Suppose a quantum computer receives an unknown state.
What should it do with it?
“Machine learning” is too broad an answer.
A useful way to understand the QQ regime is to look at concrete tasks.
Can the machine recognize the state?
Compress it?
Find its dominant structure?
Learn how it evolves?
Generate similar states?
These questions show what quantum-native learning might actually mean.
Prefer to watch? Continue with the video below, or keep reading.
Task 1: Quantum State Discrimination
Start with one of the simplest learning problems.
A source produces one of several possible quantum states.
Which one is it?
For example:
\[ \rho\in\{\rho_0,\rho_1\}. \]
Quantum state discrimination asks for the best measurement strategy to identify the state.
This already reveals something important.
The optimal measurement is not always equivalent to measuring each subsystem independently and then classifying the classical results.
Quantum measurement itself can be designed as part of the classifier.
The distinction between “model” and “measurement” starts to disappear.
For two states with known prior probabilities, the best achievable minimum-error discrimination is characterized by the Helstrom bound. That bound is a reminder that classification here is not merely classical post-processing: the measurement itself can be optimized as part of the learner, and collective measurements can sometimes outperform strategies that measure pieces independently.
Task 2: Quantum State Classification
Now generalize.
Perhaps the states come from two physical phases.
Or two classes of quantum channels.
Or two families of molecular states.
The goal is not to reconstruct each state completely.
It is to determine the class.
That is often much cheaper.
A quantum classifier may process the input state directly and output one bit:
\[ \rho \rightarrow U \rightarrow \{0,1\}. \]
The scientific question could be compact even when the state space is enormous.
This is one reason quantum learning can be useful without ever producing a full classical description.
Task 3: Learn the Dominant Structure
Classical principal-component analysis looks for directions that capture important variation in data.
A quantum analogue is quantum principal component analysis (qPCA), proposed by Seth Lloyd, Masoud Mohseni, and Patrick Rebentrost in 2014. It asks about the eigenstructure of a density matrix and seeks to access dominant eigenvectors and eigenvalues directly in quantum form.
If
\[ \rho = \sum_j \lambda_j |v_j\rangle\langle v_j|, \]
the important information may lie in the largest eigenvalues and their associated subspaces.
Quantum principal-component analysis was proposed as a way to access such structure directly in quantum form for suitable low-rank states.
The crucial phrase is in quantum form.
The goal need not be to print every component of $|v_j\rangle$ as a classical vector.
The learned subspace may instead be used inside a later quantum computation.
That changes what “learning the representation” means.
Task 4: Quantum Representation Learning
Machine learning often succeeds by finding a compressed representation.
Images become feature vectors.
Text becomes embeddings.
High-dimensional data become latent variables.
Quantum data can also contain redundancy.
Suppose a family of states occupies only a small subspace of a larger Hilbert space.
A quantum autoencoder aims to compress that family into fewer qubits while preserving the relevant information.
Conceptually:
\[ |\psi\rangle \rightarrow U_{\text{enc}} \rightarrow |\text{latent}\rangle \otimes |\text{discard}\rangle. \]
If successful, the latent quantum subsystem is a learned representation.
This is a genuinely quantum version of dimensional reduction.
Task 5: Learn Quantum Dynamics
States are only part of physics.
We also care about how systems change.
Suppose an unknown process implements a channel $\mathcal{E}$:
\[ \rho \rightarrow \mathcal{E}(\rho). \]
A learner may want to infer properties of \[ \rho \rightarrow U \rightarrow \{0,1\}. \]0.
Perhaps the goal is to predict future dynamics.
Identify a Hamiltonian.
Recognize a noise process.
Estimate a spectral property.
Or choose controls that steer the system.
Full process tomography can be expensive.
A quantum learner may target only the structure needed for the task.
This is especially relevant to AI for quantum control.
The system does not need a perfect model of everything.
It needs a model good enough to choose the next action.
Task 6: Learn Similarity
Many machine-learning algorithms depend on similarity.
Are two inputs close?
Do they belong to the same cluster?
In quantum mechanics, state overlaps and fidelities provide natural similarity measures.
A quantum processor can sometimes estimate these relationships directly.
For pure states,
\[ |\langle\psi|\phi\rangle|^2 \]
is a basic measure of overlap.
If the downstream task is itself quantum, keeping similarity information in a quantum representation may avoid unnecessary classical reconstruction.
Task 7: Generative Quantum Learning
Classical generative models learn distributions.
They produce new images, text, or samples resembling the training data.
A quantum generative model can aim at a different target:
produce quantum states or quantum measurement distributions resembling a training family.
This could be useful for:
- state preparation,
- simulation,
- modeling many-body systems,
- learning distributions that arise naturally from quantum devices.
The challenge is both computational and physical.
How do we train such a model?
How do we compare generated quantum states efficiently?
What counts as generalization when the output itself is quantum?
These questions are still developing.
Task 8: Learn for Sensing and Metrology
Now connect QQ to the physical world.
A quantum sensor interacts with a field or parameter.
The sensor state carries information about that interaction.
The conventional pipeline measures the state and then estimates the parameter classically.
But if multiple quantum sensor outputs can be stored and processed coherently, a quantum learner may participate directly in the estimation architecture.
This is where QQ becomes especially interesting for AI4S.
The learner is not analyzing a dataset after the experiment.
It is part of the experiment.
A Common Pattern: Learn Only What You Need
These tasks share a principle.
A quantum learner often should not try to reconstruct the entire quantum object.
Instead, it should extract a property useful for a downstream decision.
Classification.
A subspace.
A parameter.
A control action.
A similarity.
A compact latent state.
This mirrors an important lesson from classical shadows and modern quantum characterization:
Full information is often unnecessary.
Scientific efficiency comes from asking the right question.
The Output Can Be Classical, Quantum, or Both
Quantum learning tasks differ in what they produce.
Some end with a classical label:
\[ \rho\rightarrow y. \]
Others produce a quantum representation:
\[ \rho\rightarrow |\phi(\rho)\rangle. \]
Some feed directly into another quantum algorithm.
This matters because extracting a large quantum object into classical form can destroy the very efficiency we hoped to gain.
A quantum advantage may survive only if the learned representation remains quantum for later use.
The Most Interesting Tasks May Be Native to Quantum Experiments
This brings us back to the larger series.
CQ begins with classical data and asks whether a quantum computer can help.
QQ begins with a quantum system and asks something more natural:
What should we learn before throwing the quantum information away?
That is why the most compelling QQ applications may arise in:
- quantum simulation,
- quantum sensing,
- quantum communication,
- many-body experiments,
- quantum-device characterization.
The data are already quantum.
The scientific question is already quantum.
The learner can meet the system in its native information language.
But Capability Is Not Advantage
A quantum computer may be able to perform these tasks.
That does not automatically mean it should.
The practical question is comparative.
How many copies of the state are needed?
How many gates?
How much quantum memory?
How expensive is the interface?
How strong is the best classical measurement strategy?
Does the advantage survive noise?
This leads to the final question of the series.
Not:
Can QQ work?
But:
Where could QQ provide a real advantage over the best classical alternative?
Next in the series: Where Could QQ Provide a Real Quantum Advantage?
0 comments