Gauss’s Lemma
Lemma 63 (Gauss) Let p and q be two distinct odd primes. Then (q|p) = (−1)m, where m is the number of residues in R = { iq mod p : 1 ≤ i ≤ (p − 1)/2 } that are greater than (p − 1)/2.
• All residues in R are distinct.
– If iq = jq mod p, then p| (j − i) or p|q.
– But neither is possible.
• No two elements of R add up to p.
– If iq + jq = 0 mod p, then p|(i + j) or p|q.
– But neither is possible.
The Proof (continued)
• Replace each of the m elements a ∈ R such that a > (p − 1)/2 by p − a.
– This is equivalent to performing −a mod p.
• Call the resulting set of residues R′.
• All numbers in R′ are at most (p − 1)/2.
• In fact, R′ = {1, 2, . . . , (p − 1)/2} (see illustration next page).
– Otherwise, two elements of R would add up to p, which has been shown to be impossible.
5 1 2 3 4
6 5
1 2 3 4
6
p = 7 and q = 5.
The Proof (concluded)
• Alternatively, R′ = {±iq mod p : 1 ≤ i ≤ (p − 1)/2}, where exactly m of the elements have the minus sign.
• Take the product of all elements in the two representations of R′.
• So
[(p − 1)/2]! = (−1)mq(p−1)/2[(p − 1)/2]! mod p.
• Because gcd([(p − 1)/2]!, p) = 1, the above implies 1 = (−1)mq(p−1)/2 mod p.
Legendre’s Law of Quadratic Reciprocity
a• Let p and q be two distinct odd primes.
• The next result says their Legendre symbols are distinct if and only if both numbers are 3 mod 4.
Lemma 64 (Legendre (1785), Gauss)
(p|q)(q|p) = (−1)p−12 q−12 .
aFirst stated by Euler in 1751. Legendre (1785) did not give a correct proof. Gauss proved the theorem when he was 19. He gave at least 8 different proofs during his life. The 152nd proof appeared in 1963.
A computer-generated formal proof was given in Russinoff (1990). As of 2008, there have been 4 such proofs. According to Wiedijk (2008),
“the Law of Quadratic Reciprocity is the first nontrivial theorem that a student encounters in the mathematics curriculum.”
The Proof (continued)
• Sum the elements of R′ in the previous proof in mod2.
• On one hand, this is just ∑(p−1)/2
i=1 i mod 2.
• On the other hand, the sum equals
mp +
(p∑−1)/2 i=1
(
iq − p
⌊iq p
⌋)
mod 2
= mp +
q
(p−1)/2∑
i=1
i − p
(p−1)/2∑
i=1
⌊iq p
⌋ mod 2.
– m of the iq mod p are replaced by p − iq mod p.
– But signs are irrelevant under mod2.
The Proof (continued)
• Ignore odd multipliers to make the sum equal
m +
(p∑−1)/2 i=1
i −
(p−1)/2∑
i=1
⌊iq p
⌋ mod 2.
• Equate the above with ∑(p−1)/2
i=1 i mod 2 to obtain m =
(p∑−1)/2 i=1
⌊iq p
⌋
mod 2.
The Proof (concluded)
• ∑(p−1)/2
i=1 ⌊iqp ⌋ is the number of integral points below the line
y = (q/p) x for 1 ≤ x ≤ (p − 1)/2.
• Gauss’s lemma (p. 531) says (q|p) = (−1)m.
• Repeat the proof with p and q reversed.
• Then (p|q) = (−1)m′, where m′ is the number of integral points above the line y = (q/p) x for 1 ≤ y ≤ (q − 1)/2.
• As a result, (p|q)(q|p) = (−1)m+m′.
• But m + m′ is the total number of integral points in the
Eisenstein’s Rectangle
(p,q)
(p - 1)/2 (q - 1)/2
Above, p = 11 and q = 7.
The Jacobi Symbol
a• The Legendre symbol only works for odd prime moduli.
• The Jacobi symbol (a | m) extends it to cases where m is not prime.
• Let m = p1p2 · · · pk be the prime factorization of m.
• When m > 1 is odd and gcd(a, m) = 1, then (a| m) =
∏k i=1
(a | pi).
– Note that the Jacobi symbol equals ±1.
– It reduces to the Legendre symbol when m is a prime.
• Define (a | 1) = 1.
Properties of the Jacobi Symbol
The Jacobi symbol has the following properties, for arguments for which it is defined.
1. (ab | m) = (a | m)(b | m).
2. (a| m1m2) = (a| m1)(a | m2).
3. If a = b mod m, then (a | m) = (b | m).
4. (−1 | m) = (−1)(m−1)/2 (by Lemma 63 on p. 531).
5. (2| m) = (−1)(m2−1)/8.a
6. If a and m are both odd, then (a| m)(m | a) = (−1)(a−1)(m−1)/4.
aBy Lemma 63 (p. 531) and some parity arguments.
Properties of the Jacobi Symbol (concluded)
• These properties allow us to calculate the Jacobi symbol without factorization.
• This situation is similar to the Euclidean algorithm.
• Note also that (a | m) = 1/(a | m) because (a | m) = ±1.a
aContributed by Mr. Huang, Kuan-Lin (B96902079, R00922018) on December 6, 2011.
Calculation of (2200 |999)
(202|999) = (2|999)(101|999)
= (−1)(9992−1)/8(101|999)
= (−1)124750(101|999) = (101|999)
= (−1)(100)(998)/4
(999|101) = (−1)24950(999|101)
= (999|101) = (90|101) = (−1)(1012−1)/8(45|101)
= (−1)1275(45|101) = −(45|101)
= −(−1)(44)(100)/4
(101|45) = −(101|45) = −(11|45)
= −(−1)(10)(44)/4(45|11) = −(45|11)
= −(1|11) = −1.
A Result Generalizing Proposition 10.3 in the Textbook
Theorem 65 The group of set Φ(n) under multiplication mod n has a primitive root if and only if n is either 1, 2, 4, pk, or 2pk for some nonnegative integer k and and odd
prime p.
This result is essential in the proof of the next lemma.
The Jacobi Symbol and Primality Test
aLemma 66 If (M|N) = M(N−1)/2 mod N for all M ∈ Φ(N), then N is a prime. (Assume N is odd.)
• Assume N = mp, where p is an odd prime, gcd(m, p) = 1, and m > 1 (not necessarily prime).
• Let r ∈ Φ(p) such that (r | p) = −1.
• The Chinese remainder theorem says that there is an M ∈ Φ(N) such that
M = r mod p, M = 1 mod m.
aMr. Clement Hsiao (B4506061, R88526067) pointed out that the text- book’s proof for Lemma 11.8 is incorrect in January 1999 while he was a senior.
The Proof (continued)
• By the hypothesis,
M(N−1)/2 = (M | N) = (M | p)(M | m) = −1 mod N.
• Hence
M(N−1)/2 = −1 mod m.
• But because M = 1 mod m,
M(N−1)/2 = 1 mod m, a contradiction.
The Proof (continued)
• Second, assume that N = pa, where p is an odd prime and a ≥ 2.
• By Theorem 65 (p. 544), there exists a primitive root r modulo pa.
• From the assumption, MN−1 =
[
M(N−1)/2 ]2
= (M|N)2 = 1 mod N for all M ∈ Φ(N).
The Proof (continued)
• As r ∈ Φ(N) (prove it), we have
rN−1 = 1 mod N.
• As r’s exponent modulo N = pa is ϕ(N ) = pa−1(p − 1), pa−1(p − 1) | (N − 1),
which implies that p| (N − 1).
• But this is impossible given that p | N.
The Proof (continued)
• Third, assume that N = mpa, where p is an odd prime, gcd(m, p) = 1, m > 1 (not necessarily prime), and a is even.
• The proof mimics that of the second case.
• By Theorem 65 (p. 544), there exists a primitive root r modulo pa.
• From the assumption, MN−1 =
[
M(N−1)/2 ]2
= (M|N)2 = 1 mod N for all M ∈ Φ(N).
The Proof (continued)
• In particular,
MN−1 = 1 mod pa (13)
for all M ∈ Φ(N).
• The Chinese remainder theorem says that there is an M ∈ Φ(N) such that
M = r mod pa, M = 1 mod m.
• Because M = r mod pa and Eq. (13),
N−1 a
The Proof (concluded)
• As r’s exponent modulo N = pa is ϕ(N ) = pa−1(p − 1), pa−1(p − 1) | (N − 1),
which implies that p| (N − 1).
• But this is impossible given that p | N.
The Number of Witnesses to Compositeness
Theorem 67 (Solovay and Strassen (1977)) If N is an odd composite, then (M|N) = M(N−1)/2 mod N for at most half of M ∈ Φ(N).
• By Lemma 66 (p. 545) there is at least one a ∈ Φ(N) such that (a|N) ̸= a(N−1)/2 mod N .
• Let B = {b1, b2, . . . , bk} ⊆ Φ(N) be the set of all distinct residues such that (bi|N) = b(Ni −1)/2 mod N .
• Let aB = {abi mod N : i = 1, 2, . . . , k}.
• Clearly, aB ⊆ Φ(N), too.
The Proof (concluded)
• |aB| = k.
– abi = abj mod N implies N | a(bi − bj), which is
impossible because gcd(a, N ) = 1 and N > |bi − bj|.
• aB ∩ B = ∅ because
(abi)(N−1)/2 = a(N−1)/2b(Ni −1)/2 ̸= (a|N)(bi|N) = (abi|N).
• Combining the above two results, we know
| B |
ϕ(N ) ≤ | B |
| B ∪ aB | = 0.5.
1: if N is even but N ̸= 2 then
2: return “N is composite”;
3: else if N = 2 then
4: return “N is a prime”;
5: end if
6: Pick M ∈ {2, 3, . . . , N − 1} randomly;
7: if gcd(M, N ) > 1 then
8: return “N is composite”;
9: else
10: if (M|N) = M(N−1)/2 mod N then
11: return “N is (probably) a prime”;
12: else
13: return “N is composite”;
14: end if
Analysis
• The algorithm certainly runs in polynomial time.
• There are no false positives (for compositeness).
– When the algorithm says the number is composite, it is always correct.
• The probability of a false negative is at most one half.
– Suppose the input is composite.
– The probability that the algorithm says the number is a prime is ≤ 0.5 by Theorem 67 (p. 552).
• So it is a Monte Carlo algorithm for compositeness.
The Improved Density Attack for compositeness
All numbers < N
Witnesses to compositeness of
N via Jacobi Witnesses to
compositeness of N via common
factor
Randomized Complexity Classes; RP
• Let N be a polynomial-time precise NTM that runs in time p(n) and has 2 nondeterministic choices at each step.
• N is a polynomial Monte Carlo Turing machine for a language L if the following conditions hold:
– If x ∈ L, then at least half of the 2p(n) computation paths of N on x halt with “yes” where n = | x |.
– If x ̸∈ L, then all computation paths halt with “no.”
• The class of all languages with polynomial Monte Carlo TMs is denoted RP (randomized polynomial time).a
aAdleman and Manders (1977).
Comments on RP
• In analogy to Proposition 35 (p. 306), a “yes” instance of an RP problem has many certificates (witnesses).
• There are no false positives.
• If we associate nondeterministic steps with flipping fair coins, then we can cast RP in the language of
probability.
– If x ∈ L, then N(x) halts with “yes” with probability at least 0.5 .
– If x ̸∈ L, then N(x) halts with “no.”
Comments on RP (concluded)
• The probability of false negatives is ϵ ≤ 0.5.
• But any constant between 0 and 1 can replace 0.5.
– Repeat the algorithm k = ⌈−log1
2ϵ⌉ times and answer
“yes” only if all runs answer “yes.”
– The probability of false negatives becomes ϵk ≤ 0.5.
• In fact, ϵ can be arbitrarily close to 1 as long as it is at most 1 − 1/q(n) for some polynomial q(n).
– −log1
2ϵ = O(1−ϵ1 ) = O(q(n)).
Where RP Fits
• P ⊆ RP ⊆ NP.
– A deterministic TM is like a Monte Carlo TM except that all the coin flips are ignored.
– A Monte Carlo TM is an NTM with extra demands on the number of accepting paths.
• compositeness ∈ RP;a primes ∈ coRP;
primes ∈ RP.b
– In fact, primes ∈ P.c
• RP ∪ coRP is an alternative “plausible” notion of efficient computation.
aRabin (1976) and Solovay and Strassen (1977).
ZPP
a(Zero Probabilistic Polynomial)
• The class ZPP is defined as RP ∩ coRP.
• A language in ZPP has two Monte Carlo algorithms, one with no false positives and the other with no false
negatives.
• If we repeatedly run both Monte Carlo algorithms, eventually one definite answer will come (unlike RP).
– A positive answer from the one without false positives.
– A negative answer from the one without false negatives.
aGill (1977).
The ZPP Algorithm (Las Vegas)
1: {Suppose L ∈ ZPP.}
2: {N1 has no false positives, and N2 has no false negatives.}
3: while true do
4: if N1(x) = “yes” then
5: return “yes”;
6: end if
7: if N2(x) = “no” then
8: return “no”;
9: end if
10: end while
ZPP (concluded)
• The expected running time for the correct answer to emerge is polynomial.
– The probability that a run of the 2 algorithms does not generate a definite answer is 0.5 (why?).
– Let p(n) be the running time of each run of the while-loop.
– The expected running time for a definite answer is
∑∞ i=1
0.5iip(n) = 2p(n).
• Essentially, ZPP is the class of problems that can be solved, without errors, in expected polynomial time.
Large Deviations
• Suppose you have a biased coin.
• One side has probability 0.5 + ϵ to appear and the other 0.5 − ϵ, for some 0 < ϵ < 0.5.
• But you do not know which is which.
• How to decide which side is the more likely side—with high confidence?
• Answer: Flip the coin many times and pick the side that appeared the most times.
• Question: Can you quantify the confidence?
The Chernoff Bound
aTheorem 68 (Chernoff (1952)) Suppose x1, x2, . . . , xn are independent random variables taking the values 1 and 0 with probabilities p and 1 − p, respectively. Let X = ∑n
i=1 xi. Then for all 0 ≤ θ ≤ 1,
prob[ X ≥ (1 + θ) pn ] ≤ e−θ2pn/3.
• The probability that the deviate of a binomial random variable from its expected value
E[ X ] = E
[ n
∑
i=1
xi ]
= pn decreases exponentially with the deviation.
aHerman Chernoff (1923–). The bound is asymptotically optimal.
The Proof
• Let t be any positive real number.
• Then
prob[ X ≥ (1 + θ) pn ] = prob[ etX ≥ et(1+θ) pn ].
• Markov’s inequality (p. 503) generalized to real-valued random variables says that
prob [
etX ≥ kE[ etX ]]
≤ 1/k.
• With k = et(1+θ) pn/E[ etX ], we have
prob[ X ≥ (1 + θ) pn ] ≤ e−t(1+θ) pnE[ etX ].
The Proof (continued)
• Because X = ∑n
i=1 xi and xi’s are independent, E[ etX ] = (E[ etx1 ])n = [ 1 + p(et − 1) ]n.
• Substituting, we obtain
prob[ X ≥ (1 + θ) pn ] ≤ e−t(1+θ) pn[ 1 + p(et − 1) ]n
≤ e−t(1+θ) pnepn(et−1) as (1 + a)n ≤ ean for all a > 0.
The Proof (concluded)
• With the choice of t = ln(1 + θ), the above becomes prob[ X ≥ (1 + θ) pn ] ≤ epn[ θ−(1+θ) ln(1+θ) ].
• The exponent expands to
−θ2
2 + θ3
6 − θ4
12 + · · · for 0 ≤ θ ≤ 1.
• But it is less than
−θ2
2 + θ3
6 ≤ θ2 (
−1
2 + θ 6
)
≤ θ2 (
−1
2 + 1 6
)
= −θ2 3 .
Power of the Majority Rule
From prob[ X ≤ (1 − θ) pn ] ≤ e−θ2pn/2 (prove it):
Corollary 69 If p = (1/2) + ϵ for some 0 ≤ ϵ ≤ 1/2, then prob
[ n
∑
i=1
xi ≤ n/2 ]
≤ e−ϵ2n/2.
• The textbook’s corollary to Lemma 11.9 seems incorrect.a
• Our original problem (p. 564) hence demands, e.g.,
n ≈ 1.4k/ϵ2 independent coin flips to guarantee making an error with probability ≤ 2−k with the majority rule.
aSee Dubhashi and Panconesi (2012) for many Chernoff-type bounds.
BPP
a(Bounded Probabilistic Polynomial)
• The class BPP contains all languages L for which there is a precise polynomial-time NTM N such that:
– If x ∈ L, then at least 3/4 of the computation paths of N on x lead to “yes.”
– If x ̸∈ L, then at least 3/4 of the computation paths of N on x lead to “no.”
• So N accepts or rejects by a clear majority.
aGill (1977).
Magic 3/4?
• The number 3/4 bounds the probability (ratio) of a right answer away from 1/2.
• Any constant strictly between 1/2 and 1 can be used without affecting the class BPP.
• In fact, as with RP,
1
2 + 1 q(n)
for any polynomial q(n) can replace 3/4 (p. 559).
• The next algorithm shows why.
The Majority Vote Algorithm
Suppose L is decided by N by majority (1/2) + ϵ.
1: for i = 1, 2, . . . , 2k + 1 do
2: Run N on input x;
3: end for
4: if “yes” is the majority answer then
5: “yes”;
6: else
7: “no”;
8: end if
Analysis
• The running time remains polynomial: 2k + 1 times N’s running time.
• By Corollary 69 (p. 569), the probability of a false answer is at most e−ϵ2k.
• By taking k = ⌈ 2/ϵ2 ⌉, the error probability is at most 1/4.
• Recall that ϵ can be any inverse polynomial.
• So k remains a polynomial in n.
Aspects of BPP
• BPP is the most comprehensive yet plausible notion of efficient computation.
– If a problem is in BPP, we take it to mean that the problem can be solved efficiently.
– In this aspect, BPP has effectively replaced P.
• (RP ∪ coRP) ⊆ (NP ∪ coNP).
• (RP ∪ coRP) ⊆ BPP.
• Whether BPP ⊆ (NP ∪ coNP) is unknown.
• But it is unlikely that NP ⊆ BPP (see p. 591 and
coBPP
• The definition of BPP is symmetric: acceptance by clear majority and rejection by clear majority.
• An algorithm for L ∈ BPP becomes one for ¯L by reversing the answer.
• So ¯L ∈ BPP and BPP ⊆ coBPP.
• Similarly coBPP ⊆ BPP.
• Hence BPP = coBPP.
• This approach does not work for RP.a
aIt did not work for NP either.
BPP and coBPP
Ø\HVÙ ØQRÙ ØQRÙ Ø\HVÙ
BPP and P
Theorem 70 (Nisan and Wigderson (1994)) If every language in BPP only needs a pseudorandom generator which stretches a random seed of logarithmic length, then BPP = P.
• We only need to show BPP ⊆ P.
• Run the BPP algorithm for each of the seeds.
– There are only 2O(log n) = O(nc) seeds, a polynomial
• Accept if and only if at least 3/4 of the outcomes is a
“yes.”
• The running time is clearly deterministically polynomial.
“The Good, the Bad, and the Ugly”
P BPP ZPP
RP coRP
NP coNP