This preview shows page 1. Sign up to view the full content.
Unformatted text preview: ECE 5620 Spring 2011 Homework #6: Solutions
1. The parsing is 0,00,000,1,10,101,0000,01,1010,1. We encode the phrases as follows
Phrase
0
00
000
1
10
101
0000
01
1010
1 Pointer
∅
1
10
00
100
101
011
001
0110
0000 Added Character
0
0
0
1
0
1
0
1
0
1 One can also encode the ﬁnal phrase using only a pointer, 0011, without an added character. Using
the encoding in the table, the encoded string is 01010000110001011011000110110000001. Note that
in this example we have actually made the string longer.
2. (a) The LZ78 parsing is 1,11,111,1111,11111,. . .
(b) Let c(n) denote the number of nonnull phrases in the parsing of the string of length n. Let (n)
denote the length of the encoding for the string of length n. Since the null phrase does not need
to be communicated, and we never refer to the last phrase, we have
(n) ≤ c(n)( log[c(n)] + 1)
≤ c(n)(log[c(n)] + 2). (1) It remains to bound c(n). Using induction one can verify that
m i=
i=1 m(m + 1)
.
2 (2) Thus if n = m(m + 1)/2, then c(n) = m. More generally, if
m(m − 1)
m(m + 1)
<n≤
,
2
2 (3) then c(n) = m. But the ﬁrst inequality in (3) implies that (m − 1)2 < 2n or m = c(n) <
√
2n + 1. Substituting this into (1) yields
√
√
(n)
2n + 1 log[ 2n + 1] + 2
≤
·
→ 0.
n
n
n 1 Universal Source Coding 313 3. (a) The Lemp elZiv parsing is: 0, 00, 001, 000, 0001, 1, 01, 011, 1
(b) The sequence is: 0, 1, 00, 01, 10, 11, 000, 001, . . . concatenating all binary strings of
length 1,2,3, etc. This is the sequence where the phrases are as short as p ossible.
(c) Clearly the constant sequence will do: 1, 11, 111, 1111, . . .
10. The question assumes that we send the midpoint of the ﬁnal interval, (a source (A, P ) . For
4. Two versions of ﬁxeddatabase LempelZiv. Consider a, b), truncated to
simplicity assume that the alphab et is ﬁnite A = A < ∞ , and the symb ols are i.i.d.
∼ P . A ﬁxed database D is given, and log(b − a) + 1 decoder. The encoder parses
− is revealed to the
n into blo cks of length l , and subsequently enco des them by
the target sequence x1
n
bits. Let xnbinaryndescription of their source realizations. For idatabase.suppose thathxis leads
giving the1 and x2 denote two distinct last app earance in the = 1, 2, If a matc i
to a ﬁnal intervalentire iblo).k Letsentdenote the midpoint of the (l log iA interval, truncated to i :=
not found, the of (a , bi c is yi uncompressed, requiring ai , b ) bits. A ﬂag is used
− tell the ai ) + 1 bits. We a matc in class that
to log(bi − decoder whether showed h location is b eing describ ed, or the sequence itself.
Problems (a) and (b) give some preliminaries a ou b
y + will need in showing the optimality
i
i
yi ∈ a i ,
.
of ﬁxeddatabase LZ in (c).
2 l
(a) Let x b e a δ typical sequence of length
Without losslof generality, we may assume that a2 ≥ lb1starting athaveand let R l (x ) b e the
. Then we 0,
corresp onding recurrence index in the inﬁnite past . . . , X −2 , X−1 . Show that
1
y2 − y1 > a2 − (a1 + b1 )
l
l 2l
E Rl (X )X = x ≤ 2l(H +δ)
1
≥ b1 − (a1 + b1 )
2
where H is the entropy rate of the source.
1
= (b l(H +
(b) Prove that for any > 0 , Pr Rl (X l ) > 12− a1 )) → 0 as l → ∞ .
2
Hint: Expand the probability by= 2−(− log(b1 −a1 )+1)
conditioning on strings x l , and break things
up into typical and nontypical. Markov’slog 1
−( − inequality1)
+ and the AEP should prove
(b1 −a1 )
≥2
handy as well.
= 2− 1 .
(c) Consider the following two ﬁxed databases (i) D 1 is formed by taking all δ typical
˜
l vectors; and (ii) D2 formed by taking the most recent L = 2l(H +δ) symb ols in
Now if y1 is speciﬁed to 1 bits and y1 is a preﬁx of y , then y − y1 < 2− 1 . The above calculation
the inﬁnite past (i.e., X−L , . . . , X−1 ). Argue that the algorithm describ ed ab ove
˜
then implies that y1 cannot be a preﬁx of y2 . Since y2 > y1 , y2 cannot be a preﬁx of y1 .
is asymptotically optimal, namely that the exp ected numb er of bitsp ersymb ol
Note that ifergesdoesthe entropylog(b − a) + 1 bits in the binary representation,databaseencoder is
conv one to not use − rate, when used in conjunction with either or if the D 1
permittedD2 send any point in the ﬁnal interval, then the code may not be preﬁxfree.
or to . To see this, consider ﬁrst a uniform distribution over an alphabet of size 3. The midpoint of the second
Solution: Two versions of in binary, which is obviously
symbol’s interval is 1/2 or 0.1 ﬁxeddatabase LempelZiv a preﬁx of the third symbol’s string. If
one uses a binary expansion of − log(b − a) + 1 bits, however, then lthis δ) written as 0.100, which
is
(a) Since xl is δ typical, the AEP implies that p(x l ) ≥ 2− (H + , and the result
is not a preﬁx of the third symbol’s string.
follows from Kac’s lemma.
Consider next a binary distribution with probabilities 9/16 and 7/16. If welare not required to send the
(l)
(b) Fix > 0 , and δ ∈ (0, ) . Let Aδ b e the δ typical set for A . We divide the set
midpoint of the ﬁnal interval, then we may send 1/2 using two bits (0.10) for the ﬁrst symbol and 5/8
of sequences into the typical sequences and the nontypical sequences.
using 3 bits (0.101) for the second symbol. Observe that 10 is a preﬁx of 101. Pr sourcel ) > 2l(H + ) ) =
p(xl Pr(Rl X l ) > 2l(H + ) xl )
5. Consider two (Rl (X strings xn and xn . Suppose )that the(codeword for xn is a preﬁx of the codeword
1
2
1
xl
for xn . Then the set of phrases in the parsing of xn must form the initial set of phrases (in order) in
2
1
l
p(xl ) xn which > 2l(H + ) xl unless they are equal.
the parsing of xn . This requires that =n be a preﬁx ofPr(,Rl (X ) is impossible )
x1
2
2
(l) 6. xl ∈Aδ (a) The progression of the arithmetic encoder is as follows
2 1/ 2 5/8 5/ 8 0 23/32 5/ 8 0 1/2 1
89/128 1
1 1 1 23/32 1 Thus the ﬁnal interval is (a, b) = (89/128, 23/32) = (178/256, 184/256). The midpoint of
this interval is 181/256, or 0.10110101 in binary. Truncating to − log(b − a) + 1 = 7 bits
gives 0.1011010 as the encoded string.
(b) 0.01110 = 14/32. The progression of the arithmetic decoder is as follows 0 0 3/8
13/32 0 13/32
55/128 1/2 1 1 3/ 8
1 1/ 2 1 1/2 1/2 We choose 0 as the ﬁrst symbol because 14/32 < 1/2 We choose 1 for the second symbol
because 14/32 > 3/8, etc. The decoded string is 0111.
7. The quickest way of getting the desired string in the dictionary is to build it up character by character:
WWiWitWith...
where the vertical lines indicate the LZ parsing and are not part of the actual string. Using the formula (2) above, the required string length is then
50 i=
i=1 8. 50 · 51
= 1275.
2 (a) As shown in class, the i.i.d. distribution that assigns the maximum probability to a string is its
type, which in this case is [1/3 2/3]. The probability it assigns is
1
3 n/3 · 2
3 2n/3 ≈ 0.529n . (b) As shown in class, the Markov distribution that assigns the maximum probability to a string has
an initial distribution given by a point mass on the ﬁrst symbol
p(X1 = 0) = 1
3 and a transition probability matrix given by the Markov type
0
1
(n − 3)/(2n − 3) n/(2n − 3). . The probability that this distribution assigns to the string is
n−3
2n − 3 n/3−1 · n
2n − 3 n/3 ≈ .630n . For large n, this is much larger than the probability in (a).
(c) Consider the following secondorder Markov chain
p(X1 = 0, X2 = 1) = 1
p(00, 0) = 1 − p(10, 0) = 1 p(00, 1) = 1 − p(10, 1) = 0 p(01, 0) = 1 − p(11, 0) = 0 p(01, 1) = 1 − p(11, 1) = 1,
where p(ij, k ) refers to p(Xn = iXn−2 = j, Xn−1 = k ). This distribution assigns probability
one to the given string, which is clearly the largest possible.
(d) The distribution in (c) is secondorder Markov, and hence also thirdorder Markov, and assigns
probability one to the given string, which is clearly the largest possible. Thus it is the thirdorder
Markov distribution that assigns the largest possible probability to the string. Note that the same
reasoning implies that it is the maximumlikelihood distribution among Markov distributions of
any order. 4 ...
View
Full
Document
This note was uploaded on 10/03/2011 for the course ECE 5620 at Cornell University (Engineering School).
 '08
 WAGNER

Click to edit the document details