A quantum device runs an experiment.
A detector clicks.
The quantum state is gone.
What remains is a classical record.
Perhaps it is a bitstring:
\[ 010110\ldots \]
Perhaps a voltage trace.
A photon count.
A syndrome pattern.
A measured expectation value.
Now ordinary computing takes over.
This is the QC regime:
\[ \boxed{\text{Quantum Data Source}+\text{Classical Algorithm}}. \]
It is less glamorous than the idea of a fully quantum learner.
It is also enormously important.
Prefer to watch? Continue with the video below, or keep reading.
Quantum Experiments Naturally Become Classical Data
A quantum state is not usually handed to a data scientist as a vector of amplitudes.
The experiment is repeated.
Measurements are made.
The outcomes are recorded.
We obtain a dataset such as
\[ D=\{m_1,m_2,\ldots,m_N\}. \]
Classical statistics or machine learning then infer something about the underlying system.
This is the normal interface between quantum physics and data analysis.
The data source is quantum.
The retained data are classical.
A Simple Example: Readout Calibration
Suppose a superconducting qubit is prepared in either $|0\rangle$ or $|1\rangle$.
The readout electronics return a noisy analog signal.
Those distributions overlap.
A classifier can learn how to map the measured signal to the most likely quantum state.
The classifier itself is classical.
But it is learning from data generated by a quantum device.
This is QC in its simplest form.
Now scale the problem to many qubits, drifting hardware, correlated noise, and repeated calibration.
Machine learning becomes increasingly useful.
Quantum State Tomography
A more ambitious goal is to learn the state itself.
Quantum state tomography uses measurement statistics to reconstruct a classical description of an unknown quantum state.
For small systems, this is manageable.
For general large systems, full tomography becomes extremely expensive because the number of parameters grows rapidly.
That is why researchers often exploit structure.
Perhaps the state has low rank.
Perhaps it is generated by a known family of circuits.
Perhaps only local observables matter.
The goal is not necessarily to reconstruct everything.
It is to learn enough for the scientific task.
Neural-Network Quantum State Reconstruction
Machine learning provides one route.
Instead of storing an arbitrary table of amplitudes or density-matrix elements, a neural model can learn a structured representation consistent with the measurements.
The model acts as a compressed classical description.
This can be useful when the state belongs to a manageable family.
Again, the quantum system supplies the observations.
The learner is classical.
The challenge is to infer the hidden quantum structure from incomplete measurements.
Classical Shadows: Learn Many Properties Without Learning Everything
One of the most important ideas in modern quantum characterization is that we often do not need full tomography.
In 2020, Hsin-Yuan Huang, Richard Kueng, and John Preskill introduced the classical shadows method: randomized measurements are used to construct a compact classical description—a “shadow”—from which many different observables can later be predicted efficiently.
Suppose we want to estimate many properties of a state.
A naive strategy might reconstruct the entire state first.
That can be wasteful.
Classical shadows instead perform randomized measurements and store a compact classical representation from which many observables can later be estimated.
The conceptual pipeline is
\[ \rho \rightarrow \text{randomized measurements} \rightarrow \text{classical shadow} \rightarrow \text{many predictions}. \]
This is a striking example of how clever measurement design can reduce the amount of quantum data required.
It also establishes an important baseline for QQ.
Before claiming that quantum processing is necessary, we should ask how far an optimized classical measurement-and-inference strategy can already go.
Quantum Sensors Also Live in QC
Imagine a quantum sensor measuring a magnetic field.
The sensor state is quantum.
The environment changes that state.
Then the sensor is measured.
A classical estimator infers the field.
This is QC.
Machine learning can help when the relationship between raw measurement records and the target parameter is complicated.
It can also help choose controls or measurement settings adaptively.
So QC is not just about quantum computers.
It includes quantum metrology and sensing more broadly.
Quantum Error Correction Produces a QC Data Stream
A quantum error-correcting code repeatedly generates syndrome measurements.
Those measurements are classical.
A decoder processes them to infer likely errors.
So even though the protected information is quantum, the decoding problem is often a classical learning problem over data from a quantum source.
The loop is
\[ \text{quantum system} \rightarrow \text{syndrome measurements} \rightarrow \text{classical decoder} \rightarrow \text{correction or logical update}. \]
This is one reason machine learning and quantum error correction fit naturally together.
The Measurement Choice Determines the Dataset
Classical machine learning usually treats the dataset as given.
QC forces us to ask where the dataset came from.
Suppose two scientists receive identical quantum states.
One measures only in the $Z$ basis.
The other uses a carefully designed set of measurements.
They do not obtain equally informative datasets.
The data-collection policy matters.
In quantum experiments, learning begins before the classical data exist.
This is a crucial idea.
The measurement is part of the learning architecture.
Adaptive Measurement Makes QC More Powerful
We can go further.
Instead of choosing all measurements in advance, use the results of early measurements to decide later ones.
The system may learn:
We already know this observable accurately.
The uncertainty is concentrated elsewhere.
Measure this different basis next.
This creates a closed loop:
\[ \text{measure} \rightarrow \text{update classical model} \rightarrow \text{choose next measurement}. \]
That is active learning applied to quantum experiments.
It can dramatically improve efficiency when experiments are expensive.
The Fundamental Limitation: Measurement Discards Information
Despite these powerful strategies, QC has a boundary.
Once the state is measured and only the classical record remains, coherent quantum information that was not captured is gone.
We can store every bit of the measurement record perfectly.
We can run the world's best classical algorithm afterward.
But we cannot generally reconstruct information that the measurement destroyed or never extracted.
This is not a weakness of machine learning.
It is a consequence of quantum measurement.
This Sets Up the QQ Question
Classical shadows demonstrate that measurement can be much smarter than naive tomography.
Adaptive experiments can be smarter still.
So QQ should not be motivated by saying:
Classical methods cannot learn quantum systems.
They can, often remarkably well.
The real question is narrower and more interesting:
Are there learning tasks for which processing the quantum data coherently before measurement provides an advantage that no measurement-first classical pipeline can match?
That is the boundary between QC and QQ.
It is not a difference in branding.
It is a difference in when quantum information is allowed to become classical.
Next in the series: QQ: What Does It Mean to Learn Quantum Data Quantumly?
0 comments