1: %3.02.03
2: \documentclass{article}
3: \usepackage{amstext,amsmath,amsthm,latexsym}
4: \setlength{\textwidth}{6.5in}
5: \setlength{\textheight}{9.3in}
6: \setlength{\oddsidemargin}{0in}
7: \setlength{\evensidemargin}{0in}
8: \setlength{\topmargin}{-0.1in}
9: \setlength{\headheight}{0in}
10: \setlength{\headsep}{0in}
11: \setlength{\footskip}{0.5in}
12: \renewcommand{\arraystretch}{1.2}
13: \newcommand{\pair}[1]{\langle #1 \rangle}
14: \newcommand{\CC}{CC}
15: \newcommand{\PCC}{PCC}
16: \newcommand{\TCC}{TCC}
17:
18: \newtheorem{definition}{Definition}
19: \newtheorem{observation}{Observation}
20: \newtheorem{claim}{Claim}
21: \newtheorem{corollary}{Corollary}
22: \newtheorem{theorem}{Theorem}
23: \newtheorem{lemma}{Lemma}
24: \newtheorem{fact}{Fact}
25: \newtheorem{conjecture}{Conjecture}
26: \newtheorem{proposition}{Proposition}
27: \newtheorem{rem}{\sc Remark}
28: \newenvironment{remark}{\begin{rem}}{\end{rem}}
29:
30: \newtheorem{ex}{\sc Example}
31: %\newenvironment{proof}{\par \sf Proof.\rm}{\hspace*{\fill}$\Box$\vspace{1ex}}
32: \newenvironment{example}{\begin{ex}}{\hspace*{\fill}$\Diamond$\end{ex}}
33:
34:
35: \date{}
36: \begin{document}
37:
38: %% \pagestyle{headings}
39: %In order to omit page numbers and running heads
40: %please change this line to
41: %\pagestyle{empty}
42: %and change the first command line too, see above.
43:
44:
45: \title{Individual Communication Complexity}
46:
47:
48: \author{Harry Buhrman
49: \thanks{
50: CWI and University of Amsterdam}
51: \and
52: Hartmut Klauck
53: \thanks{IAS, Princeton}
54: \and
55: Nikolai Vereshchagin
56: \thanks{Moscow State University}
57: \and
58: Paul Vit\'{a}nyi
59: \thanks{CWI and University of Amsterdam}
60: }
61:
62:
63:
64: \maketitle
65:
66:
67: \begin{abstract}
68: We initiate the theory of communication complexity of individual
69: inputs held by the agents, rather than worst-case
70: or average-case. We consider total, partial, and partially correct
71: protocols, one-way versus two-way, with and without help bits.
72: \end{abstract}
73:
74:
75: \section{Introduction}
76: Assume Alice has input $x$ and Bob has input $y$ and they want
77: to compute a function $f(x,y)$ by communicating information
78: and local computation according to a fixed protocol $P = (P_A,P_B)$
79: where $P_A$ is the protocol executed by Alice, and $P_B$
80: is the protocol executed by Bob. For definiteness assume that
81: the requirement is that Alice outputs $f(x,y)$.
82: We are only interested in minimizing
83: the number of bits communicated between Alice and Bob as in
84: \cite{Ya79,KN97}.
85: In the usual setting one considers the worst-case or average-case over all
86: inputs $x,y$ of given length $n$. However, in current
87: situations like replicated file systems, and cache coherence algorithms,
88: in multiprocessor systems and computer networks, the
89: worst-case or average-case are not necessarily significant.
90: The files or updates can be very large; but in real life
91: they may typically be non-random
92: and have considerable
93: regularities or correlations that allow the communicated information
94: to be greatly compressible.
95: Neither the worst-case not the
96: average-case may be relevant; one wants to analyze the individual case.
97: This gives also much more information: from the individual
98: case-analysis one can easily derive the worst-case and the average-case,
99: but not the other way around. Indeed, certain phenomena have
100: no counterpart in more traditional settings: For example,
101: there are inputs for Bob such
102: that irrespective of Alice's input,
103: every ``simple'' total protocol requires arbitrarily higher
104: communication complexity than some more ``complex'' total protocol.
105: %Taking into consideration
106: %the allowed time to compress and
107: %uncompress, the (time-limited) Kolmogorov
108: %complexity \cite{LiVi97} measures the optimal number of bits
109: %involved.
110: Our results are expressed in terms of Kolmogorov
111: complexity \cite{LiVi97}, the minimal number of bits from which the data
112: can be decompressed by effective computation. We use the ``plain''
113: Kolmogorov complexity denoted as $C(x), C(x|y), C(x,y)$
114: for the absolute complexity of $x$, the conditional complexity of $x$ given $y$,
115: and the joint complexity of $x,y$.
116: Increased compression of the data approximates the
117: Kolmogorov complexity more and more, but the actual value is uncomputable
118: in general.
119: Given $x,y$, and assuming that Alice and Bob
120: have a protocol $P$ that works correctly on
121: $x,y$,
122: we study the
123: %begin changes
124: {\em individual communication complexity} $\CC^P(x,y)$
125: defined as the
126: number of bits Alice with input $x$ and Bob with input $y$
127: exchange using protocol $P$.
128: %{\em individual communication complexity} $\TCC_f^P(x,y)$
129: %defined as the minimum
130: %number of bits Alice with input $x$ and Bob with input $y$
131: %need to exchange to compute $f(x,y)$ using protocol $P$ as defined
132: %below.
133: %end changes
134: We refer to a standard definition of communication protocol \cite{KN97}.
135: We assume that the protocol identifies the length $n$ of the strings
136: on which it works. By the complexity of a protocol $P$
137: we mean its plain Kolmogorov complexity conditional to $n$,
138: denoted as $C(P|n)$.
139:
140: {\bf Results and Related Work:}
141: We use the framework of communication complexity as in \cite{Ya79,KN97}.
142: As far as we are aware there is no previous work on individual communication
143: complexity.
144: We formulate a theory of individual communication
145: complexity, and first analyze the "mother" problem, the indentity function,
146: where Alice outputs the input
147: of Bob.
148: We look at special functions such as the inner product, random functions,
149: and the equality
150: function.
151: We then turn to the question of analyzing the communication complexity,
152: with respect to the best protocol of given complexity,
153: for the mother problem (identity function).
154: For total protocols that are always correct, the power of one-way protocols
155: equals that of two-way protocols, but for partially correct protocols
156: or partial protocols, two-way protocols are remarkably more powerful.
157: We establish a relation with Kolmogorov's Structure function,
158: and the existence of strange ``non-communicable'' inputs of possibly
159: low Kolmogorov complexity for total
160: protocols---for which
161: the communication complexity of every total protocol
162: is necessarily very large (almost the literal uncompressed input
163: needs to be communicated) unless all
164: of the input is hard-wired in the protocol.
165: It is shown that for partial protocols
166: two-way is more powerful than one-way when we use help
167: bits.
168:
169:
170: \section{The mother function: Identity}
171: We start with listing some easy facts that establish lower and upper
172: bounds on individual communication
173: complexity with respect to individual protocols $P$
174: expressed in terms of $C(y|n)$, $C(y|P)$ and compared to
175: $C(y|x)$.
176: We assume that the protocols do not depend on $x,y$,
177: they are uniform, and they compute
178: the function concerned on strings of length $n$.
179: Let $C$ be a constant such that $C(y|n)\le n+C$
180: for all $y$.
181: %begin change Paul:
182: %The following is not even applied in (1), so not consistent.
183: %The Kolmogorov complexity will always be conditional to
184: %$n$, so in this subsection the expression $C(\cdot)$
185: %will always mean $C(\cdot|n)$.
186: %end change
187:
188: Let $I(x,y)=y$ be the identity function: Alice with input $x$ and
189: Bob with input $y$ compute output $y$ by Alice.
190: This is the ``mother'' function:
191: for if Alice can compute
192: $I$ then she can compute every computable function $f$.
193:
194: (1) For all $n$ there is a protocol $P$ of complexity $n+O(1)$ to compute
195: the identity function such that for all $x,y$ of length $n$ we have
196: $\CC_I^P(x,y)\le C(y|n)$.\\
197: Indeed, assume Bob knows $L_n=|\{p\mid |p|\le n+C,\ U(p)\text{ halts}\}|$.
198: ($U$ is the reference universal Turing machine.)
199: Then Bob can find all halting programs of length at most
200: $n+C$ by enumerating them until he obtains $L_n$ halting programs.
201: This allows him to find a shortest program $y^*$ for $y$.
202: He transmits that program to Alice and Alice computes $y$.
203: The complexity of this protocol is $C(L_n)+O(1)=n+O(1)$.
204:
205: (2) The complexity bound $n+O(1)$ on $C(P|n)$ in item (1) is tight.
206: For every protocol of complexity less than $n$ the assertion of item (1) is false:
207: for all $P$ there are $x,y$ such that
208: $\CC_I^P(x,y)\ge n$ but $C(y|P)=O(1)$ (and hence $C(y|n)\le C(P|n)+O(1)$, that
209: is $C(y|n)$ is much smaller than $\CC_I^P(x,y)$ if $C(P|n)$ is much smaller than $n$).\\
210: Indeed, let $\epsilon$ be the empty string and let $y$ be the first string such that
211: $\CC_I^P(\epsilon,y)\ge n$ (by counting arguments there is such $y$).
212:
213: (3) For every protocol $P$ to compute
214: identity function and for every $x,y$ we have
215: $\CC_I^P(x,y)\ge C(y|P)-O(1)$.\\
216: Let $c$ be the conversation between Alice and Bob on inputs $x,y$.
217: It suffices to prove that given $P,c$ we can find $y$.
218: It is known \cite{KN97} that the set of all pairs $(x',y')$ such that
219: the conversation between Alice and Bob
220: on input $(x',y')$ is equal to $c$ is a rectangle, that is, has the form $X\times Y$,
221: for some $X,Y\subset\{0,1\}^n$.
222: The set $Y$ is a one-element set, as for every $y'\in Y$ Alice
223: outputs $y$ also on the input $(x,y')$ (the output of Alice depends on $c,P,x$ only).
224: We can find $Y$ given $P,c$ and since $Y=\{y\}$ we are done.
225:
226: By item (2), for every protocol there are
227: $x,y$ such that the right hand side of the inequality
228: $\CC_I^P(x,y)\ge C(y|P)-O(1)$
229: is much less than its left hand side, more specifically,
230: $C(y|P)=O(1)$ and $\CC_I^P(x,y)\ge n$.
231:
232: (4) How is $\CC_I^P(x,y)$ related to $C(y|x)$?
233: By item (3) we have $\CC_I^P(x,y)\ge C(y|x)-C(P)-O(\log C(P))$ for all $x,y$.
234: For all $P$ this inequality is not tight for some $x,y$: there are $x,y$
235: such that $C(y|x)=O(1)$ but $\CC_I^P(x,y)\ge n$.\\
236: Indeed, let $x=y$. We need to prove that for some $x$ it holds
237: $\CC_I^P(x,x)\ge n$. For every $x$ let $c(x)$ denote the conversation
238: on the pair $(x,x)$. For every $x$ the set of input pairs $(x',y')$
239: producing the conversation $c(x)$ is a rectangle of height 1, as we have seen in
240: item (3). Therefore $c(x)$ are pairwise different for different
241: $x$ hence for some $x$ we have $|c(x)|\ge n$.
242:
243:
244: (5) However, for some $P,x,y$ the inequality
245: $\CC_I^P(x,y)\ge C(y|x)-C(P|n)-O(\log C(P|n))$ is close to an equality:
246: for all $\alpha$ there are $P,x,y$ such that $CC_I^P(x,y)=C(y|x)-\alpha+O(1)$ and
247: $C(P|n)\le\alpha +O(1)$.\\
248: Indeed, let $x$ be some string. Let $y$ be a random string of length
249: $n$, independent of $x$, that is, $C(y|x)=n+O(1)$.
250: Let $P$ be the following protocol:
251: Bob first looks whether his string $y'$ has the same prefix of length
252: $\alpha$ than $y$. If this is the case he sends to
253: Alice 0 and then $n-\alpha$ remaining bits of $y'$ and Alice
254: prefixes the $n-\alpha$ received bits by $\alpha$ first
255: bits of $y$ and outputs the resulting string.
256: Otherwise Bob sends to
257: Alice 1 and then $y'$. The complexity of this protocol is at most
258: $\alpha+O(1)$, as both Alice and Bob need to know only the
259: first $\alpha$ bits of $y$. And we have $CC_I^P(x,y)=n-\alpha=
260: C(y|x)-\alpha+O(1)$.
261:
262: \section{Other functions}
263: Because Alice can compute every computable function once she knows Bob's
264: input, we have
265: $\CC^{P'}_f (x,y) \leq \CC^{P}_I(x,y)$, with $C(P') \leq C(P,f)+O(1)$.
266: %$\TCC^{\alpha+O(1)}_f (x,y) \leq \TCC^{\alpha}_I(x,y)$.
267:
268: The trivial lower bound on the individual communication complexity
269: of a function $f$ is $\CC^P(x,y)\ge C(f(x,y) \mid x,P)-O(1)$
270: (and hence $\TCC^{\alpha}_f(x,y)\ge
271: C(f(x,y) \mid x)-\alpha-O(\log\alpha)$ anticipating
272: on a later defined notion).
273: In this section we establish some nontrivial lower bounds
274: on $\CC^P(x,y)$ for $P$
275: computing $f$ on all arguments for the inner product function, the equality function
276: and for
277: random Boolean functions.
278:
279:
280: \subsection{Inner Product}
281: We extend an argument introduced in \cite{BJLV00}.
282: Initially, Alice has a string $x =
283: x_1,\ldots,x_n$ and Bob has a string $y = y_1,\ldots,y_n$ with
284: $x,y \in \{0,1\}^n$. Alice and Bob
285: compute the inner product of $x$ and $y$ modulo 2
286: \[f(x,y) = \sum_{i=1}^n x_i \cdot y_i \bmod 2\]
287: with Alice ending up with the result.
288:
289: \begin{theorem}
290: Every deterministic protocol $P_f$ computing
291: the inner product function $f$ (without help bits)
292: requires at least $\CC^P (x,y) \geq C(x,y \mid P)-n-O(1)$ bits of communication
293: on all $x,y$.
294: \end{theorem}
295:
296: \begin{proof}
297: Fix a communication protocol $P$ that computes the inner product.
298: Let Alice's input be $x =
299: x_1\ldots x_n$ and Bob's input be $y_1\ldots y_n$.
300: Run the communication protocol $P$ on $x,y$ and
301: let $c(x,y)$ be the communication between Alice and Bob.
302: Note that $P$ can be viewed as a tree with $c(x,y)$ a path in that
303: tree~\cite{KN97}. Hence $c(x,y)$ form a prefix free set.
304: Consider the set $S=S(x,y)$ defined by
305: \[
306: S := \{ (a,b) \mid C(a,b) = C(x,y),\
307: \text{and Alice outputs $f(x,y)$ having the conversation $C(x,y)$ and input $a$}\}.
308: \]
309: We claim that $|S|\le 2^n$.
310: To prove the claim assume first that $f(x,y)=0$.
311: Let $X$ be the first projection $X$ of $S$ and
312: $Y$ be the second projection of $S$.
313: Being an intersection of two rectangles, $S$ is a rectangle too.
314: As $P$ computes $f$ we know that $f(a,b)=0$ for all $(a,b)\in S$.
315: In other words, every
316: element of $X$ is orthogonal to every element in $Y$ hence
317: $\text{rank}(X)+\text{rank}(Y)\le n$.
318: Thus $|S|=|X|\cdot|Y|\le 2^{\text{rank}(X)+\text{rank}(Y)}\le 2^n$.
319: Assume now that that $f(x,y)=1$.
320: Again $S=X\times Y$ for some $X,Y$ and
321: $f(a,b)=1$ for all $(a,b)\in S$.
322: Subtracting $x$ from the first component of all pairs in $S$
323: we obtain a rectangle $S'$ such that $f(a,b)=0$ for all $(a,b)\in S'$.
324: By above argument, we have $|S'|\le 2^n$. As $|S'|=|S|$ we are done.
325:
326: Given $P$, $C(x,y)$, $f(x,y)$ and the index of $(x,y)$ in $S$
327: we can compute $(x,y)$.
328: By the prefix free property,
329: $C(x,y)$ and the index of $(x,y)$ can be concatenated without delimiters.
330: Consequently,
331: $C(x,y|P) \leq |c(x,y)|+n+O(1)$.
332: \end{proof}
333:
334: \begin{remark}
335: \rm
336: The result of the theorem
337: is only significant for $C(x,y) > n$, but it cannot be improved.
338: Namely, $\CC^P (x,y) \geq C(x,y \mid P)-n-O(1)
339: = C(y \mid x,P) - (n- C(x \mid P))-O(\log n)$, where
340: the equality follows from the ``symmetry of information'' property of
341: Kolmogorov complexity \cite{LiVi97}. The term $n- C(x \mid P)$
342: is a called the {\em randomness deficiency} of $x$ with respect to $P$.
343: Clearly,
344: for $x=00 \ldots 0$
345: it is maximal with $C(x \mid P)= O(\log n)$
346: and Alice knows already from her input $x$ that $f(x,y)=0$
347: and no bits, or only one bit, depending on the protocol conventions,
348: need to be exchanged: $\CC_f^P (x,y)=0$ irrespective of the complexity of $y$
349: which can be $n$.
350: \end{remark}
351:
352: %\begin{remark}
353: %\rm
354: %Thus, $\TCC_f^{\alpha} (x,y) \leq \TCC_I^{\alpha} (x,y)$ for all $x,y,{\alpha}$
355: %and $\TCC_f^{\alpha} (x,y) \ll \TCC_I^{\alpha} (x,y)$ for some $x,y,{\alpha}$, where $I$ is the mother
356: %problem defined before.
357: %\end{remark}
358:
359: \subsection{Random Functions}
360: Alice has $x = x_1 \ldots x_n$ and Bob has $y_1 \ldots y_n$,
361: and $f: \{0,1\}^{2n} \rightarrow \{0,1\}$ satisfies
362: \begin{equation}\label{eq.randomf}
363: C(f \mid n) \geq 2^{2n}- n.
364: \end{equation}
365: The latter condition means that the truth table describing the outcomes
366: of $f$ for the $2^n$ possible inputs $x$ (the rows) and the $2^n$
367: possible inputs for $y$ (the columns) has high Kolmogorov complexity.
368: If we flip the truth table for a prospective $f$ using a fair coin, then
369: with probability at least $1- 2^{-n}$ it will satisfy (\ref{eq.randomf}).
370:
371:
372: \begin{theorem}
373: Every deterministic protocol $P$ computing
374: a function $f$ satisfying (\ref{eq.randomf}) (without help bits)
375: requires at least
376: $\CC_f^P (x,y) \geq \min \{C(x \mid P), C(y \mid P)\}
377: - \log n-O(1)$.
378: \end{theorem}
379:
380: \begin{proof}
381: Run the communication protocol $P$ on $x,y$ and
382: let $c(x,y)$ be the communication between Alice and Bob.
383: Consider the set $S=S(x,y)$ defined by
384: \[
385: S= \{ (x',y') \mid c(x',y') = c(x,y),\
386: \text{and Alice outputs $f(x,y)$ having the conversation $c(x,y)$ and input $x'$}\}.
387: \]
388: Then $S$ is a
389: {\em monochromatic rectangle} in
390: the function table of $f$ (that is, $f(x',y')=f(x,y)$ for all $(x',y')\in S$).
391: Suppose the rectangle $S$ has dimensions $a \times b$.
392: Then we can describe $f$ by giving $f(x,y)$, the value of $a$ in $2\log a +O(1)$
393: bits, the value of $b$ in $2\log b+O(1)$ bits,
394: the positions of the rows of the rectangles $an$ bits,
395: the positions of the columns of the rectangles in $bn$
396: bits, all of the table except the rectangle,
397: in row-major order, in $2^{2n} - ab$ bits.
398: This description must have length at least the
399: Kolmogorov complexity, so by (\ref{eq.randomf}) we find
400: $$
401: 2^{2n} - ab + (a+b)n + 2 \log ab+O(1)
402: \geq 2^{2n}-n.
403: $$
404: Assume w.l.o.g. that $b\ge a$.
405: Then $a<3n$ if $n$ is large enough.
406: (Otherwise we would have
407: $3bn\le ab\le (2b+1)n+2\log b^2+O(1)$.)
408: Given the communication
409: sequence $c(x,y)$, $n$ and $f(x,y)$
410: we can find the rectangle $S$ that it defines. Then, we can
411: reconstruct $x$ by indicating its row in the rectangle.
412: Then
413: $C(x \mid P ) \leq |c(x,y)| + \log n+O(1)$.
414: \end{proof}
415:
416:
417: \subsection{Equality and Functions with Large Monochromatic Rectangles}
418: Let $f$ be the equality function, with $f(x,y)=1$ if $x=y$ and 0 otherwise.
419:
420: \begin{theorem}
421: For every deterministic protocol $P$ computing $f$ we have
422: $\CC^P(x,x) \geq C(x \mid P)-O(1)$ for all $x,y$.
423: On the other hand there is $P$ of complexity $O(1)$ such that
424: there are $x,y$ ($x \neq y$) with
425: $C(x \mid P), C(y \mid P) \geq n-1$ for which
426: $\CC_f^P(x,y) = 2$.
427: \end{theorem}
428: \begin{proof}
429: Lower bound:
430: Since trivially the communication sequence must be different and
431: uniquely identify $x$ if both Alice and Bob have input $x$, we have
432: also $\CC_f^P(x,x) +O(1)\geq C(x \mid P)$.
433:
434: Upper bound: In the function table the lower left rectangle consisting
435: of all $x$'s beginning with 0 and all $y$'s beginning with 1
436: is monochromatic (entries are all 0). Thus, a protocol where Bob communicates
437: one bit to Alice indicating whether $x$ starts with 0 allows Alice, in case
438: $y$ starts with 1, to output 0.
439: Otherwise Alice and Bob start the default protocol.
440: Thus, for such $x,y$ and $P$ we have $\CC_f^P(x,y) = 2$.
441: By simple counting for some such inputs
442: we have $C(x \mid P),C(y \mid P) \geq n-1$.
443: \end{proof}
444:
445: Generalizing this idea,
446: every function that contains large monochromatic rectangles,
447: of size say $2^{2n}/n^{O(1)}$, has many pairs $x,y$ of complexity
448: close to $n$ for which the individual communication complexity
449: drops to $O(\log n)$, as follows:
450:
451: In round 1 Bob tells Alice in which large rectangle (if any) his input
452: is situated, by sending the index of the rectangle to Alice, and 0 otherwise.
453: If Bob did send an index, and Alice's input is in that rectangle as well,
454: then Alice outputs the color (``0'' or ``1'') of the rectangle. Otherwise,
455: Alice starts a default protocol.
456:
457:
458: \section{Total protocols}
459:
460: Let $f$ be a function defined on pairs of strings
461: of the same length. Assume that
462: Alice has $x$, Bob has $y$
463: and Alice wants to compute $f(x,y)$.
464: As the complexity measure we consider
465: the number of bits communicated
466: between Alice and Bob.
467: The naive definition of the individual communication
468: complexity of the value of the function $f$ on the argument $(x,y)$
469: is the number of communicated bits in the ``best'' communication protocol.
470: %However we do not see how to define the best protocol for
471: %given $f$. Note that
472: Then, for every $x,y$ there is a protocol with no communication
473: at all on $(x,y)$: the string $y$ is hard wired into the protocol.
474: To meaningfully capture the individual communication
475: complexity of computing a function $f(x,y)$ we define now the following
476: notion.
477: \begin{definition}
478: \rm
479: Let $\alpha$ be a natural number parameter. Let $\TCC^{\alpha}_f(x,y)$ stand
480: for the minimum $\CC^{P}(x,y)$ over all total protocols
481: $P$ of complexity at most $\alpha$ that always
482: compute $f$ correctly (being total such a protocol
483: terminates for
484: all inputs, and not only for $(x,y)$).
485: \end{definition}
486:
487: For $\alpha=n+O(1)$ we have $\TCC^{\alpha}_f(x,y)=0$ for all computable
488: $f$ and all $x,y$, since we can hard wire $y$ into the protocol.
489: Therefore it is natural to consider only $\alpha$
490: that are much smaller than $n$, say $\alpha=O(\log n)$.
491: Since computation of the Identity function suffices to compute
492: all other (recursive) functions we have
493: $\TCC^{\alpha+O(1)}_f (x,y) \leq \TCC^{\alpha}_I(x,y)$.
494: The trivial lower bound
495: is $\TCC^{\alpha}_f(x,y)\ge C(f(x,y) \mid x)-\alpha-O(\log\alpha)$.
496: For $f=I$ this gives
497: $\TCC^{\alpha}_I(x,y)\ge C(y \mid x)-\alpha-O(\log\alpha)$.
498:
499: \subsection{One-way equals two-way for Identity}
500: Let $\TCC_{f,\text{1-way}}^\alpha(x,y)$ stand
501: for the minimum $\TCC^{P}(x,y)$ over all one-way (Bob sends a message to Alice)
502: total protocols $P$ of complexity at most $\alpha$
503: computing $f$ (over
504: all inputs, and not only on $(x,y)$).
505: It is clear that
506: $\TCC_{f,\text{1-way}}^\alpha(x,y)$ does not depend on $x$:
507: indeed, consider for given $(x,y)$ the best protocol
508: $P$; that protocol sends the same message
509: on every other pair $(x',y)$ hence
510: $\TCC^{\alpha}_{f,\text{1-way}}(x',y)\le \TCC^{\alpha}_{f,\text{1-way}}(x,y)$.
511: Therefore
512: we will use the notation $\TCC_{f}^\alpha(y)$ dropping both $x$ and ``1-way''.
513: %\subsection{Identity function}
514: Obviously,
515: $$
516: \TCC_{f}^\alpha(x,y)\le \TCC_f^\alpha(y)
517: $$
518: for all $\alpha,x,y,f$.
519:
520: Surprisingly, for $f=I$, the Identity function,
521: this inequality is an equality.
522: That is, for total protocols ``1-way'' is as powerful as ``many-way.''
523: More specifically, the following holds.
524:
525: \begin{theorem}\label{th1}
526: There is a constant $C$ such that for all $\alpha,x,y$ we have
527: $$
528: \TCC_I^{\alpha+C}(y) \le \TCC_{I}^{\alpha}(x,y).
529: $$
530: \end{theorem}
531: \begin{proof}
532: Pick a two-way protocol $P$ witnessing $\TCC_{I}^{\alpha}(x,y)=l$.
533: Let $c=c(x,y)$ be the conversation according to $P$ between Alice and Bob on inputs $x,y$.
534: It is known that the set of all pairs $(x',y')$ such that
535: the conversation between Alice and Bob
536: on input $(x',y')$ is equal to $c$ is a rectangle, that is, has the form $X\times Y$,
537: for some $X,Y\subset\{0,1\}^n$.
538: The set $Y$ is a one-element set, as for every $y'\in Y$ Alice
539: outputs $y$ also on the input $(x,y')$ (the output of Alice depends on $c,P,x$ only).
540:
541: Consider the following 1-way protocol $P'$: find an $x'$ with minimum
542: $c(x',y)$ and send $c(x',y)$ to Alice. Alice
543: then finds the set of all pairs $(x'',y')$ such that
544: the conversation between Alice and Bob
545: on input $(x'',y')$ is equal to $c(x',y)$.
546: As we have seen that set has the form $X\times\{y\}$ for some
547: $X$. Thus Alice knows $y$. As $|c(x',y)|\le|c(x,y)|=\TCC_{I}^{\alpha}(x,y)$
548: and $C(P'|P)=O(1)$ we are done.
549: \end{proof}
550:
551: \subsection{Non-Communicable objects}
552:
553: The function $\TCC^{\alpha}_I(y)$, as a function of $y,\alpha$,
554: essentially coincides with
555: \emph{Kolmogorov structure function} $h_y(\alpha)$
556: studied in \cite{GTV01,VV02}.
557: The latter is defined by
558: $$%\begin{equation}\nolabel
559: h_{y}(\alpha) = \min_{S} \{\log | S| : S \ni y,\; C(S) \leq \alpha\},
560: $$%\end{equation}
561: where $S$ is a finite set
562: and $C(S)$ is
563: the length (number of bits) in the
564: shortest binary program from which the reference universal
565: machine $U$
566: computes a listing of the elements of $S$ and then
567: halts.
568: More specifically we have
569: \begin{align}\label{eq.equiv}
570: \TCC^{\alpha+O(1)}_I(y) &\le h_{y}(\alpha),\\
571: h_{y}(\alpha+O(\log n))&\le \TCC^{\alpha}_I(y).
572: \nonumber
573: \end{align}
574:
575: To prove the first inequality
576: we have to transform a finite set $S\ni y$
577: into a one-way protocol $P$ of complexity at most
578: $\alpha = C(S)+O(1)$ witnessing
579: $\TCC^{\alpha}_I (y)\le\log|S|$. The protocol just
580: sends the index of $y$ in $S$, or $y$ literally if $y \not\in S$.
581:
582: To prove the second inequality
583: we have to transform
584: a one-way total protocol $P$ into a finite set $S\ni y$
585: of complexity at most $C(P)+O(\log n)$
586: with $\log|S|\le \CC^P(y)$.
587: The set consists of all $y'$ on which $P$ sends
588: the message of the same length $l$ as the length of the message on $y$.
589: Obviously, $|S|\le 2^{l}=2^{\CC^P(y)}$ and to specify
590: $S$ we need a program describing $P$ and $l$.
591: Thus $C(S)\le C(P)+O(\log \TCC^{\alpha}_I(y) )\le C(P)+O(\log n)$.
592:
593: For the properties of $h_{y}(\alpha)$, which by Theorem~\ref{th1}
594: are also properties of $\TCC_{I}^\alpha(x,y)$, its relation with
595: Kolmogorov complexity $C(y)$ of $y$ and possible shapes
596: of the function $\alpha\mapsto h_{y}(\alpha)$
597: we refer to~\cite{VV02}.
598:
599: We will present here only a few properties. First, two
600: easy inequalities:
601: For all $\alpha\ge O(1)$ and all $x,y$ we have
602: \begin{equation}\label{eq1}
603: C(y|n)-\alpha-O(\log\alpha)\le \TCC^{\alpha}_I(y)\le n-\alpha+O(1).
604: \end{equation}
605: The first inequality is the direct consequence of the
606: inequality $C(y|n)\le \CC^P(y)+C(P|n)+O(\log C(P|n))$,
607: which is trivial.
608: To prove the second one consider the protocol that sends $n-\alpha+C$
609: bits of $y$ (for appropriate constant $C$) and the remaining
610: $\alpha$ bits are hardwired into the protocol.
611: Its complexity is at most $\alpha-C+O(1)\le\alpha$ for appropriate
612: choice of $C$.
613:
614: The second property is not so easy. Given $y$, consider values of $\alpha$
615: such that
616: \begin{equation}\label{eq.ss}
617: TCC^{\alpha}_I(y)+\alpha = C(y)+O(1).
618: \end{equation}
619: That is, the protocol $P$ witnessing (\ref{eq.ss}) together with
620: the one-way communication
621: record Bob sends to Alice form a two-part code for $y$ that is---up to
622: an independent additive constant---as
623: concise as the shortest one-part code for $y$ (that has length $C(y)$
624: by definition).
625: Following the usage in \cite{VV02} we call
626: $P$ a ``sufficient'' protocol for $y$.
627: The descriptions of the protocol plus the communication precisely
628: describe $y$, and in fact, it can be shown that the converse holds
629: as well (up to a constant additive term).
630: There always exists such a protocol, since the protocol that
631: contains $y$ hard wired in the form of a shortest program of
632: length $C(y)$ satisfies the equality with $\alpha = C(y)+O(1)$
633: and $TCC^{\alpha}_I(y)=0$.
634: By definition we cannot have
635: $TCC^{\alpha}_I(y)+\alpha < C(y)-O(1)$, but for $\alpha$ sufficiently
636: small we have $TCC^{\alpha}_I(y)+\alpha > C(y)+O(1)$.
637: %Let $\alpha_0$ be a least $\alpha$ satisfying (\ref{eq.ss}).
638: %The protocol witnessing $\alpha_0$ is a ``minimal'' sufficient statistic.
639: In fact, for every form of function satisfying the obvious constraints
640: on $TCC^{\alpha}_I$ there is a $y$ such
641: that $TCC^{\alpha}_I (y)$ realizes that function up to logarithmic
642: precision. This shows that there are essentially non-communicable strings.
643: More precisely:
644:
645: \begin{theorem}\label{th11}
646: For every $k \leq n$ and
647: monotonic decreasing function $h(\alpha)$ on
648: integer domain $[0,k]$ with $h(0)=n$, $h(k)=0$, $C(h)=O(\log n)$, and
649: $h(\alpha)+\alpha \geq k$ for $\alpha \in [0,k]$,
650: there is a string $y$ of length $n$ and
651: $C(y)=k$ such that
652: \begin{align*}
653: \TCC^{\alpha+O(\log n)}_I(y) &\le h(\alpha),\\
654: h(\alpha+O(\log n))&\le \TCC^{\alpha}_I(y).
655: \end{align*}
656: %that $TCC^{\alpha + a}_I(y) = h(\alpha) +b$ with $a,b = \pm O(\log n)$.
657: \end{theorem}
658:
659: The proof is by combining Theorem 1 of \cite{VV02} with (\ref{eq.equiv}).
660: In particular, for every $k < n- O(\log n)$
661: there are strings $y$ of length $n$
662: and complexity $k$ such that $TCC^{\alpha}_I(y) > n - \alpha$
663: for all $\alpha < k - O(\log n)$ while $TCC^{\alpha}_I(y) = O(1)$
664: for $\alpha \geq k+O(1)$. We call such strings $y$ {\em non-communicable}.
665: For example,
666: with $k= (\log n)^2$ this shows that there are $y$ of complexity
667: $C(y)= (\log n)^2$
668: with $TCC^{\alpha}_I(y) = n - (\log n)^2$ for all
669: $\alpha < C(y)-O(\log n)$ and $O(1)$ otherwise.
670: That is, Bob can hold a highly compressible string $y$,
671: but cannot use that fact to reduce the communication
672: complexity significantly below $|y|$! Unless {\em all} information
673: about $y$ is hard wired in the (total) protocol the communication between
674: Bob and Alice requires sending $y$ almost completely literally.
675: For such $y$, irrespective of $x$, the communication complexity is
676: {\em exponential} in the complexity of $y$ for all protocols of
677: complexity less that that of $y$; when the complexity of the protocol
678: is allowed to pass the complexity of $y$ then the communication complexity
679: suddenly drops to 0.
680:
681: %\subsection{General bounds}
682: \begin{corollary}\label{cor11}
683: For every $n,k$ with $k \leq n$
684: there are $y$ of length $n$ and $C(y)=k$ such that for every $x$
685: %(i) $C(y|n)-\alpha-O(\log\alpha)\le \TCC^{\alpha}_I(x,y)\le n-\alpha+O(1)$.
686: $TCC_I^{\alpha} (x,y) \geq n - \alpha$ for $\alpha < C(y)-O(\log n)$; while
687: for every $x,y$ we have
688: $TCC_I^{\alpha} (x,y) = O(1)$ for $\alpha \geq C(y)+O(1)$.
689: \end{corollary}
690: This follows by combining Theorems~\ref{th1},~\ref{th11}.
691: If we relax the requirement of total and correct protocols to partial and
692: partially correct protocols then we obtain the significantly weaker
693: statements of Theorem~\ref{th61} and Corollary~\ref{co21}.
694:
695:
696: \section{Partially correct and partial protocols}
697:
698: The individual communication complexity can decrease
699: if we do not require the communication protocol to be correct on all the
700: input pairs.
701: Let $\CC_{f}^{\alpha}(x,y)$ stand
702: for the minimum $CC^{P}(x,y)$ over all $P$ of complexity at most $\alpha$
703: computing $f$ correctly on input $(x,y)$
704: (on other inputs $P$ may output incorrect result).
705: The minimum of the empty set is defined as $\infty$.
706: Let $\CC_{f,\text{1-way}}^{\alpha}(x,y)$
707: stand
708: for the minimum $CC^{P}(x,y)$ over all one-way (Bob sends a message to Alice)
709: $P$ of complexity at most $\alpha$
710: computing $f(x,y)$ (again, on other inputs $P$ may work incorrectly).
711: For instance, if $f$ is a Boolean function then
712: $\CC_{f,\text{1-way}}^{O(1)}(x,y)=0$ for all $x,y$
713: (either the protocol outputting always 0 or
714: the protocol outputting always 1 is computes $f(x,y)$ for
715: specific pair $(x,y)$).
716:
717:
718: \subsection{Partially correct and partial protocols versus total ones}
719:
720:
721: It is easy to see that in computing the Identity function for some $(x,y)$
722: total, partially correct, protocols
723: are more powerful than totally correct ones.
724: A total partially correct protocol $P$
725: computes $f(x,y)$ correctly for some $(x,y)$,
726: but may err on some inputs $(u,v)$, in which case we set $\CC^P(x,y) = \infty$.
727: Being total such a protocol
728: terminates for
729: all inputs.
730:
731: \begin{definition}
732: \rm
733: Let $\alpha$ be a natural number parameter.
734: Let $\CC^{\alpha}_f(x,y)$ stand
735: for the minimum $\CC^{P}(x,y)$ over all total partially correct protocols
736: $P$ of complexity at most $\alpha$.
737: \end{definition}
738:
739: For instance, for every $n$ there is a total protocol
740: $P=P_n$ computable from $n$ such
741: that
742: $\CC^P_{I,\text{1-way}}(x,x)=0$
743: (Alice outputs her string), thus $\CC^{O(1)}_I(x,x)=0$.
744: On the other hand, for random $x$ of length $n$
745: we have $\TCC^{\alpha}_I(x,x)\ge \TCC^{\alpha-O(1)}_I(x)\ge
746: C(x|n)-\alpha-O(\log\alpha)\ge n-\alpha-O(\log\alpha)$.
747:
748:
749: We also consider partial protocols that
750: on some $x,y$ are allowed to get stuck, that is, give no instructions at
751: all about how to proceed. Formally, such a protocol is a pair of programs
752: $(P_A,P_B)$. The program $P_A$ tells Alice what to do
753: for each $c$ (the current part of the conversation) and
754: $x$: either wait the next bit from Bob, or to send
755: a specific bit to Bob, or to output a certain string and halt.
756: Similarly, the program $P_B$ tells Bob what to do
757: for each $c$ and
758: $y$: either to wait the next bit from Alice or to send
759: a bit to Alice, or to halt.
760: This pair must satisfy the following requirements
761: for all $(x,y)\in\{0,1\}^n$ and all $c$:
762: if a party gets the instruction to send a bit
763: then another party gets the instruction to wait for a bit.
764: However we do not require that for all $(x,y)\in\{0,1\}^n$ and all $c$
765: both parties get some instruction, it is allowed that
766: $P_A,P_B$ start some endless computation.
767: In particular, Alice may wait for a bit and at the same time
768: Bob has no instruction at all.
769:
770: \begin{definition}
771: \rm
772: The complexity of a partial protocol $P=(P_A,P_B)$
773: is defined as $C(P|n)$.
774: We say that $P$ computes $f$ on the input $(x,y)$
775: if Alice outputs $f(x,y)$ when $P_A,P_B$ are
776: run on $(x,y)$. On other pairs Alice is allowed to
777: output a wrong answer or not output anything at all.
778: If protocol $P$
779: does not terminate, or gives an incorrect answer, for
780: input $(x,y)$, then $\CC^{P}(x,y)= \infty$.
781: Two-way and one-way individual communication complexities
782: with complexity of the partial protocol upper bounded by $\alpha$
783: are denoted as
784: $\PCC_{f}^{\alpha}(x,y)$
785: and
786: $\PCC_{f,\text{1-way}}^{\alpha}(x,y)$
787: respectively.
788: \end{definition}
789: Since the total, partially correct, protocols
790: are a subset of the partial protocols, we always have
791: $\PCC_{f}^{\alpha}(x,y) \leq CC_{f}^{\alpha}(x,y) \leq \TCC_f^{\alpha}(x,y)$.
792: Consider again the Identity function.
793: We have the following obvious lower bound
794: \begin{equation}\label{eq51}
795: C(y|x)-\alpha-O(\log\alpha)\le\PCC_{I}^{\alpha}(x,y)
796: \end{equation}
797: for all $\alpha,x,y$.
798: On the other hand we have the following upper
799: bound if $\alpha$ is at least $\log C(y)+O(1)$:
800: \begin{equation}\label{eq2}
801: \PCC_{I,\text{1-way}}^{\log C(y)+O(1)}(x,y)\le C(y).
802: \end{equation}
803: Indeed, we hardwire the value $C(y)$ in the protocol
804: using $\log C(y)$ bits. This enables
805: $P_B$ to find a shortest description $y^*$ of $y$
806: and to send it to Alice; subsequently $P_A$ decompresses the message
807: received from Bob. Note that the program $P_B$
808: gives no instruction to Bob if the complexity of Bob's
809: input is greater than $C(y)$. Therefore, this protocol is not total.
810: Comparing Equation~\eqref{eq2} to Equation~\eqref{eq1}
811: we see that for $\PCC$
812: we have a better upper bound than for $\TCC$. It turns out
813: that for some pairs $(x,y)$ the communication complexity
814: for totally correct (and even for partially correct) protocols
815: is close to the upper bound $n-\alpha$ while
816: the communication complexity for
817: partial protocols is close to the lower bound $C(y|x)\approx \alpha\ll n$.
818:
819: \begin{theorem}\label{th61}
820: For all $\alpha,n,x$ there are $y$ of length $n$ such that
821: $\CC^{\alpha}_I(x,y)\ge n-\alpha$ and
822: $C(y|x)\le\alpha+O(1)$.
823: \end{theorem}
824: \begin{proof}
825: Fix a string $x$. By counting arguments,
826: there is a string $y$ with $\CC^{\alpha}_I(x,y)\ge n-\alpha$.
827: Indeed, there are less than $2^{\alpha+1}$ total protocols of
828: complexity at most $\alpha$. For each total protocol $P$ there are
829: at most $2^{n-\alpha-1}$ different $y$'s with
830: $CC^P(x,y)<n-\alpha$. Therefore the total number of $y$'s with
831: $\CC^{\alpha}_I(x,y)< n-\alpha$ is less than $2^{\alpha+1}2^{n-\alpha-1}=2^n$.
832:
833: Let $y$ be the first string with $\CC^{\alpha}_I(x,y)\ge n-\alpha$.
834: To identify $y$ conditional to $x$ we only need to now the
835: number of total protocols
836: of complexity at most $\alpha$: given that number
837: we enumerate all such protocols until we find all them.
838: Given all those protocols and $x$ we run all of them on all pairs $(x,y)$
839: to find $\CC^{\alpha}_I(x,y)$ (here we use that the protocols are total)
840: for every $y$, and determine the first $y$ for which it is at least $n-\alpha$.
841: Hence $C(y|x)\le\alpha+O(1)$.
842: \end{proof}
843:
844: \begin{corollary}\label{co21}
845: Fix constants $C_1,C_2$ such that
846: $\CC^{\log C(y)+C_1}_{I,\text{1-way}}(x,y)\le C(y)\le n+C_2$.
847: Applying the theorem to the empty string $\epsilon$ and to (say)
848: $\alpha=2\log n$
849: we obtain a $y$ of length $n$ with exponential
850: gap between $\CC^{2\log n}_I(\epsilon,y)\ge n-2\log n-O(1)$ and
851: $\PCC^{\log (n+C_2)+C_1}_{I,\text{1-way}}(\epsilon,y)\le C(y)\le\log n+O(1)$.
852: \end{corollary}
853:
854: Using a deep result of An. Muchnik~\cite{Mu02}
855: we can prove that
856: $\PCC^{\alpha}_{I,\text{1-way}}$ is close to $C(y|x)$
857: for $\alpha\ge O(\log n)$ .
858:
859: \begin{theorem}[An. Muchnik]
860: For all $x,y$ of length $n$ there is $p$ such that
861: $|p|\le C(y|x)+O(\log n)$,
862: $C(p|y)=O(\log n)$ and
863: $C(y|p,x)=O(\log n)$, where the constants in
864: $O(\log n)$ do not depend on $n,x,y$.
865: \end{theorem}
866:
867: \begin{corollary}\label{co31}
868: For all $x,y$ of length $n$ we have
869: $\PCC^{O(\log n)}_{I,\text{1-way}}(x,y)\le C(y|x)+O(\log n)$.
870: \end{corollary}
871: \begin{proof}
872: Let $p$ be the program of Muchnik's theorem,
873: let $q$ be the program of length $O(\log n)$
874: for the reference computer to reconstruct $p$ from
875: $y$ and let $r$ the program of length $O(\log n)$
876: for the reference computer to reconstruct $y$ from the pair $(x,p)$.
877: The protocol is as follows: Bob finds $p$ from $y,q$
878: and sends $p$ to Alice; Alice reconstructs $y$ from
879: $x,r$. Both $q$ and $r$ are hardwired into the protocol,
880: so its complexity is $O(\log n)$.
881: This protocol is partial, as both Bob and Alice may be stuck
882: when reconstructing $p$ from $y',q$ and
883: $y$ from $x',r$.
884: \end{proof}
885:
886: For very small values of $C(y|x),C(y)$ we can do even better using the
887: coloring lemma 3.9 and theorem 3.11 from \cite{BGLVZ}.
888:
889:
890:
891: \begin{lemma}\label{lem.color}
892: Let $k_1,k_2$ be such that $C(x) \leq k_1$ and $C(y \mid x) \leq k_2$,
893: and let $m = |\{(x,y): C(x) \leq k_1, \; C(y \mid x) \leq k_2 \}|$.
894: For $M=2^{k_1}$, $N= 2^{k_2}$
895: and every $1 \leq B \leq N$
896: Bob can compute the recursive function
897: $R(k_1,k_2,m, y) = r_y \leq (N/B)e(MN)^{1/B}$ such that
898: Alice can reconstruct $y$ from $x, r_y,m$ and at most $b \leq \log B$
899: extra bits.
900: \end{lemma}
901:
902: Using $k_1,k_2,m,y$, Bob can compute $r_y$ and
903: send it in $\log r_y$ bits to Alice. The latter computes $y$ from
904: $x,m,r_y$ using additionally $b \leq \log B$ special bits provided
905: by the protocol.
906: Then, the number of bits that need to be communicated,
907: 1 round, from Bob to Alice, is
908: \[
909: \log r_y \leq k_2 - \log B + \frac{k_2 + k_1}{B}.
910: \]
911: The protocol $P=(P_A,P_B)$ uses $\leq 2(k_1+k_2) + b +O(1)$ bits.
912: \begin{corollary}
913: If $C(x),C(y|x) = O(\log \log n)$ and $b= \Theta (\log \log n)$
914: then
915: $PCC^{\Theta (\log \log n)}_{I,1-way} (x,y) \leq C(y|x)- \Theta (\log \log n)$.
916: \end{corollary}
917:
918:
919:
920: \subsection{Two-way is better than one-way for partially correct protocols}
921:
922: Note that for the Identity function all our upper bounds hold
923: for one-way protocols and all our lower bounds hold
924: for two-way protocols. The following question arises:
925: are two-way protocols more powerful than one-way ones
926: (to compute the Identity function)?
927: Theorem~\ref{th1} implies that for total
928: protocol it does not matter whether the communication is one-way or two-way.
929: For partially correct total protocols and partial protocol
930: the situation is different.
931: It turns out that partially correct total
932: two-way protocols are stronger than even
933: partial one-way protocols.
934:
935: \begin{theorem}\label{th7}
936: For every $k,l,s$ such that
937: $k\ge s+l2^{s}$ there are strings $x,y$ of length
938: $(2^{s}+1)k$ such that
939: $\CC_I^{O(1)}(x,y)\le 2^{s}\log(2k)$ but $\PCC^s_{I,\text{1-way}}(x,y)\ge l$.
940: \end{theorem}
941: \begin{proof}
942: We let $x=z_0z_1\dots z_{2^{s}}$ where $z_0,\dots,z_{2^{s}}$ have length $k$
943: and $y=z_j00\dots0$ for some $j$.
944:
945: To prove the upper bound
946: consider the following two-way protocol:
947: Alice
948: finds a set of indexes
949: $I=\{i_1,\dots,i_{2^{s}}\}$ such that for every distinct $j,m$ there is $i\in I$
950: such that $i$th bit of $z_j$ is different from $i$th bit of $z_m$ (such set
951: does exist, which may be shown by induction).
952: Then she sends to Bob the string $i_1\dots i_{2^{s}}$ and Bob sends
953: to Alice $i$th bit of $y$ for all $i\in I$. Alice knows now $y$.
954:
955: We need to find now particular $z_0,z_1,\dots, z_{2^{a+b+s}}$ such that no one-way
956: protocol is effective on the pair $(x,y)$ obtained from
957: them in the specified way. To this end let $P_1,\dots, P_N$ be
958: all the one-way partial protocols of complexity less than $s$
959: computing the identity function.
960: For every $z$ and $i\le N$
961: let
962: $c(z,i)$ denote the message sent by Bob in protocol $P_i$ when he receives
963: $z00\dots 0$ as help bits provided the length of the message is less than $l$.
964: Otherwise let $c(z,i)=\infty$. Let $c(z)$ stand for the concatenation
965: of $c(z,i)$ over all $i$.
966: The range of $c(z)$ has $(2^l)^{N}<2^{l2^{s}}$ elements.
967: Hence there
968: is $c$ such that for at least $2^{k-2^{s}l}>2^{s}$ different $z$'s
969: we have $c(z)=c$. Pick such $c$ and pick different
970: $z_0,z_1,\dots, z_{2^{s}}$ among those $z$'s.
971: Let $y_j$ stand for the string obtained from $z_j$ by appending 0s.
972: We claim that $\CC^{P_i}_I(x,y_j)\ge l$ for some $j$ for all $i\le N$.
973: Assume that this is not the case. That is,
974: for every $j$ there are $i$ such that
975: $\CC^{P_i}_I(x,y_j)<l$.
976: There are $j_1\ne j_2$ for which $i$ is the same.
977: As $c(z_{j_1},i)=c(z_{j_2},i)\ne\infty$
978: Alice receives the same message in $P_i$ on inputs $(x,y_{j_1})$, $(x,y_{j_2})$
979: and should output both answers
980: $y_{j_1},y_{j_2}$, which is a contradiction.
981: \end{proof}
982:
983: \begin{corollary}
984: Let in the above theorem $s=(\log k)/3$ and $l=k^{2/3}/\log k$.
985: These values satisfy the condition $k\ge s+l2^{s}$ and
986: hence there are $x,y$ of length about $k^{4/3}$
987: %begin changes Paul: changed
988: %with almost quadratic gap between $\TCC^{O(1)}_I(x,y)\le k^{1/3}\log k$
989: %and $\TCC^{(\log k)/3}_{I,\text{1-way}}(x,y)\ge k^{2/3}/\log k$.
990: %to
991: with almost quadratic gap between $\CC^{O(1)}_I(x,y)\le k^{1/3}\log k$
992: and $\PCC^{(\log k)/3}_{I,\text{1-way}}(x,y)\ge k^{2/3}/\log k$.
993: Letting $s=\log\log k$ and $l=k/(2\log k)$
994: we obtain
995: $x,y$ of length about $k\log k$
996: %and
997: %with an exponential gap between $\TCC^{O(1)}_I(x,y)\le \log k\log(2k)$
998: %and $\TCC^{\log k}_{I,\text{1-way}}(x,y)\ge k/(2\log k)$.
999: %to
1000: with an exponential gap between $\CC^{O(1)}_I(x,y)\le \log k\log(2k)$
1001: and $\PCC^{\log k}_{I,\text{1-way}}(x,y)\ge k/(2\log k)$.
1002: \end{corollary}
1003:
1004: \section{Summary of some selected results for comparison}
1005: \begin{align*}
1006: \bullet \; \; &\forall_{x,y,\alpha} [ \TCC_I^{\alpha} (x,y) \geq
1007: \CC_I^{\alpha} (x,y) \geq \PCC_I^{\alpha} (x,y) ] \; \text{by definition}.
1008: \\
1009: \bullet \; \; & \forall_{\alpha,x,y}
1010: [\TCC_I^{\alpha + O(1)} (y) = \TCC_I^{\alpha} (x,y) + O(1) ]\;
1011: \text{Theorem~\ref{th1} and discussion}.
1012: \\
1013: \bullet \; \; &\forall_{n,k,\alpha} \exists_{y, |y|=n, C(y)=k} \forall_{x} [ \alpha < C(y) - O(\log n)
1014: \Rightarrow \TCC_I^{\alpha} (x,y) \geq n- \alpha ]\;
1015: \text{Corollary~\ref{cor11}}.
1016: \\
1017: \bullet \; \; &\forall_{x,y,\alpha}
1018: [ \alpha \geq C(y) - O(1)
1019: \Rightarrow \TCC_{I}^{\alpha} (x,y) = O(1) ] \; \text{Corollary~\ref{cor11}}.
1020: \\
1021: \bullet \; \; &\forall_{n,x,\alpha} \exists_{y, |y|=n}
1022: [ \alpha \geq C(y|x) - O(1) \& \CC_{I}^{\alpha} (x,y) \geq n - \alpha ]
1023: \; \text{Theorem~\ref{th61}}.
1024: \\
1025: \bullet \; \; &\forall_{x,y,\alpha}
1026: [\PCC_{I}^{\alpha} (x,y) \geq C(y|x) - \alpha - O(\log \alpha) ] \;
1027: \text{ (\ref{eq51})}.
1028: \\
1029: \bullet \; \; &\forall_{n,x,y}
1030: [\PCC_{I, 1-way}^{\log C(y)+O(1)} (x,y) \leq C(y) ] \;
1031: \text{ (\ref{eq2})}.
1032: \\
1033: \bullet \; \; &\forall_{n} \forall_{x,y, |x|=|y|=n}
1034: [\PCC_{I, 1-way}^{O(\log n)} (x,y) \leq C(y|x)+O(\log n) ]
1035: \; \text{ Corollary~\ref{co31}}.
1036: \\
1037: \bullet \; \; &\forall_{k,l,s: k \geq s+l2^s} \exists_{x,y, |x|=|y|= (2^s+1)k}
1038: [ \CC^{O(1)}_I(x,y) \leq 2^s \log (2k) \& \PCC^s_{I, 1-way} (x,y) \geq l]
1039: \; \text{Theorem~\ref{th7}}.
1040: \end{align*}
1041:
1042: \section{Protocols using help bits}
1043: A (partial) protocol with $a$ help bits for Alice $b$ help bits for Bob
1044: on strings of length $n$ may be defined
1045: as a regular (partial) protocol $P$ on inputs $u,v$ of length
1046: $n+a,n+b$, respectively. We say that
1047: $P$ computes $f$ on the input $(x,y)$ if
1048: there are
1049: $h_A\in\{0,1\}^a$ and
1050: $h_B\in\{0,1\}^b$ such according to $P$ on input $(xh_A,yh_B)$ Alice outputs $f(x,y)$.
1051: The crucial point in this definition is that
1052: the help bit sequences $h_A,h_B$ may depend on the input pair $(x,y)$.
1053: We say that $P$ computes $f$ if it computes $f$ on
1054: all input pairs $(x,y)\in\{0,1\}^n$.
1055: Thus $P$ may compute many different functions. For instance there is a protocol
1056: with 1 help bit for Alice and no help bits for Bob that
1057: computes every Boolean function: Alice just receives the value
1058: of the function as the help bit and outputs it.
1059: Define $\CC^P(x,y)$ as
1060: the minimum of the length of conversation on input $(xh_A,yh_B)$
1061: according to $P$ over $h_A\in\{0,1\}^a$,
1062: $h_B\in\{0,1\}^b$ such that on input $(xh_A,yh_B)$ Alice outputs $f(x,y)$
1063: (the minimum of the empty set is defined as $\infty$).
1064: We define $\TCC^{\alpha,a,b}_{f}(x,y)$ as
1065: the minimum $CC^{P}(x,y)$ over all $P$ of complexity at most $\alpha$
1066: computing $f$ (over
1067: all inputs, and not only on $(x,y)$).
1068: Define $\TCC_{f,\text{1-way}}^{\alpha,a,b}(x,y)$
1069: $\CC^{\alpha,a,b}_{f}(x,y)$, $\CC^{\alpha,a,b}_{f,\text{1-way}}(x,y)$
1070: analogously (in the latter two we minimize over all $P$ of complexity at
1071: most $\alpha$).
1072:
1073:
1074: \subsection{Partially correct protocols versus total ones---with help bits}
1075: In contrast to the no-help-bit
1076: case, now the difference between totally and partially correct
1077: protocols in not essential:
1078: allowing only one extra help bit
1079: we can effectively transform a protocol $P$ computing
1080: $f$ on specific input $(x,y)$
1081: into a protocol $P'$ computing $f$ an all inputs
1082: so that $\TCC^{P'}_f(x,y)\le CC^{P}_f(x,y)$:
1083: at first Alice and Bob receive 1 bit of help information
1084: (or only one of them, and in that case he/she resends
1085: that bit to the other; in this case the right hand side of the inequality should be
1086: incremented by 1). If this is the case then they start $P$. Otherwise they start
1087: the default protocol.
1088: So we obtain
1089: \begin{align*}
1090: \TCC^{\alpha+O(1),a+1,b+1}_{f}(x,y)&\le \CC^{\alpha,a,b}_{f}(x,y),\\
1091: \TCC^{\alpha+O(1),a,b+1}_{f}(x,y)&\le \CC^{\alpha,a,b}_{f}(x,y)+1,\\
1092: \TCC^{\alpha+O(1),a+1,b}_{f}(x,y)&\le \CC^{\alpha,a,b}_{f}(x,y)+1.
1093: \end{align*}
1094:
1095:
1096: The same applies to $\TCC_{\text{1-way}}$ and $\CC_{\text{1-way}}$
1097: (except the last inequality, as now Alice is unable to send to Bob).
1098: Therefore we will not consider specially totally correct protocols.
1099: We will study only values
1100: $\CC^{\alpha,a,b}_{f}(x,y)$, $\CC^{\alpha,a,b}_{f,\text{1-way}}(x,y)$.
1101:
1102: Moreover,
1103: we can decrease $a$ by $a'$ at the expense of increasing
1104: $\alpha$ by $a'+b'+O(\log b')$ (help bits are appending
1105: to the program specifying the protocol), and similarly for $b$:
1106: \begin{align*}
1107: \CC^{\alpha+a'+b'+O(\log a'b'),a,b}_{f}(x,y)&\le \CC^{\alpha,a+a',b+b'}_{f}(x,y),
1108: \end{align*}
1109: but not vice verse. The same is true for 1-way protocols.
1110:
1111: \subsection{Partial protocols with help bits}
1112: For partial protocol we can even decrease $\alpha$
1113: at the expense of increasing
1114: both $a$ and $b$: indeed let $p$ be the shortest
1115: program for $(P_A,P_B)$ and let $q$ we the prefix of $p$ and
1116: $r$ be the remaining bits of $p$. Consider now the following
1117: programs $P_A',P_B'$; both receive $r$ as help an both
1118: have $p$ hard wired.
1119: $P_A$ appends computes $p=qr$ and decompresses $p$
1120: and then executes $P_A$. The program $P_B$ acts in a similar way.
1121: Note that $C(P_A',P_B')\le|q|+O(1)$.
1122: Thus we obtain
1123: \begin{align*}
1124: \PCC^{\alpha+O(1),a+\alpha'+O(\log\alpha'),b+\alpha'+O(\log\alpha')}_{f}(x,y)&
1125: \le \PCC^{\alpha+\alpha',a,b}_{f}(x,y).
1126: \end{align*}
1127:
1128: \subsection{Two-way is better than one-way with help bits}
1129:
1130: \begin{theorem}
1131: For every $k,l,s,a,b$ such that
1132: $k\ge a+b+s+l2^{s+b}$ there are strings $x,y$ of length
1133: $(2^{a+b+s}+1)k$ such that there is a two-way protocol
1134: of complexity $O(1)$ with $1$ help bit (either for Alice or for Bob)
1135: such that $\CC^P(x,y)\le 2^{a+b+s}\log(2k)+1$ but for every
1136: one-way protocol $P$ of complexity less than $s$
1137: with $a$ help bits for Alice and $b$ help bits for
1138: Bob we have $\CC^P(x,y)\ge l$.
1139: \end{theorem}
1140: \begin{proof}
1141: The proof is similar to the proof of the previous Theorem~\ref{th7}.
1142:
1143: We let $x=z_0z_1\dots z_{2^{a+b+s}}$ where $z_0,\dots,z_{2^{a+b+s}}$ have length $k$
1144: and $y=z_j00\dots0$ for some $j$.
1145:
1146: To prove the upper bound
1147: consider the following two-way protocol: if $x,y$ has not the above form
1148: Alice receives 0 as the help bit and starts the default protocol.
1149: Otherwise she receives 1 as the help bit and finds a set of indexes
1150: $I=\{i_1,\dots,i_{2^{a+b+s}}\}$ such that for every distinct $j,m$ there is $i\in I$
1151: such that $i$th bit of $z_j$ is different from $i$th bit of $z_m$.
1152: Then she sends to Bob the string $1i_1\dots i_{2^{a+b+s}}$ and Bob sends
1153: to Alice $i$th bit of $y$ for all $i\in I$. Alice knows now $y$.
1154:
1155:
1156: We need to find now particular $z_0,z_1,\dots, z_{2^{a+b+s}}$ such that no one-way
1157: protocol is effective on the pair $(x,y)$ obtained from
1158: them in the specified way. To this end let $P_1,\dots, P_N$ be
1159: all the one-way protocols of complexity less than $s$
1160: with $a$ help bits for Alice and $b$ help bits for Bob computing the identity function.
1161: For every $z$, $i\le N$ and $h_B$ where $h_B$ is a binary sequence
1162: of length $b$
1163: let
1164: $c(z,i,h_B)$ denote the message sent by Bob in protocol $P_i$ when he receives
1165: $z00\dots 0$ as the input and
1166: $h_B$ as help bits provided the length of the message is less than $l$.
1167: Otherwise let $c(z,i,h_B)=\infty$. Let $c(z)$ stand for the concatenation
1168: of $c(z,i,h_B)$ over all $i,h_B$.
1169: The range of $c(z)$ has $(2^l)^{N2^b}<2^{l2^{s+b}}$ elements.
1170: Hence there
1171: is $c$ such that for at least $2^{k-2^{s+b}l}>2^{a+b+s}$ different $z$'s
1172: we have $c(z)=c$. Pick such $c$ and pick different
1173: $z_0,z_1,\dots, z_{2^{a+b+s}}$ among those $z$'s.
1174: Let $y_j$ stand for the string obtained from $z_j$ by appending 0s.
1175: We claim that $\TCC^{P_i}(x,y_j)\ge l$ for some $j$ for all $i\le N$.
1176: Assume that this is not the case. That is,
1177: for every $j$ there are $i,h_A,h_B$ such that
1178: $\TCC^{P_i}(x,y_j)<l$ with help bit sequences $h_A,h_B$.
1179: There are $j_1\ne j_2$ for which the triples $(i,h_A,h_B)$
1180: coincide. As $c(z_{j_1},i,h_B)=c(z_{j_2},i,h_B)\ne\infty$
1181: Alice receives the same message in $P_i$ on inputs $(x,y_{j_1})$, $(x,y_{j_2})$,
1182: with the help bit sequences $h_A,h_B$ and should output both answers
1183: $y_{j_1},y_{j_2}$, which is a contradiction.
1184: \end{proof}
1185:
1186: \begin{corollary}
1187: Let in the above theorem $a=b=s=(\log k)/6$, $l=k^{1/2}/\log k$.
1188: These values satisfy the condition $k\ge a+b+s+l2^{s+b}$ and
1189: hence there are $x,y$ of length about $k^{1.5}$ for which
1190: there is a two-way protocol of complexity $O(1)$ with only one help bit
1191: with $\CC^P(x,y)\le k^{1/3}\log k$ but
1192: there is no one-way protocol of complexity $(\log k)/6$ with $(\log k)/6$ help bits
1193: both for Alice and Bob
1194: with $\CC^P(x,y)<k^{1/2}/\log k$.
1195: \end{corollary}
1196:
1197:
1198: \begin{thebibliography}{4}
1199:
1200: \bibitem{BGLVZ}
1201: C.H. Bennett, P. G\'acs, M. Li, P.M.B. Vit\'anyi, and W. Zurek,
1202: Information Distance,
1203: {\em IEEE Trans. Information Theory}, IT-44:4(1998) 1407--1423.
1204:
1205: \bibitem{BJLV00}
1206: H. Buhrman, T. Jiang, M. Li, P. Vitanyi,
1207: New applications of the incompressibility method: Part II,
1208: {\em Theoretical Computer Science}, 235:1(2000), 59--70.
1209:
1210: \bibitem{GTV01} P. G\'acs, J. Tromp, P.M.B. Vit\'anyi.
1211: Algorithmic statistics, {\em IEEE Trans. Inform. Th.}, 47:6(2001), 2443--2463.
1212:
1213:
1214: \bibitem{KN97}
1215: E. Kushilevitz, N. Nisan, {\em Communication Complexity},
1216: Cambridge University Press, 1997.
1217:
1218: \bibitem{Le86}
1219: L. A. Levin, Average Case Complete Problems,
1220: {\em SIAM J.Comput.}, 15:1(1986), 285--286.
1221:
1222: \bibitem{LiVi97}
1223: M. Li and P.M.B. Vit\'anyi, %
1224: \it An Introduction to Kolmogorov Complexity
1225: and its Applications%
1226: \rm , Springer-Verlag, New York, 2nd Edition, 1997.
1227:
1228: \bibitem{Mu02}
1229: An.A. Muchnik, Conditional complexity and codes, {\em Theoret.
1230: Comput. Sci.}, 271:1/2(2002), 97--111.
1231:
1232: \bibitem{VV02}
1233: N.K. Vereshchagin and P.M.B. Vit\'anyi,
1234: Kolmogorov's structure functions and an application to the foundations of
1235: model selection, {\em Proc. 47th IEEE Symp. Found. Comput. Sci.},
1236: 2002, 751--760.
1237:
1238: \bibitem{Ya79}
1239: A.C. Yao. Some complexity questions related to distributive computing.
1240: In: {\em Proc.
1241: 11th ACM Symposium on Theory of Computing}, 1979, 209--213.
1242: \end{thebibliography}
1243:
1244: \end{document}
1245: