Suppose a laboratory prepares an unknown quantum state $\rho$.
You have two choices.
Choice one:
measure it immediately,
store the outcome,
repeat the experiment many times,
then let a classical computer analyze the records.
Choice two:
do not measure immediately.
Keep the quantum state coherent.
Bring it into a quantum processor.
Perhaps combine it with another copy.
Perform a quantum operation.
Only then measure.
The distinction is the heart of the QQ regime.
Prefer to watch? Continue with the video below, or keep reading.
QQ in One Line
QQ means
\[ \boxed{\text{Quantum Data}+\text{Quantum Algorithm}}. \]
The data are quantum before the learning begins.
The learner can process those data coherently.
Only later do we convert the final result into classical information.
The pipeline is
\[ \rho \rightarrow U_{\text{learn}} \rightarrow \text{measurement}. \]
Or, when multiple copies are processed jointly,
\[ \rho^{\otimes N} \rightarrow U_{\text{learn}} \rightarrow \text{measurement}. \]
That looks like a small rearrangement.
It is not.
Why Not Measure First?
This is the natural question.
If the final answer is going to be classical anyway, why not measure the state and analyze the data later?
Because quantum measurement is not a neutral storage operation.
It selects information.
Suppose the state contains relationships that are easier to access through a joint quantum operation across several copies.
If we measure each copy separately first, those coherent relationships may no longer be available.
No classical postprocessing can restore information that the measurement did not preserve.
So the timing of measurement matters.
Quantum Data Are Not Classical Data with More Numbers
It is tempting to think of a quantum state as an enormous hidden vector.
If only we had enough storage, perhaps we could simply read all the amplitudes.
Quantum mechanics does not work that way.
An unknown state cannot be fully read out in one shot.
Measurements return probabilistic outcomes.
Different observables may be incompatible.
The no-cloning theorem prevents arbitrary perfect copying of an unknown state.
Learning about $\rho$ therefore requires repeated access to preparations of the state or to a quantum process that generates it.
Quantum data are physical resources.
They are not ordinary files.
Collective Measurements Change What Can Be Learned
Suppose we receive two copies:
\[ \rho\otimes\rho. \]
A QC strategy measures each copy separately and keeps the classical outcomes.
A QQ strategy can allow the two copies to interact before measurement.
This enables collective measurements.
The measurement acts on the joint quantum state rather than on each sample independently.
For some learning tasks, collective access can reduce the number of samples needed.
That is one of the most important theoretical motivations for QQ.
The advantage may come not from faster arithmetic, but from a more informative way of interrogating the data.
Learning Becomes Part of the Experiment
In QC, we can draw a line:
\[ \text{experiment} \mid \text{classical analysis}. \]
The detector sits on one side.
The computer sits on the other.
In QQ, that boundary moves.
The quantum processor may sit inside the measurement architecture.
The system under study produces quantum information.
The quantum learner transforms that information.
The final detector measures the transformed state.
So computation and measurement become intertwined.
This is a very different view of a quantum computer.
It is not only a machine for solving abstract algorithms.
It can become part of a scientific instrument.
QQ Avoids One CQ Problem
Recall the CQ regime.
Classical data \[ \rho \rightarrow U_{\text{learn}} \rightarrow \text{measurement}. \]0 must be encoded:
\[ x \rightarrow |\phi(x)\rangle. \]
That loading step can be expensive.
QQ starts with quantum data already in quantum form.
There is no need to manufacture a quantum state from a giant classical table.
This is one reason quantum data are often considered a more natural target for quantum machine learning than arbitrary classical datasets.
The input interface is already quantum.
But QQ Creates a New Interface Problem
Where does the quantum state come from?
If it is generated on the same quantum processor, the situation is relatively simple.
But suppose the state comes from:
- a quantum sensor,
- a molecule,
- a remote quantum network,
- another processor,
- a many-body experiment.
Now the quantum information must be transferred into a quantum memory or processor without destroying the coherence we want to exploit.
That is a serious engineering challenge.
One important technology is quantum transduction: coherently converting quantum information between different physical carriers or frequency ranges. For example, superconducting processors naturally operate with microwave-frequency excitations, while long-distance fiber networks favor optical photons. A useful interface must convert between them while adding very little noise.
The elegant arrow
\[ \rho \rightarrow \text{quantum processor} \]
may require quantum transduction, networking, quantum memory, and eventually fault-tolerant interfaces.
So QQ removes the classical data-loading problem but introduces a quantum data-ingestion problem.
What Is the Output?
A QQ learner does not usually need to output a complete classical description of the state.
That would defeat much of the purpose.
Instead, it may answer a targeted question:
Which class does this state belong to?
What is the dominant spectral subspace?
Are these two states similar?
Which channel generated the data?
What control action should be applied next?
The output may be only a few classical bits.
The internal processing can still exploit the full quantum structure.
This is a recurring theme in quantum algorithms:
A quantum state can contain enormous structure even when the final answer is compact.
QQ Is Especially Natural for Quantum Science
Suppose the scientific object is already quantum.
A many-body state.
A channel.
A quantum field mode.
A sensor output.
Then measuring everything immediately into classical records may be an unnecessary restriction.
If the laboratory has coherent quantum processing available, it can ask richer questions before measurement.
This suggests an ambitious future architecture:
\[ \text{quantum system} \rightarrow \text{quantum memory} \rightarrow \text{quantum learner} \rightarrow \text{classical decision} \rightarrow \text{next experiment}. \]
That is AI4S with a quantum section inside the loop.
QQ Does Not Mean “No Classical Computing”
Real systems will still be hybrid.
Classical computers may:
- choose circuits,
- optimize parameters,
- schedule experiments,
- interpret outputs,
- maintain memory,
- coordinate agents.
The distinction is simply that the learning pipeline includes a stage where unmeasured quantum data are processed quantum mechanically.
The classical layer remains essential.
QQ is not a declaration of quantum purity.
It is an architectural choice about where the quantum information is used.
The Right Comparison Is QC vs. QQ
If we want to establish a QQ advantage, the fair competitor is not naive tomography.
It is the best QC pipeline.
That may include:
- adaptive measurements,
- classical shadows,
- compressed sensing,
- strong prior models,
- efficient classical learning.
Only after accounting for those methods should we ask whether coherent processing provides something additional.
This is a high bar.
It should be.
Quantum advantage is meaningful only relative to a serious classical alternative.
The Question Is Now Concrete
We have finally reached the point where the abstract phrase “quantum machine learning” becomes a specific scientific question:
If the data are quantum, what can a quantum learner discover directly that would be harder to learn after measuring first?
To answer that, we need examples.
What can a quantum learner actually learn?
Next in the series: What Can a Quantum Learner Learn Directly?
0 comments