not have measure
0 in REC, it follows that Detn \ DTIMEx (s) 6=
. Fix a sequence y 2
Detn \ DTIMEx (s). Then there is an s-time-bounded oracle Turing machine
M such that y DTIME( s) x. Fix a constant c for M as in Lemma 5.5.
M
Dene g (n) = c for all n 2 N and dene the functions bt, and gb from t and
T
g as in Lemma 5.5. Since gb and cM are constant, we have te(n) > tb(n) a.e.
and n > gb(n) a.e., so y 2 Detn Dbbtg . It follows by Lemma 5.5 that x 2 Dtc .
2
Notation. Let H and K be the characteristic sequences of the halting
problem and the diagonal halting problem, respectively. That is, the se-
quences H K 2 f0 1g are dened by
1
34
x 2 f0 1g is recursively enumerable (r.e.) if x = A for some r.e. set
1
the set
K x = fn 2 N M x (n) haltsg
n
where Mn is the nth oracle Turing machine in a standard enumeration. The
jump of a sequence x 2 f0 1g is the sequence
1
jump(x) = Kx :
A sequence x 2 f0 1g is high if x T K and jump(x)
T jump(K ). A
1
(i) y T x.
(ii) X fyk j k 2 Ng, where each yk 2 f0 1g is dened by yk n] =
1
yhk ni] for all; n 2 N . (Here we are using the standard pairing func-
k +n+1
tion hk ni = 2 + n:)
We use the following two known facts.
Theorem 5.13 (Sacks 46]). There exist r.e. sequences that are high and
not Turing equivalent to K .
Theorem 5.14 (Martin 38]). A sequence y 2 f0 1g satises jump(K )
1
T
jump(y) if and only if there exists x
T y such that REC is uniformly re-
cursive in x.
Corollary 5.15. Every high Turing degree contains a strongly deep se-
quence.
Proof. The key observation, pointed out to the third author by Stuart
Kurtz, is that every high Turing degree contains a weakly useful sequence.
To see this, let a be a high Turing degree. By Theorem 5.14, there is a
sequence x 2 a such that REC is uniformly recursive in x. Then there is a
sequence y T x such that REC fyk j k 2 Ng. Dene z 2 f0 1g by 1
( k
z k] = yx k2 ] 1 ] ifif kk isis even
;
odd
2
35
Then z
T x, so z 2 a. Also, there is a constant c 2 N such that
REC fyk j k 2 Ng DTIMEz (cn2 + c)
so z is weakly useful. This conrms that every high Turing degree contains a
weakly useful sequence. By Theorem 5.11, the corollary follows immediately.
2
Taken together, Theorem 5.13 and Corollary 5.15 show that Theorem
5.11 does indeed strengthen Bennett's result, Corollary 5.12.
We conclude this section by proving that strongly deep sequences are
extremely rare, both in the sense of Lebesgue measure and in the sense of
Baire category.
Theorem 5.16. The set strDEEP is meager and has measure 0. In fact, if
t and c are as in Theorem 5.2, then Dtc is meager and has measure 0.
Proof. Let t and c be as in Theorem 5.2. Then RAND \ Dtc =
. Since
RAND has measure 1, it follows that Dtc has measure 0.
For each n 2 N, the complement of Dtc (n) can be written as a (nite)
union of cylinders Cw , with each jwj = n. (This is because membership or
nonmembership of a sequence x in Dtc (n) depends only upon x0::n ; 1].)
Thus, for each n 2 N, the set Dtc (n) is closed. It follows that, for each
m 2 N, the set T Dtc (n) is closed, whence the set Dtc = S T Dtc(n)
1 1 1
If t and c are as in Theorem 5.2, then it is not dicult to show that FIN \
Dtc =
. It follows that Player II could use the sequence 0 in place of z in
1
the above strategy. That is, Player II could win by playing an appropriate
number of 0's, instead of random bits, during each turn.
36
6 Weak Computational Depth
In Theorem 5.16, we saw that strongly deep sequences are very rare, both
in the sense of Lebesgue measure and in the sense of Baire category. In
this brief section, we show that the situation is dierent for weakly deep
sequences. We rst recall the denition.
Denition (Bennett 5]). A sequence x 2 f0 1g is weakly deep, and we
1
We thus have
wkDEEP = RECtt(RAND)c :
Since REC RAND RECtt(RAND), it follows immediately that
wkDEEP \ REC = wkDEEP \ RAND =
i.e., that no weakly deep sequence can be recursive or algorithmically ran-
dom.
As the terminology suggests, every strongly deep sequence is weakly
deep.
Theorem 6.1 (Bennett 5]). strDEEP wkDEEP.
Proof. Assume that x 2 strDEEP and x y. To see that x 2 wkDEEP,
tt
it suces to show that y 62 RAND. But this follows immediately from
Theorems 5.2 and 5.6. 2
In particular, Theorems 5.11 and 6.1 imply that weakly deep sequences
exist. It should be noted that Gacs 16] has proven that, for every sequence
x 2 f0 1g , there exists a sequence z 2 RAND such that x T z . Thus
1
comeager. 2
Bennett 5] noted that there exist sequences that are weakly deep, but
not strongly deep. The following corollary shows that such sequences are,
in the sense of Baire category, commonplace.
Corollary 6.3. The set wkDEEP ; strDEEP is comeager.
Proof. This follows immediately from Theorems 5.16 and 6.2. 2
Thus, in the sense of Baire category, almost every sequence x 2 f0 1g1
Proof. This follows immediately from Theorem 6.1 and Corollary 6.2. 2
Figure 4 summarizes the relationships among REC, RAND, wkDEEP,
and strDEEP. In the sense of Lebesgue measure, almost every binary se-
quence is in RAND. On the other hand, in the sense of Baire category,
almost every binary sequence is in wkDEEP ; strDEEP.
38
RAND
REC tt (RAND)
REC wkDEEP
strDEEP
39
7 Conclusion
We have shown that every weakly useful sequence is strongly deep. This
result generalizes Bennett's observation that K is strongly deep, and gives
support to Bennett's thesis that the computational usefulness of K is re-
lated to its computational depth. We mention two open questions that are
suggested by this result.
Recall that a sequence x 2 f0 1g is weakly useful if there is a recursive
1
time bound s : N ! N such that DTIMEx (s) does not have measure 0
in REC. Dene a sequence x 2 f0 1g to be strongly useful if there is a
1
jump(n)(x)
T jump(n) (K ), where jump(n) is the n-fold iteration of the
jump operation. A Turing degree a is then highn if it contains a highn
sequence. (See 50], for example.) If a sequence or degree is highn , then it
is clearly highn+1 . The Turing degree of K is clearly the only high0 degree.
It is also clear that a sequence or degree is high1 if and only if it is high.
Thus, by Corollary 5.15, every high1 Turing degree contains a strongly deep
sequence.
Question 7.2. For n > 1, is it necessarily the case that every highn Turing
degree contains a strongly deep sequence?
Answers to Question 7.1 and 7.2 may well improve our understanding
of computational depth vis-a-vis computational usefulness. More generally,
further investigation of Bennett's fundamental notions may yield profound
insights into the role of depth in the organization of computational, physical,
and biological information.
Acknowledgments
The third author thanks Charles Bennett for several helpful discussions,
and Stuart Kurtz for pointing out Theorem 5.14. We also thank Ron Book,
40
Josep D"az, and two anonymous referees for suggestions that have improved
the exposition of this paper.
References
1] L. Adleman. Time, space, and randomness. Technical Report
MIT/LCS/79/TM-131, Massachusettes Institute of Technology, Lab-
oratory for Computer Science, March 1979.
2] J. L. Balcazar, J. D"az, and J. Gabarro. Structural Complexity I.
Springer-Verlag, 1988.
3] Y. M. Barzdin . Complexity of programs to determine whether natural
0
42
25] M. Koppel. Structure. In R. Herken, editor, The Universal Turing
Machine: A Half-Century Survey, pages 435{452. Oxford University
Press, 1988.
26] L. A. Levin. On the notion of a random sequence. Soviet Mathematics
Doklady, 14:1413{1416, 1973.
27] L. A. Levin. Laws of information conservation (nongrowth) and as-
pects of the foundation of probability theory. Problems of Information
Transmission, 10:206{210, 1974.
28] L. A. Levin. On the principle of conservation of information in intu-
itionistic mathematics. Soviet Mathematics Doklady, 17:601{605, 1976.
29] L. A. Levin. Uniform tests of randomness. Soviet Mathematics Doklady,
pages 337{340, 1976.
30] L. A. Levin. Various measures of complexity for nite objects (ax-
iomatic description). Soviet Mathematics Doklady, 17:522{526, 1976.
31] L. A. Levin. Randomness conservation inequalities information and in-
dependence in mathematical theories. Information and Control, 61:15{
37, 1984.
32] L. A. Levin and V. V. V'jugin. Invariant properties of informational
bulks. Proceedings of the Sixth Symposium on Mathematical Founda-
tions of Computer Science, pages 359{364, 1977.
33] M. Li and P. M. B. Vitanyi. Kolmogorov complexity and its applica-
tions. In J. van Leeuwen, editor, Handbook of Theoretical Computer
Science, Volume A, pages 187{254. Elsevier, 1990.
34] M. Li and P. M. B. Vitanyi. Learning simple concepts under simple
distributions. SIAM Journal on Computing, 20:911{935, 1991.
35] M. Li and P. M. B. Vitanyi. An Introduction to Kolmogorov Complexity
and its Applications. Springer, 1993.
36] J. H. Lutz. Resource-bounded measure. in preparation.
37] J. H. Lutz. Almost everywhere high nonuniform complexity. Journal
of Computer and System Sciences, 44:220{258, 1992.
43
38] D. A. Martin. Classes of recursively enumerable sets and degrees of
unsolvability. Z. Math. Logik Grundlag. Math., 12:295{310, 1966.
39] P. Martin-Lof. On the denition of random sequences. Information and
Control, 9:602{619, 1966.
40] P. Martin-Lof. Complexity oscillations in innite binary sequences.
Zeitschrift fur Wahrscheinlichkeitstheory und Verwandte Gebiete,
19:225{230, 1971.
41] K. Mehlhorn. The \almost all" theory of subrecursive degrees is decid-
able. In Proceedings of the Second Colloquium on Automata, Languages,
and Programming, pages 317{325. Springer Lecture Notes in Computer
Science, vol. 14, 1974.
42] Y. N. Moschovakis. Descriptive Set Theory. North-Holland, 1980.
43] J. C. Oxtoby. Measure and Category. Springer-Verlag, 1980. second
edition.
44] H. Rogers, Jr. Theory of Recursive Functions and Eective Computabil-
ity. McGraw - Hill, 1967.
45] H. L. Royden. Real Analysis, third edition. Macmillan Publishing
Company, 1988.
46] G. E. Sacks. Degrees of Unsolvability. Princeton University Press, 1966.
47] C. P. Schnorr. Process complexity and eective random tests. Journal
of Computer and System Sciences, 7:376{388, 1973.
48] A. Kh. Shen . The frequency approach to dening a random sequence.
0
45