Introduction to Quantum Computation

量子計算導論

Topic 4 Quantum algorithms

1
 Preliminary

• The advantages of quantum computers are NOT universal.

• This depends on the appropriate quantum algorithms for particular computational tasks.
• One of the important feature of quantum algorithms is the quantum parallelism.

Classical parallelism:


Uˆ
1
2 n

0  1    2n  1 
Quantum parallelism:

1 ˆ
2 n
U 0  Uˆ 1    Uˆ 2n  1 

• Advantage in the investigation of the global property of the outputs.

• No advantage if, e.g., one wants to know all the details of the outputs. 2
 Quantum parallelism
• A simple example
Classical parallelism:

We need four computational resources.

Quantum parallelism:

We “have” all results with one operation!

• Actually, we do not obtain each single result in hand, but the superposition.
2
p (00)   , f (00)
2
p (01)   , f (01) Probabilistically, rather than
2
p (10)   , f (10) deterministically
2
p (11)   , f (11)
Not necessarily extract each single result with 4 runs!
• Advantage in the investigation of the global property of the outputs.
• No advantage if, e.g., one wants to know all the details of the outputs. 3
 Deutsch-Jozsa algorithm

• Showing the advantages of quantum parallelism and quantum interference.

• Deutsch game:

X   x | x  0,1, 2, 2  1
n
x Constant: f(X) = 0, or f(X) = 1

Task: which f ?
f ( x)  f ( X / 2)  0
Balanced: 
 f ( X / 2)  1

• Example: for n = 3, at worst, Alice will need to query Bob with x = 0, 1, 2, 3, 4 to obtain
f(0), f(1), f(2), f(3), f(4), to determine f.
• For the classical case, at worst, Alice will need to query Bob at least 2n-1 + 1 times to
determine f.

4
 Deutsch-Jozsa algorithm
• Quantum circuit of the Deutsch-Jozsa algorithm:

 0  0 00
  
• n-qubit computational basis in the decimal representation: 
n

1   0 01
 
 n
n 1 n2 0 
  x  x1  2  x2  2    xn 1  21
 xn  2  2  010
n 1  
 x  x1 , x2 , x3 ,  , xn | x  0,1, 2,  , 2   n
      3  011
n
  
 n


 n
 2  1  
111
5
 n
 Deutsch-Jozsa algorithm
• Quantum circuit of the Deutsch-Jozsa algorithm:

• Alice superposes all the queries together:

 0  00 00  1  0  1


n
1
 1  Hˆ  n 0  Hˆ 1  Hˆ 0    Hˆ 0  Hˆ 1 




0  1      0  1   Hˆ 1

n
2n n

0 1
n
2 1

1
2n
 n

0  1  2    2  1  Hˆ 1 
1
2n

x 0
x 
2
• Important technique:
0  1    0  1    0  1   000  001  010  011  100  101  110  111
6
 0 1  2  3  4  5  6  7
 Deutsch-Jozsa algorithm
• Quantum circuit of the Deutsch-Jozsa algorithm:

• Bob performs the operation Uˆ f x, y  x, y 2 f ( x) , f : X  0,1 once. → Quantum parallelism!

1 2n 1
 0 1  1 2n 1
0  2 f ( x)  1 2 f ( x)
2  Uˆ f  1   f
Uˆ x 
2 
   x 
2
2n x 0  2n x 0

0 1
n
1 2 1
  (  1) f ( x)
x 
2
2n x 0

 0 2 0  12 0 0 1

0  2 f ( x)  1  2 f ( x)  2 2 0 1
  (1) f ( x )
2  0 2 1  1 2 1

0 1 2
7
 2 2
 Deutsch-Jozsa algorithm
• Quantum circuit of the Deutsch-Jozsa algorithm:

• Alice determines the type of the function f:

2n 1
0 1 2n 1 2n 1
0 1
1
 (1)   1
  (1) xz  f ( x ) z 

 3  Hˆ n
2  f ( x)
Hˆ n
x   n
2 n
x 0 2 2 z 0 x 0 2
1  2 1  0 1 0 1
n
2n 1 2n 1
1
 n   (1) f ( x )  0    (1) xz  f ( x ) z 

 n
2  x 0  2 2 z 1 x  0 2
2n 1
1 1
  (1) xz z

Hˆ  n x  Hˆ x1  Hˆ x2    Hˆ xn  (1) x1z1  x2 z2  xn zn z1 z2  zn 
  2 n
z1 , z2 ,, zn
 2 n
z 0
n n

• f is constant. → All register qubits are measured to be 0.

→ Quantum interference! 8
• f is balanced. → Some of the register qubits are 1.
 Deutsch-Jozsa algorithm
• Experimental demonstration.

9
Nature 536, 63 (2016)
 Grover’s algorithm

• Also known as quantum search algorithm. It was proposed by Lov Grover in 1996.
• How do you search M objects through N elements? For a classical computer, this can be
represented by a function f:{0, 1,…, N-1}→{0, 1}, which returns value f(x) = 1 if x is a
solution to the search problem, and f(x) = 0 otherwise. Typically, this takes O(N) steps.
• For a quantum computer with a quantum oracle Ô capable of recognizing the solution
to the search problem, then this can be represented with one more oracle qubit as:
Oˆ
x 0  x 0  f ( x)(mod 2) .
• The oracle qubit is |1⟩ if x is a solution, and is |0⟩ otherwise.
• The representation can be more concise if the oracle qubit is prepared in a superposition:
 0  1  Oˆ  0 1  Oˆ
 (1)   x (1)
f (x) f (x)
x  x  x
 2   2 

• Namely, the oracle marks the solutions to the search problem by flipping the phase of
the solution.
• For an N-item search problem with M solutions, a quantum computer need only O( N / M )
steps, faster than a classical one.

10
 Grover’s algorithm
• Schematic of the overall quantum search algorithm:

Grover operator

N 1
1
• The n-qubit state after the Hadamard transform:   Hˆ n
0 
N
 x ,N  2 .
x 0
n

• Step 1 is the oracle Ô performing a reflection about the α-axis:

Step 1

Steps 2~4

• The Grover operator Ĝ consists of four steps. The combined effect of steps 2~4 is
Hˆ  n (2 0 0  Iˆ) Hˆ  n  2    Iˆ → Reflection about the state |ψ〉!
• The overall effect of the Grover operator is written as Ĝ = (2|ψ〉
〈ψ|-Î)Ô. 11
 Grover’s algorithm

• Why this circuit works? This can be explained in a geometric

manner.
1 N 1 1 1 N M M
• Note that    x  " x   ' x 
N x0 N x
 
N x

N N
 
 cos   sin 
2 2
1
 
N M x
" x : not solution
1
 
M x
' x : desired solution
• Step 1 is the oracle Ô performing a reflection about the state |α〉:Oˆ (a   b  )  a   b  .
• Steps 2~4 are equivalent to perform a reflection about the state |ψ〉.
3 3
• The overall effect of Ĝ is Gˆ   (2    Iˆ)Oˆ   cos   sin  .
2 2
• Each Grover iteration rotates |𝜓⟩ an angle θ toward |𝛽 ⟩. Repeated application of Ĝ k-times
takes the initial state |𝜓⟩ to
 2k  1   2k  1 
Gˆ k   cos      sin    .
 2   2 
• Repeating Grover iteration R = CloInt[(arccos M / N ) /  ] times rotates |𝜓⟩ most
approaching |𝛽 ⟩ withing a angle θ/2 ≤ π/4. Measuring the state then yields an solution
to the search problem with probability higher than ½ (not deterministically). 12
 Grover’s algorithm

• A 2-qubit example; N = 4, M = 1.
• The oracles corresponding to each desired solution :
 f (0)  1  f (1)  1  f (2)  1  f (3)  1
   
 f ( x )  0, x  0  f ( x )  0, x  1  f ( x )  0, x  2  f ( x)  0, x  3

• Quantum circuit of the quantum search algorithm:

0

1 3 3 1  /3  /3
• The state before the Grover iteration is   
4 x0
x 
2
 
2
  cos
2
  sin
2
 .

• Single Grover iteration moves |𝜓⟩ to |𝛽 ⟩. The measurement on the top two qubits gives
exactly the desired solution.
13
 Grover’s algorithm
• Experimental demonstrations.

14
Phys. Rev. A 72, 050306(R) (2005) Nat. Commun. 8, 1918 (2017)
 Quantum Fourier transform

• (Classical) discrete Fourier transform:

 x0 , x1 , x2 , , xN 1   N 
F
  y0 , y1 , y2 , , y N 1   N
 1 N 1
1
 y0 
N
 xj  N
( x0  x1  x2    xN 1 )
 j 0

 1 N 1 i 2 j
1 2 2
i 2
2
i ( N 1)
e
i
1 N 1 i 2 jk
 y1  N
xj  ( x0  e x1  e
N N
x2    e N xN 1 )
yk  
N j 0
e N
x j   N j 0 N
 N 1 i 2 2 j 2 2 2
1 1 i 2 i 4 i (2 N  2)
 y2 
 N
e
j 0
N
xj 
N
( x0  e N
x1  e N
x2    e N xN 1 )



15
 Quantum Fourier transform
• n-qubit quantum Fourier transform in the decimal representation:
Orthonormal basis in the decimal representation:  0 , 1 , 2 , 3 , , 2  1 
n

2n 1 i 2 jk
1
j 
F
 Fˆ  j  : e 2n
k
n
2 k 0

2n 1 
n
2 1 
n
2 1 2n 1  1 2n 1 i 2 jk 
 
   x j j 
F
 Fˆ   x j j    x j Fˆ  j    x j  n e 2n
k 
j 0  j 0  j 0 j 0  2 k 0 
2
 2n 1 1
2n 1 i n jk  2n 1
    x j e 2  k   yk k
k 0  j 0 2n  k 0

2n 1 2n 1
  x j j 
  yk kF

j 0 k 0

16
 Quantum Fourier transform
• n-qubit quantum Fourier transform in the binary representation:
2n 1 i 2 j k 1 1 k1 2n1  k2 2n2  kn2 22  k n1 21  kn 20
1 1 i 2 j
j  j1 j2  jn 
F
 Fˆ  j  : e 2n
k    e 2n
k1k2  kn  2 k n 1kn
 2 n
k 0 2 n
k1  0 kn  0
n

1  
1 2 2 1 2n1 22  2 1
i 2 j n i 2 j n i 2 j n i 2 j
Fˆ  j   n / 2 0 00  e
 2
0 01  e 2
010  e 2
011    e 2n
111 
2  
  
  
  
 
  
 n n n n n 
1       
n1 n 2 2
2
i 2 j n
2
i 2 j n
2
i 2 j n
2
i 2 j n   1
i 2 j n 
 n/2  0  e 2 1    0  e 2
1      0  e 2 1    0  e 2 1    0  e 2 1 
2   
 




 
   
      
1 n2 1 n 1
j 2  j j 2  jn1 2  jn j 2  jn1  21  jn
1  j
i 2 n i 2 n1 2 n i 2 2 i 2 1
 n/2  0  e 2
1  0  e 2
1     0  e 2n1
1  0  e 2n
1
2  
      
    
1
    
 n / 2 0  ei 2 (0. jn ) 1  0  ei 2 (0. jn1 jn ) 1    0  ei 2 (0. j2 j3 jn1 jn ) 1 
2
 0  ei 2 (0. j1 j2 j3 jn1 jn ) 1 
2 2 j1  2n j2  2n  2    jn 1  21  jn
i 2 j n  i 2 ( j1  2  j2  2    jn 1  2  jn ) n  i 2
n 1 n2 1
 i 2
2 2 2n 2n 1
• Meaning of digits:
decimal: 1234  1103  2 102  3 101  4 100
0.123  1 101  2 102  3  103
binary: 1101  1 23  1 22  0  21  1 20  13
1 0 1
0.101  1 21  0  22  1 23     0.625 17
2 4 8
 Quantum Fourier transform
• n-qubit quantum Fourier transform in the binary representation:
     
1 2 1 n 1 1
j 2  j j 2  jn1 2  jn j 2  jn1 2  jn
1  j
 i 2 n1 2 n i 2  n2 i 2 1
Fˆ  j1 j2  jn   n / 2  0  e
i 2 n
1  0  e 1  0  e 1     0  e 1
3 n
2 2 2 2
2  
      
    
1
      
 n / 2 0  ei 2 (0. jn ) 1  0  e i 2 (0. jn1 jn ) 1  0  ei 2 (0. jn2 jn1 jn ) 1    0  ei 2 (0. j1 j2  jn ) 1
2

• Quantum circuit of the quantum Fourier transform (up to swap operations):

1 0  1 0  1 0  1 0  1 0 
where Rk   2  R1   2     R2   2  R3   2 
0 e 2 
i k
 0 ei 2   0 1 0 e 2 
i 2
0 e 2 
i 3
       
 1
 H 0  (0 1) j
21/ 2 1 i 2 k 1
• Note that   H jk  1/ 2 ( 0  e 2 1 )  1/ 2 ( 0  ei 2 (0. jk ) 1 )
H 1  1 ( 0  1 ) 2 2
 21/ 2
• We need n(n+1)/2 gates plus n/2 swap gates. → a Θ(n2) algorithm.
18
• Classical FFT needs Θ(n2n) gates.
 Quantum Fourier transform
• Experimental demonstration.

19
Nature 536, 63 (2016)
 Phase estimation

• Given a unitary operator (of any dimension) Û, it can be expressed in terms of the
eigenvalue decomposition Uˆ   ei 2 u u .
u

• Suppose that we have an unknown unitary operator Û (sometimes known as oracles) with
an eigenstate |u〉 (could be also unknown) prepared in advance, such that Uˆ u  ei 2 u .
What is the corresponding phase φ?
• Not that φ lies between 0 and 1; 0 < φ < 1.
• Schematic of the overall phase estimation procedure:

t-qubit register

Second stage

First stage
20
 Phase estimation

• First stage of the phase estimation.

• Output state:
 
1

2n / 2
0  e  
i 2 2t 1 
1  0  e 
i 2 2t 2 
1   
 0  e i 2 2 2 
 
1  0  e i 2 21 
1 
 0  e i 2 20 
1 u 
• Suppose that φ can be exactly expressed with t digits in the binary representation
as φ = 0.φ1φ2…φt-1φt. Then
1
 
2 n/2        
0  ei 2 (0.t ) 1  0  ei 2 (0.t1t ) 1    0  ei 2 (0.2 t 1t ) 1  0  ei 2 (0.12t1t ) 1  u

• In fact, we do not know the exact number of t in advance. The number of qubits t depends
21
on the accuracy we wish to have in the estimation of φ.
 Phase estimation

• Second stage of the phase estimation.

• The input state of the inverse FT is

1
 in 
2n /2
0  e i 2 (0.t )
 
1  0  e i 2 (0.t 1t )
1   0  e 
i 2 (0. 2 t 1t )
1  
0  e i 2 (0.12 t 1t )
1 u .

• The output state of the inverse FT is

 out  1 2 t 1t  u .

• The final measurements on each qubit determine the values of φj’s to be 0 or 1.

1 2 t 1 t
• The desired result is   0.12 t 1t     .
2 22 2t 1 2t 22
 Order-finding

• A very important application of phase estimation algorithm.

• To understand the quantum algorithm for order-finding requires a little background in
number theory.
• Mathematical supplement of number theory:
1. Given positive integers x and n, the result of dividing x by n with remainder q can be
uniquely written as:
x = kn + q.
2. This can be expressed in terms of the modular notation, x equals q modulo n, as:
x = q(mod n).
3. We use the modular notation to indicate that we are working in modular arithmetic in
which we only pay attention to remainders. For example, 2 = 5 = 8 = 11(mod 3).
4. For positive integers x and N, x < N, with no common factors, the order of x modulo
N is the least positive integer r, such that xr = 1(mod N).
5. For example, 51 = 5 (mod 31), 52 = 25 (mod 31), 53 = 1 (mod 31).
6. The order-finding problem is to determine the order for given x and N.
7. [Theorem] 𝑟 ≤ 𝑁.
8. Order-finding is believed to be hard for a classical computer, in the sense that no
algorithm is known to solve the problem using resources polynomial in the O(L) bits
needed to specify the problem, where L  log 2 ( N )   log 2 ( N ) is the number of bits needed
23
to specify N.
 Order-finding

• [Problem] Suppose that x and N are given, we want to find the order r such that xr =
1(mod N).
• The basis of L-qubit register in the decimal representation:
 
 y | y  0,1, 2, , N  1, N , , 2  1   0  0
L
00 , 1  0 01 , 2  010 , 3  011 , , 2  1  111 
      
 
L
 
 L L L L L 
 xy (mod N ) , 0  y  N  1
• Consider the unitary operator: Uˆ y  
 y (mod N ) , N  y  2  1
L

1 r 1 2  s  k  k
• For integer 0 ≤ 𝑠 ≤ 𝑟 − 1, the state us   exp  i  x (mod N ) is an eigenstate
r k 0  r
of Û with eigenvalue exp(i2πs/r).
Example: x = 5, N = 31
2 s 1 2   s 2
1  0 i i 
us   5 (mod 31)  e 3
51
(mod 31)  e 3
5 2
(mod 31) 
3 
2  s1 2  s2
1  i i 
  1(mod 31)  e 3
5(mod 31)  e 3
25(mod 31) 
3 
2 s 1 2  s  2
1  i i 
Uˆ us   51
(mod 31)  e 3
5 2
(mod 31)  e 3
53
(mod 31) 
3 
2 s 1 2  s 1 2  s 2 2  s
i 1  i i  i
e 3 e
3
5 (mod 31)  e 3 5 (mod 31)  1(mod 31)   e 3 us
1 2
24
3 
 Order-finding

• Order-finding problem becomes a phase estimation problem:

s
i 2
xr = 1 (mod N) Uˆ us  e r us

• Application of the phase estimation algorithm:

s s

• There are still three issues:

j
1. Efficient computation of controlled-U2 .
2. Preparation of eigenstate |us〉 requires that we know r.
3. Obtaining r from the estimated phase.

25
 Order-finding

• [Issue 1] Suppose that we know Û|y〉= |xy(mod N), 0 ≤ y ≤ N-1. How can we compute
j
the sequence of controlled-U2 ?
• Modular exponential xz(mod N)
z y  z Uˆ zt 2 Uˆ z3 2 Uˆ z2 2 Uˆ z1 2 y  z x zt 2    x z3 2  x z2 2  x z1 2 y (mod N )
t 1 2 1 0 t 1 2 1 0

t 1
 z3 2 2  z2 21  z1 20
 z x zt 2 y (mod N )  z x z y (mod N ) .

• [Issue 2] How to circumvent the dependence on r of the initial eigenstate preparation?

• Note that 1 r 1
r
 u
s 0
s 1.

• We use t  2 L  1  log 2 2  (2 ) 1  qubits in the first register and prepare the second
register in |1〉.
• Schematic of the overall order-finding procedure:

26
 Order-finding
r 1
1
• The output state of the inverse FT (before performing the measurement) is
r

s 0
s us .
s
i 2
Uˆ us  e r us

s s

r 1 r 1
s 1 1
0 us   s
PhaseEstimation
us ,  s 
r
0 1 
r

s 0
0 us 
PhaseEstimation

r

s 0
s us

• After the measurement, we will obtain certain one 𝜑 ∈ {𝜑 , 𝜑 , 𝜑 , … , 𝜑 }.

• φs is an estimation of the phase φs ≈ s/r accurate to 2L+1 bits, with probability at least (1-
ε)/r, for certain one 0 ≤ s ≤ r-1.
Example: x = 5, N = 31, L = 5, t = 12
0 0 0 0 0
0  0000 0  0   2
 3  4    12  0
2 2 2 2 2
0 1 0 1 1
1  01011  1   2  3  4    12  0.333252...
2 2 2 2 2
1 0 1 0 0
 2  1010 0   2   2  3  4    12  0.666504... 27
2 2 2 2 2
 Order-finding

• [Issue 3] Suppose that we have measured one φs. It is an approximation to the correct s/r
accurate to 2L+1 bits. How can we retrieve r from φs?
• The continued fraction is the representation of a real number by a sequence of integers:
1
 a0 , a1 , , aM  : a0  e.g. 31  2  5  2  1  2  1    2  1
  2, 2,1,1, 2
1 13 13 13 3 1
a1  2 2
1 5 5 1
 1
aM 1
1
2
• Suppose that φs = [a0, a1,…, aM] is a good approximation to s/r, i.e., |(s/r) - φs | ≤ 1/(2r2),
then s/r is a convergent of the continued fraction, i.e., [a0, a1,…, am] for 0 ≤ m ≤ M.
• [Final solution] By cutting the continued fraction for φs at different lengths, we obtain
several convergents for the candidate s’/r’. We then check whether these r’ are the
answer to the order-finding problem xr = 1(mod N).
Example: x = 5, N = 31, L = 5, t = 12
1365 1 1
1  0.333252...   0
1
  0, 3,1365  0  0  0,3  0 
4096 3 3
1365
1365 1 1 1 1 2
 2  0.666504...    0,1,1,1, 682 0,1  0   1 0,1,1  0  1  0,1,1,1  0  1

2048 1 1 2 1 3
1 1
1
1 28
 Factoring

• To compute the product of two integers is (in principle) easy.

• However, given a positive composite integer, how can one efficiently find out the prime
factors? This problem, prime factorization, cannot be solved efficiently with classical
computer.
• It turns out that prime factorization is an important problem from various aspect,
particularly the cryptography and the associated applications.
• Factoring problem is shown to be equivalent to the order-finding problem!
• To see this, we need two theorems.

29
 Factoring
Theorem
Suppose N is an L-bit composite number, and 1 ≤ x ≤ N is a non-trivial
solution to the equation x2 = 1 (mod N), that is, neither x = 1 (mod N) nor
x = -1 = N-1 (mod N). Then at least one of gcd(x-1, N) and gcd(x+1, N) is
a non-trivial factor of N that can be computed using O(L3) operations.

Theorem
Suppose N = p1a1… pmam is the prime factorization of an odd composite
positive integer. Let 1 ≤ x ≤ N-1 be an integer chosen at random, subject
to the requirement that x is co-prime to N. Let r be the order of x modulo
N. Then
p(r is even and xr/2 ≠ -1(mod N)) ≥ 1 – 2-m.

30
 Factoring
Factoring algorithm
1. Determine if N is trivially factorisable.
2. Randomly choose 1 ≤ x ≤ N-1. If gcd(x, N) > 1, return it.
3. To find the order r of x modulo N.
4. If r is even and xr/2 ≠ -1(mod N), then compute gcd(xr/2 -1, N) and
gcd(xr/2 +1, N). Each of these two integers may be a non-trivial factor
of N.
5. If not, repeat step 2.~4.

Example [factoring 91]:

1. Suppose that we wish to factor N = 91.
2. [Step 2] We choose x = 9, gcd(9,91)=1.
3. [Step 3] The order of 9 is 3: 91 = 9 (mod 91), 92 = 81 (mod 91), 93 = 1 (mod 91).
4. [Step 4] The order of 9 is odd. The algorithm fails.
5. [Step 2] We choose x = 4, gcd(4,91)=1.
6. [Step 3] The order of 4 is 6: 43 = 64 (mod 91), 44 = 74 (mod 91), 45 = 23 (mod 91), 46 = 1
(mod 91).
7. [Step 4] The order of 4 is even. Furthermore, 46/2 = 64 ≠ -1(mod 91).
8. [Step 4] The algorithm succeeds, giving gcd(64 -1, 91) = 7.
9. [Successful factorizing] One can find out that 91 = 7 x 13. 31
 Factoring
• Experimental demonstrations of factoring 15 = 3 x 5.

Phys. Rev. Lett. 99, 250504 (2007).

Nature 414, 883-887 (2001).

Phys. Rev. Lett. 99, 250505 (2007).

32
 Factoring
• Experimental demonstrations of factoring 21 = 3 x 7.

Nat. Photonics 6, 773 (2012). Sci. Rep. 11, 16599 (2021).

33
 Amplitude estimation

• Given a unitary operator 𝑈 acting on n+1 qubits such that 𝑈|0⊗ ⊗ |0⟩ =
1 − 𝑎|𝜓 ⟩ ⊗ |0⟩ + 𝑎|𝜓 ⟩ ⊗ |1⟩, where 𝑎 ∈ [0,1] is unknow and |𝜓 ⟩ and |𝜓 ⟩
are two n-qubit (normalized) states, not necessarily orthogonal. How to estimate 𝑎?
• The AE algorithm was first introduced in [Contem. Mathemat. 305, 53-74 (2002)]. It
combines the phase estimation (PE) and the Grover’s algorithm.
• Defining the oracle ˆ ˆ Uˆ † Sˆ ,
Qˆ  US 0 0 where Sˆ 0  Iˆ  2  0 n
 0 n  0 0 and Sˆ0  2 0 n 0 n  Iˆ.
• Quantum circuit for the AE algorithm (notice the similarity to the PE algorithm):

P̂  34
 Amplitude estimation 0  0

Uˆ 0 n 1
• Similarly, the AE algorithm can be understood in a geometric manner.
 a
• Note that Uˆ 0 n  0  1  a  0  0  a  1  1
1  1
 sin  a  0  0  cos  a  1  1 , 0   a  1

• The oracle Qˆ  (USˆ ˆ Uˆ † )( Sˆ ) can be divided into two steps.

0 
0

• The first step Sˆ  Iˆ  2  0 n  0 n  0 0 is a reflection about the state Qˆ  1  1

ˆ ˆ Uˆ † is a reflection about the Qˆ  0
0

|𝜓 ⟩ ⊗ |1⟩, while the second step US  0 Uˆ 0 n 1

0
⊗  a
state 𝑈|0 ⊗ |0⟩.
• The overall effect of the oracle can be understood from the action on
the two basis states Qˆ   1  cos 2   1  sin 2   0
 1 a 1 a 0
 .
ˆ
Q  0  0 =  sin 2 a  1  1  cos 2 a  0  0

• The oracle can also be expressed as an eigenvalue problem

Qˆ    ei 2a  
 ,
ˆ
Q    e
 i 2 a

1 i
with eigenvalues e  i 2a  cos 2 a  i sin 2 a and eigenvectors    0  0  1  1 .
2 2
• Recast the problem into the basis of the two eigenvectors:
i i a
Uˆ 0 n  0 
2
 
e    e  ia   , e  ia  cos  a  i sin  a 35
 Amplitude estimation

• Go back to the quantum circuit

P̂ 

• The first stage of the algorithm is to prepare the eigen state |𝜓 ⟩ with Pˆ 
 0 n 1   .

36
 Amplitude estimation

• Go back to the quantum circuit

P̂ 

• Recall the action of the oracle on the eigenstate: 𝑄 |𝜓 ⟩ = 𝑒 |𝜓 ⟩.

• Output state of the second stage of the algorithm:
 
1
2n / 2
0  e i 2 2m1  a
 
1  0  e 
i 2  2m 2  a
1    0 e i 2 2 2  a
 
1  0  e  
i 2 21  a
1  0  e 
i 2  20  a
1  

1
       
 n / 2 0  ei 2 (0.m ) 1  0  ei 2 (0. m1m ) 1    0  ei 2 (0.2  m1 m ) 1  0  ei 2 (0.12 m1m ) 1    .
2
37
(binary representation of θa in m digits)
 Amplitude estimation

• Go back to the quantum circuit

P̂ 

• The final stage is the inverse FT.

• Input state of the inverse FT is
1
 
2 n/2        
0  ei 2 (0.m ) 1  0  ei 2 (0.m1 m ) 1    0  ei 2 (0. 2  m1m ) 1  0  ei 2 (0.12 m1 m ) 1    ,
while the output state of the inverse FT is  out  1 2  m 1 m .
• The final measurements on each qubit determine the values of θj’s to be 0 or 1.
• The desired result is  a  0.1 2  m 1 m  1   22    mm11   mm . 38
2 2 2 2
 HHL algorithm

• A quantum algorithm for solving linear system of equations proposed by

Harrow, Hassidim, and Lloyd.

• May not necessary efficient to extract full information of the solution 𝑥⃗.
• Typically for solving the expectation value 𝑥⃗ ⋅ 𝑀 ⋅ 𝑥⃗.

39
 HHL algorithm
• A quick review how we solve linear system of equations.
• Linear system of equations: 𝐴𝑥⃗ = 𝑏

• Assumption: 1. 𝐴 is an Hermitian matrix and 2. 𝑏 is a unit vector.

• Diagonalization of 𝐴:

where and

• Solution:

• Solving 𝐴𝑥⃗ = 𝑏 amounts to find out the eigenvalues {𝜆 , 𝜆 , … } of 𝐴.

40
→ Application of quantum phase estimation algorithm.
 HHL algorithm

• Assumption 2.: 𝑏 is a unit vector. → |𝑏⟩ = 𝑏 |1⟩ + 𝑏 |2⟩ + ⋯ is a quantum state.

and are the same quantum states in different basis.

• Now working with the basis given by the eigenvectors of A , {|𝑢 ⟩, |𝑢 ⟩,…}.
|𝑏⟩ = 𝑏 |𝑢 ⟩ + 𝑏 |𝑢 ⟩ + ⋯

𝐴 = (2𝜋𝜆 )|𝑢 ⟩⟨𝑢 | + (2𝜋𝜆 )|𝑢 ⟩⟨𝑢 | + ⋯ → Rescaling the eigenvalues such
that0 ≤ 𝜆 < 1.
• Recall the binary representation of 𝜆 in t digits.
( ) ( ) ( )
𝜆 = 0. 𝜆 𝜆 𝜆 …𝜆 = + + ⋯+ , 𝜆 ,𝜆 ,…,𝜆 ∈ {0,1}

|𝜆 = |𝜆 𝜆 …𝜆

• The goal of the HHL algorithm is to fabricate a state |𝑥 ⟩ = 𝐴 |𝑏⟩ = ∑ |𝑢 .

41
 HHL algorithm

• Schematic of the overall phase HHL algorithm:

Phase estimation Inverse phase estimation

Initial state: |Ψ ⟩ = |0⟩|0⟩⊗ |𝑏⟩ .

|𝑏⟩ = 𝑏 |1⟩ + 𝑏 |2⟩ + ⋯ can be prepared by using the
amplitude encoding with respect to the coefficients {𝑏 , 𝑏 , … }.
Phase estimation: |Ψ ⟩ = 𝑃𝐸{|Ψ ⟩} = |0⟩𝑃𝐸{|0⟩⊗ ∑ 𝑏 |𝑢 } = |0⟩ ∑ 𝑏 |𝜆 |𝑢 .

• Compare with our goal |𝑥 ⟩ = ∑ |𝑢 , we need to put the eigenvalues at the denominator.
• This is achieved by the controlled Ry-rotation subroutine. 42
 HHL algorithm

• Subroutine: controlled Ry-rotation

Phase estimation: |Ψ ⟩ = |0⟩ ∑ 𝑏 |𝜆 |𝑢 .

Recall that |𝜆 = |𝜆 𝜆 …𝜆 , and 𝜆 ,𝜆 ,…,𝜆 ∈ {0,1}.
Controlled Ry-rotation:
( ) ( )
|Ψ ⟩ = ∑ 𝑏 𝑐𝑜𝑠 |0⟩ + 𝑠𝑖𝑛 |1⟩ |𝜆 |𝑢 , where 𝜆⃗( )
𝜃⃗ = 𝜆 𝜃 + ⋯+ 𝜆 𝜃.
• The parameters 𝜃’s are determined by setting
( ) ( )
𝑐𝑜𝑠 |0⟩ + 𝑠𝑖𝑛 |1⟩ = 1− |0⟩ + |1⟩, ∀ 𝑗.
• Note that the 𝜃’s may not necessarily be unique when t is larger than #j.
• 𝐶 is a constant to make 0 ≤ ≤ 1 as larger as possible. 43
 HHL algorithm

• Subroutine: controlled Ry-rotation

Controlled Ry-rotation: |Ψ ⟩ = ∑ 𝑏 1− |0⟩ + |1⟩ |𝜆 |𝑢 .

• Measuring the ancillary qubit and postselecting the outcome 1. Then we have

|Ψ ⟩ = |1⟩ ∑ |𝜆 |𝑢 , with normalization constant 𝑁 = ∑ .

• Finally, the IPE maps all |𝜆 |𝑢 to |0⟩⊗ |𝑢 . Then we have

|Ψ ⟩ = |1⟩|0⟩⊗ ∑ |𝑢 .

44
 HHL algorithm

• Experimental realization

45
 Simulating open quantum systems

• Every quantum gate (single qubit gate or multiqubit gate) provided by quantum computing
platforms, e.g., IBM Q, is an unitary operator.

• Ideally, neither decoherence nor incoherent states on an ideal quantum computer.

• An authentic quantum system inevitably interacts with a surrounding environment, leading

to decoherence. The reduced state is described by a dynamical linear map:
û(  )  TrenvUˆ (    env )Uˆ † .
• Mathematical description: operator-sum representation (Kraus operators).
û(  )   Eˆ k  Eˆ k† .
k

• A prototypical quantum circuit for simulating an open quantum system:

46
 Simulating open quantum systems
Iˆ
• Depolarizing channel: û(  )  p  (1  p ) 
2
• Bloch sphere representation of the effect of the depolarizing channel:

• Operator-sum representation: û(  )  p Xˆ  Xˆ  p Yˆ Yˆ  p Zˆ  Zˆ  (1  p) 

3 3 3

• A prototypical quantum circuit for depolarization channel:

û(  )

47
 Simulating open quantum systems
1 0  0 
• Amplitude damping channel: û(  )  Eˆ1  Eˆ1†  Eˆ 2  Eˆ 2† , Eˆ1   , ˆ
E
 2   
 0 1    0 0 
• Bloch sphere representation of the effect of the amplitude damping channel:

• A prototypical quantum circuit for amplitude damping channel:

 out  Trq[ 2]  2
a c  1 0 
1     2   
c

b   0 0   a  b sin c cos
 2 2
    
 a c sin c cos 0  c  cos  
b cos 2 
2 2  2 2 
 
 c  sin  b sin 2
 b
sin

0
2   2 2 2 2  
  sin 2
1 2
 
 c cos  b
sin

b cos 2

0 2
 2 2 2 2 
 0 0 0 0  48

 Simulating open quantum systems
1 0  0 0 
• Phase damping channel: û(  )  Eˆ1  Eˆ1†  Eˆ 2  Eˆ 2† , Eˆ1   , ˆ
E
 2   
 0 1     0  
• Bloch sphere representation of the effect of the phase damping channel:

• A prototypical quantum circuit for phase damping channel:

 out  Trq[2]  2
a c  1 0 
1     
c

b   0 0   a c cos 
 2
    
 a 0 c cos c sin  c cos  b 
2 2  2 
 
 0 0 0 0 
2      b  
0 b cos 2 sin     sin 2
1 2
 c cos
 2 2 2  2
   b 
 c sin 0 sin  b sin 2  49
 2 2 2
 Simulating open quantum systems
• Further sophisticated simulations.

npj Quantum Inf. 5, 106 (2019).

Total Hamiltonian of a spin chain:

npj Quantum Inf. 6, 1 (2020).

Discretized time steps: Markovian environment
Uˆ (t )  e  Uˆ ( t )Uˆ ( t ) Uˆ ( t )
ˆ
 iHt
engineering
Trotter decomposition
Hˆ  Aˆ  Bˆ ,[ Aˆ , Bˆ ]  0
Uˆ ( t )  e i ( A B ) t  e  iAt e  iBt  Oˆ ( t 2 )
ˆ ˆ ˆ ˆ

non-Markovian AD non-Markovian DP

50
 Hybrid quantum-classical algorithms

• The computing power of near-term quantum computers remains limited due to, e.g.,
noises, qubit lifetime, intrinsic limitations of the physical devices, lack of appropriate
quantum algorithms, …
• An alternative approach is to combine quantum and conventional computers.
• We briefly introduce three examples:
1. Quantum support vector machine
2. Variational quantum eigensolver
Variational quantum algorithms
3. Quantum neural network

51
 Quantum support vector machine
• Support vector machine (SVM) is a well-developed machine algorithm.
The basic idea underlying SVM is to construct a hyperplane to separate the datasets belonging to
distinct classes. More specifically, considering a binary classification task with training dataset
{(𝑥⃗ , 𝑦 )}, where 𝑥⃗ ∈ ℝ is the feature and 𝑦 ∈ {𝑅𝑒𝑑, 𝐵𝑙𝑢𝑒} is the label. Suppose that the data is
linearly separable, then goal of SVM is to find a vector 𝑤 ∈ ℝ and a parameter 𝑤 ∈ ℝ such that
𝑤 𝑥⃗ + 𝑤 ≥ 0, 𝑓𝑜𝑟 𝑦 = 𝑅𝑒𝑑
𝑤 𝑥⃗ + 𝑤 ≤ 0, 𝑓𝑜𝑟 𝑦 = 𝐵𝑙𝑢𝑒
,
meanwhile maximize the margin, which is the distance between the closest training points (the
support vectors) and the separating hyperplane 𝑤 𝑥⃗ + 𝑤 = 0.
• In most practical problems, the datasets are not linear separable. Then we need further
techniques.
• Feature map: mapping the data to higher dimensional parameter space.

• Kernel matrix 𝐾 = [𝑘(𝑥⃗ , 𝑥⃗ )]: estimation of the similarity between data points in the
higher dimensional parameter space without explicitly refer to the feature map.
52
• Then the classifier can be trained based on the kernel matrix 𝐾 .
 Quantum support vector machine
• QSVM was first proposed in 2014. However, one of the critical issues is the assumption
that the data can be efficiently accessed quantum mechanically.

• Quantum feature map was introduced in 2019 to overcome the data accessibility.

• Implementation of quantum feature map with parameterized quantum circuits on IBMQ.

Nature 567, 209 (2019)

53
 Quantum support vector machine
• Advantage of QSVM was proved for a heuristic problem.

Nat. Phys. 17, 1013 (2021)

• Estimating kernel function with quantum circuits.

parametrized quantum circuits: 𝑈(𝑥⃗ )
feature map: |𝑥⃗ ⟩ = 𝑈(𝑥⃗ )|0⊗
quantum kernel: 𝑘 𝑥⃗ , 𝑥⃗ = 𝑥⃗ 𝑥⃗ = 0⊗ |𝑈 (𝑥⃗ )𝑈(𝑥⃗ )|0⊗

• An explicit example: zz feature map.

54
 Variational quantum algorithms

Nat. Rev. Phys. 3, 625 (2021)

• VQAs are hybrid algorithm combining the computing power of quantum and
classical computers.
• VQAs implement parametrized quantum circuits on quantum computers and
optimize the parameters with classical computers.
• Prominent applications include the variational quantum eigensolver and quantum
neural network.

55
 Variational quantum eigensovler

Nat. Commun. 5, 4213 (2014) Phys. Rev. X 6, 031007 (2016)

• Schrödinger equation: an eigenvalue problem 𝐻 |𝜓 = 𝐸 |𝜓

• Trial ansatz with parametrized quantum circuit: |𝜓(𝜃⃗ ) = 𝑈(𝜃⃗ )|0⊗
• Objective function: expectation value 𝑓(𝜃⃗ ) = 𝐻 = 𝜓(𝜃⃗ ) 𝐻 𝜓(𝜃⃗ ) ≥ 𝐸
• Goal: optimizing 𝜃⃗ such that 𝑓(𝜃⃗ ) → 𝐸 56
 Quantum neural network

• Artificial neural network (ANN)

• Quantum neural network (QNN)

57