Quantum computing with kernel methods for machine learning
Methods, systems, and apparatus for quantum machine learning. In one aspect, a method includes obtaining, by a quantum computing device, a training dataset of quantum data points; computing, by the quantum computing device, a kernel matrix that represents a similarity between the quantum data points included in the training dataset, comprising computing a value of a kernel function for each pair of quantum data points in the training dataset, wherein the kernel function is based on reduced density matrices for the quantum data points; and providing, by the quantum computing device, the kernel matrix to a classical processor, wherein the classical processor performs a training algorithm using the kernel matrix to construct a machine learning model.
1. A computer-implemented method comprising:
obtaining, by a quantum computing device, a training dataset of quantum data points;
computing, by the quantum computing device, a kernel matrix that represents similarities among the quantum data points included in the training dataset, wherein the quantum data points are encoded as quantum states of qubits by the quantum computing device, comprising computing, for each pair of quantum data points in the training dataset, a corresponding value of a kernel function, wherein the kernel function is based on reduced density matrices for the quantum data points, wherein the kernel function takes a first and a second quantum data points as input and computes a similarity between the first and the second quantum data points as an element in the kernel matrix, and wherein the reduced density matrices are computed for quantum states of a subset of qubits; and
providing, by the quantum computing device, the kernel matrix to a classical processor.
2. The method of claim 1 , further comprising:
receiving, from the quantum computing device and by the classical processor, the kernel matrix; and
performing, by the classical processor, a training algorithm using the kernel matrix to construct a machine learning model.
3. The method of claim 1 , further comprising:
obtaining, by the quantum computing device, a validation dataset of quantum data points;
computing, by the quantum computing device, new elements of the kernel matrix, wherein the new elements comprise entries representing similarities between the quantum data points in the validation dataset and the quantum data points in the training dataset, wherein the quantum data points in the validation dataset are encoded as quantum states of qubits by the quantum computing device and computing the new elements comprises computing, for each pair of quantum data points in the training dataset and the validation dataset, corresponding values of the kernel function; and
providing, by the quantum computing device, the new elements of the kernel matrix to the classical processor.
4. The method of claim 3 , further comprising processing, by the classical processor, the new elements of the kernel matrix to output predictions for each quantum data point in the validation dataset.
5. The method of claim 1 , wherein the kernel function is based on single-body reduced density matrices for the quantum data points in the training dataset.
6. The method of claim 5 , wherein the kernel function comprises a linear kernel function.
7. The method of claim 6 , wherein the linear kernel function i) takes a first quantum data point and second quantum data point as input, ii) produces a numerical output, and iii) includes a sum of terms, wherein the sum runs over each of N-qubits for N>1 and the summand corresponds to a respective qubit and is equal to the trace of a product of a) a reduced density matrix for the first quantum data point on a subsystem corresponding to the respective qubit b) a reduced density matrix for the second quantum data point on a subsystem corresponding to the respective qubit.
8. The method of claim 6 , wherein the linear kernel function is given by
Q
(
x
i
x
j
)
=
∑
l
Tr
[
Tr
m
≠
l
[
ρ
(
x
i
)
]
]
[
Tr
n
≠
l
[
ρ
(
x
j
)
]
]
where x i , x j represent a first and second quantum data point, l represents an index that runs from 1 to the number of qubits N and labels each qubit, ρ(x i )=|x i x x i | and Tr m≠k [ρ(x i )] represents a 1-reduced density matrix (RDM) on qubit k.
9. The method of claim 6 , wherein computing a value of the kernel function for a pair of quantum data points in the training dataset, the pair comprising a first N-qubit quantum state wirth N>1 and a second N-qubit quantum state with N>1, comprises:
repeatedly and for each qubit index:
computing a 1-reduced density matrix (RDM) for the first N-qubit quantum state on a subsystem corresponding to qubit l, comprising obtaining a copy of an N-qubit quantum system in the first N-qubit quantum state and measuring each qubit in the quantum system except the l-th qubit to obtain a first reduced quantum state of the quantum system;
computing a 1-RDM for the second N-qubit quantum state on a subsystem corresponding to qubit l, comprising obtaining a copy of an N-qubit quantum system in the second N-qubit quantum state and measuring each qubit in the quantum system except the l-th qubit to obtain a second reduced quantum state of the quantum system;
determining a trace of the product of the first reduced quantum state and the second reduced quantum state; and
summing averages of the determined traces for each qubit index.
10. The method of claim 5 , wherein the kernel function comprises a squared exponential kernel function.
11. The method of claim 10 , wherein the squared exponential kernel function i) takes a first quantum data point and second quantum data point as input, ii) produces a numerical output, and iii) includes an exponential function of a sum of terms, wherein the sum runs over each of N-qubits for N>1 and the summand corresponds to a respective qubit and is equal to norm of a) a reduced density matrix for the first quantum data point on a subsystem corresponding to the respective qubit minus b) a reduced density matrix for the second quantum data point on a subsystem corresponding to the respective qubit.
12. The method of claim 10 , wherein the squared exponential kernel function is given by
Q
1
E
(
x
i
x
j
)
=
exp
(
-
γ
∑
l
Tr
m
≠
l
[
ρ
(
x
i
)
]
-
Tr
n
≠
l
[
ρ
(
x
j
)
]
2
)
where x i , x j represent a first and second quantum data point, I represents an index that runs from 1 to the number of qubits N and labels each qubit, ρ(x i )=|x i x i | and Tr m≠k [ρ(x i )] represents a 1-RDM on qubit k.
13. The method of claim 10 , wherein computing a value of the kernel function for a pair of quantum data points in the training dataset, the pair comprising a first N-qubit quantum state with N>1 and a second N-qubit quantum state with N>1, comprises, repeatedly and for each qubit index:
computing a 1-reduced density matrix (RDM) for the first N-qubit quantum state on a subsystem corresponding to qubit l, comprising obtaining a copy of an N-qubit quantum system in the first N-qubit quantum state and measuring each qubit in the quantum system except the l-th qubit to obtain a first reduced quantum state of the quantum system;
computing a 1-RDM for the second N-qubit quantum state on a subsystem corresponding to qubit l, comprising obtaining a copy of an N-qubit quantum system in the second N-qubit quantum state and measuring each qubit in the quantum system except the l-th qubit to obtain a second reduced quantum state of the quantum system;
subtracting the second reduced quantum state from the first reduced quantum state to obtain a third reduced quantum state and determining a norm of the third reduced quantum state; and
summing averages of the determined norms for each qubit index and computing an exponent of the summed averages.
14. The method of claim 1 , wherein the kernel function is based on k-body RDMs for the quantum data points, wherein k is less than a predetermined value.
15. The method of claim 14 , wherein the kernel function comprises a linear kernel function.
16. The method of claim 15 , wherein the linear kernel function i) takes a first quantum data point and second quantum data point as input, ii) produces a numerical output, and iii) comprises a sum of terms, wherein the sum runs over each subset of k-qubits taken from the N qubits and each summand corresponds to a respective subset and is equal to the trace of a product of a) a reduced density matrix for the first quantum data point on a subsystem corresponding to the respective subset of k qubits and b) a reduced density matrix for the second quantum data point on a subsystem corresponding to the respective subset of k qubits.
17. The method of claim 15 , wherein the linear kernel function is given by
Q
L
k
(
x
i
x
j
)
=
∑
K
∈
S
k
(
n
)
Tr
[
Tr
m
∉
K
[
ρ
(
x
i
)
]
]
[
Tr
n
∉
K
[
ρ
(
x
j
)
]
]
where S k (n) represents a set of subsets of k qubits, ρ(x i )=|x i x i | and Tr m≠K [ρ(x i )] represents a k-RDM.
18. The method claim 15 , wherein computing a value of the kernel function for a pair of quantum data points in the training dataset, the pair comprising a first N-qubit quantum state and a second N-qubit quantum state, comprises:
repeatedly and for each set of k-qubits:
computing a k-RDM for the first N-qubit quantum state on a subsystem corresponding to qubits in the set, comprising obtaining a copy of an N-qubit quantum system in the first N-qubit quantum state and measuring each qubit in the quantum system except for the qubits included in the set to obtain a first reduced quantum state of the quantum system;
computing a k-RDM for the second N-qubit quantum state on a subsystem corresponding to qubits in the set, comprising obtaining a copy of an N-qubit quantum system in the second N-qubit quantum state and measuring each qubit in the quantum system except for the qubits included in the set to obtain a second reduced quantum state of the quantum system;
determining a trace of the product of the first reduced quantum state and the second reduced quantum state; and
summing averages of the determined for each set of k qubits.
19. The method of claim 14 , wherein the kernel function comprises an exponential kernel function.
20. The method of claim 19 , wherein the exponential kernel function is given by
Q
s
(
x
i
x
j
)
=
∑
k
=
0
∞
γ
k
k
!
n
k
Q
l
k
(
x
i
x
j
)
=
𝔼
exp
(
γ
N
∑
h
=
1
n
(
9
δ
s
h
i
s
h
j
δ
b
h
i
b
h
j
-
4
)
)
where the expected value E is taken over n s samples from randomly chosen Pauli frames that are measured on a first and second system i and j,
δ
s
h
i
s
h
j
represents a first indicator function for an agreement between a random Pauli measurement result performed independently on the first system i and the second system j and
δ
b
h
i
b
h
j
represents a second indicator function for a measurement basis agreement.
21. The method of claim 20 , wherein computing a value of the kernel function for a pair of quantum data points in the training dataset, the pair comprising a first N-qubit quantum state and a second N-qubit quantum state, comprises repeatedly:
obtaining a first measurement result, comprising measuring each qubit in the first system in a random Pauli basis to obtain values s h i and b h i for the h-th qubit, wherein s h i is either 1 or −1 and by is a random basis X, Y, or Z;
obtaining a second measurement result, comprising measuring each qubit in the second system in a random Pauli basis to obtain values s h j and b h j for the h-th qubit, wherein s h j is either 1 or −1 and b h j is a random basis X, Y, or Z;
comparing, for the h-th qubit in the N qubit system, the first measurement result and second measurement result to determine a value of the first indicator function;
determining, for the h-th qubit in the N qubit system, a value of the second indicator function; and
multiplying, summing and averaging the determined values of the first indicator function and the second indicator function.
22. The method of claim 1 , wherein the quantum data points comprise N-qubit quantum states with N>1.
23. The method of claim 1 , wherein obtaining the training dataset of quantum data points comprises:
receiving a training dataset of classical data points; and
generating the training dataset of quantum data points, comprising embedding each classical data point in a respective quantum state by applying a respective encoding circuit to a reference quantum state.
24. An apparatus comprising:
one or more classical processors; and
one or more quantum computing devices in data communication with the one or more classical processors, wherein the quantum computing hardware comprises:
one or more qubit registers, each qubit register comprising one or more qubits, and
a plurality of control devices configured to operate the one or more qubit registers;
wherein the apparatus is configured to perform operations comprising:
obtaining, by a quantum computing device, a training dataset of quantum data points;
computing, by the quantum computing device, a kernel matrix that represents similarities among the quantum data points included in the training dataset, wherein the quantum data points are encoded as quantum states of qubits by the quantum computing device, comprising computing, for each pair of quantum data points in the training dataset, a corresponding value of a kernel function, wherein the kernel function is based on reduced density matrices for the quantum data points, wherein the kernel function takes a first and a second quantum data points as input and computes a similarity between the first and the second quantum data points as an element in the kernel matrix, and wherein the reduced density matrices are computed for quantum states of a subset of qubits; and
providing, by the quantum computing device, the kernel matrix to a classical processor.