This open source implementation allows the prediction of many properties: few-body observables, two-point correlation functions, subsystem entanglement entropy, from very few measurements.
We require g++ and python version 3.
On the experimental side, we require single-qubit Pauli measurements (i.e., each measurement measures all qubits in some Pauli X, Y, or Z- basis). This should be readily available in most quantum machines.
An introduction to this procedure and the underlying theory can be found in our papers:
-
Efficient estimation of Pauli observables by derandomization: https://arxiv.org/abs/2103.07510.
-
Predicting Many Properties of a Quantum System from Very Few Measurements: https://arxiv.org/abs/2002.08953, https://www.nature.com/articles/s41567-020-0932-7.
The following is for using a command line interface to run the C++ implementation. The most important code in this repo is the creation of measurement schemes, see data_acquisition_shadow.cpp or data_acquisition_shadow.py.
# Compile the codes
> g++ -std=c++0x -O3 data_acquisition_shadow.cpp -o data_acquisition_shadow
> g++ -std=c++0x -O3 prediction_shadow.cpp -o prediction_shadow
# Generate observables you want to predict
> g++ -O3 -std=c++0x generate_observables.cpp -o generate_observables
> ./generate_observables
# Create measurement scheme (stored in scheme.txt) using derandomized version of classical shadows
> ./data_acquisition_shadow -d 100 generated_observables.txt 1> scheme.txt
# Do the physical experiments
# Store the data in measurement.txt
# Predicting many few-body observables
> ./prediction_shadow -o measurement.txt observables.txt
# Predicting many subsystem entanglement entropy
> ./prediction_shadow -e measurement.txt subsystems.txtSince many people are using Python, we have implemented data_acquisition_shadow.py which is the Python version of data_acquisition_shadow.cpp and prediction_shadow.py which is the Python version of prediction_shadow.cpp. The purpose of the two codes is only to facilitate understanding of the procedure and it could be orders of magnitude slower than the C++ implementation. It can be used through the command line interface
> python data_acquisition_shadow.py -d 10 observables.txt
> python prediction_shadow.py -o measurement.txt observables.txtor by importing into a Python code
import data_acquisition_shadow
import prediction_shadow
# randomized classical shadow consisting of 100 parallel measurements in a 20-qubit system
measurement_procedure = data_acquisition_shadow.randomized_classical_shadow(100, 20)
# measurement_procedure = [a list of 100 parallel measurements, each being [a list of 20 single-qubit Pauli bases]]
print(measurement_procedure)
# full_measurement = [a list of [list of (XYZ basis, +-1 outcome) for each qubit]]
# one_observable = [a list of (Pauli-XYZ, index for the qubit)]
estimate_exp(full_measurement, one_observable)Currently, prediction_shadow.py only support the -o option for prediction expectation value of observables.
In your terminal, perform the following to compile the C++ codes to executable files:
> g++ -std=c++0x -O3 data_acquisition_shadow.cpp -o data_acquisition_shadow
> g++ -std=c++0x -O3 prediction_shadow.cpp -o prediction_shadowThe executable data_acquisition_shadow could be used to produce an efficient measurement scheme for predicting many local properties from very few measurements. There are two ways to use this program:
> ./data_acquisition_shadow -r [number of measurements] [system size]This generates random Pauli measurements. There would be [number of measurements] repetitions on a system with [system size] qubits.
You may then use this set of randomized measurements to perform the experiment.
> ./data_acquisition_shadow -r 5 3
X X Z
Y Z X
Y Z Z
X X Z
Y X ZA random output of 5 measurement repetitions for a system of 3 qubits is outputted.
> ./data_acquisition_shadow -d [measurements per observable] [observable file]If you already know a list of local observables you want to measure, this option creates a derandomized version of the randomized measurement.
We consider local observables as observables that act on few qubits but do not have to be geometrically local.
Furthermore, since every local observables could be decomposed to tensor product of Pauli matrices, we only focus on tensor product of Pauli matrices.
The generated measurement scheme would allow you to measure all the local observables given in [observable file] for at least [measurements per observable] times.
An important note is that: if we increase [measurements per observable], the ratio number of measurement repetitions / [measurements per observable] would not be constant, but would actually decrease.
This is because the derandomization procedure would find a more efficient approach to measure the observables.
The format of the [observable file] is given as follows.
An example of a list of local observables could be found in observables.txt.
[number of qubits / system size]
[k-local] X/Y/Z [ith qubit] X/Y/Z [jth qubit] ...
[k-local] X/Y/Z [ith qubit] X/Y/Z [jth qubit] ...
...
We consider [ith qubit] to take value from 0 to [number of qubits / system size] - 1.
One could also consider Pauli observables with weights given as follows.
[number of qubits / system size]
[k-local] X/Y/Z [ith qubit] X/Y/Z [jth qubit] ... [weight]
[k-local] X/Y/Z [ith qubit] X/Y/Z [jth qubit] ... [weight]
...
We consider [weight] to be a floating point number in between 0.0 to 1.0. A smaller weight means the observable is less important and will be measured less often in the derandomization procedure.
This example consider measuring all the observables in the example file observable.txt at least 1 time. The output is the measurement basis for each repetition interleaved with [Status T: X]. T stands for T-th measurement repetitions and X stands for the minimum number of measurements in all the observables we hope to predict.
# A pre-specified [observable file]: observables.txt
> ./data_acquisition_shadow -d 1 observables.txt
X Y X Y Y X Y X Y X
[Status 1: 0]
X X Y X Y Y X Y X Y
[Status 2: 0]
X X X X X Y X X Y X
[Status 3: 0]
X X X X X X X X X X
[Status 4: 1]The following example consider a much longer list of local observables.
We first generate the file generated_observables.txt consisting of a large number of 3- and 4-local observables.
Then we use data_acquisition_shadow to generate an efficient measurement scheme for measuring all local observables for at least 100 times.
# Automatic generation of [observable file] using Python
> python generate_observables.py
> ./data_acquisition_shadow -d 100 generated_observables.txtAlternatively, we show an example using C++ to generate the file containing the list of observables. C++ is much more efficient than Python, because C++ is a lower-level language.
# Automatic generation of [observable file] using C++
> g++ -O3 -std=c++0x generate_observables.cpp -o generate_observables
> ./generate_observables
> ./data_acquisition_shadow -d 100 generated_observables.txtBecause the generated measurement schemes could be quite long, we provide the following options based on shell commands (1> and 2>).
# Some more advanced options using shell commands
# + Output the measurement scheme to a file
> ./data_acquisition_shadow -d 100 generated_observables.txt 1> scheme.txt
# + Don't output [Status X: X]
> ./data_acquisition_shadow -d 100 generated_observables.txt 2> /dev/null
# + Do both
> ./data_acquisition_shadow -d 100 generated_observables.txt 1> scheme.txt 2> /dev/nullPerform physical experiments using the generated scheme to gather the measurement data. The [measurement file] should be structured as follows. An example of the format is given in measurement.txt.
[system size]
[X/Y/Z for qubit 1] [-1/1 for qubit 1] ...
[X/Y/Z for qubit 1] [-1/1 for qubit 1] ...
...
The first line consists of the number of qubits [system size] in the quantum system.
A single-shot measurement result is recorded in each of the following lines.
A single-shot measurement result consists of whether we measure in X, Y, Z-basis for each qubit, and the corresponding binary measurement outcome.
Advanced tips:
In practice, it may be more economic to run the same measurement schemes for multiple times. Our framework / program could also operate properly in such a scenario.
As the repetition increases, the prediction would also become more accurate.
However, the ratio between the number of measurement bases (Nb) and the number of repetitions (Nr) for each basis is an important hyper-parameter that should be properly tuned.
For local observables, when N = Nb x Nr is fixed, it is preferred to have Nr = 1. However, if changing the bases every time would be less economic, then you could also try a larger Nr > 1.
For entanglement entropy, when N = Nb x Nr is fixed, Nr = 1 may no longer be preferable. We should consider Nr as a hyper-parameter, and try Nr = 1, 2, 4, 8, 16, 32, 64, 128, ... to see which yields the best performance.
The executable prediction_shadow could be used to predict many local properties from the measurement data obtained in Step 3. There are two ways to use this program:
> ./prediction_shadow -o [measurement.txt] [observable.txt]This command allows the prediction of local observables, where the observables act on few qubits but do not have to be geometrically local.
The prediction is based on the measurement data [measurement.txt] obtained in Step 3. An example of the measurement txt file is given in measurement.txt.
The local observables are given in [observable.txt]. The format can be found in Step 1. An example of the observable txt file is given in observables.txt.
> ./prediction_shadow -o measurement.txt observables.txt
0.049015
0.006748
0.014625
-0.005204
0.017512
-0.024036
-0.022046
-0.002698
-0.038271
0.003201
-0.023298
-0.003155
-0.003189
0.002247
1.000000
1.000000This predicts 16 local observables given in observables.txt from the randomized measurements given in measurement.txt.
The randomized measurements are performed on a system of 10 qubits, where two consecutive qubits form a singlet state (a total of 5 singlet states).
> ./prediction_shadow -e [measurement.txt] [subsystem.txt]This command allows the prediction of subsystem entanglement entropy. In particular, we output Renyi entropy of order 2.
[measurement.txt] is the measurement data obtained in Step 3. An example of the measurement txt file is given in measurement.txt.
[subsystem.txt] contains a list of subsystems that we want to predict their entanglement entropy. The format of [subsystem.txt] is
[system size]
[subsystem 1 size] [position of qubit 1] [position of qubit 2] ...
[subsystem 2 size] [position of qubit 1] [position of qubit 2] ...
The first line [system size] indicates the number of qubits in the entire system.
In each of the following line, we specify one subsystem.
[subsystem T size] is the size of the T-th subsystem.
Then we specify what are the positions for the qubits in the T-th subsystem. [position of qubit X] is the position of the X-th qubit in the subsystem.
An example can be found in subsystems.txt.
> ./prediction_shadow -e measurement.txt subsystems.txt
0.000000
2.000000
0.000277
1.996378
1.035726
0.003031This predicts the entanglement entropy for six subsystems given in subsystems.txt from the randomized measurements given in measurement.txt.
The randomized measurements are performed on a system of 10 qubits, where two consecutive qubits form a singlet state (a total of 5 singlet states).