Quantum Ml

Tensor Networks and Quantum ML

Tensor networks efficiently represent low-entanglement quantum states, serving both as classical machine learning models and as the benchmark quantum methods must beat.

A bridge between quantum and classical

A tensor network factorizes a large high-dimensional array into a contracted web of small tensors. In quantum physics they represent many-body states with limited entanglement using far fewer parameters than the full exponential-size amplitude vector. The best-known example is the matrix product state (MPS), which represents a one-dimensional chain of qubits with a cost that grows with the entanglement across each cut, called the bond dimension.

Tensor networks as classical models

Kronos motion — classical vs quantum

Tensor networks are also standalone classical machine learning models. An MPS can serve as a classifier or generative model, trained by sweeping optimization along the chain in a manner reminiscent of the density-matrix renormalization group. These models are interpretable, their capacity is controlled by the bond dimension, and they run entirely on classical hardware, making them a strong practical baseline.

The simulability connection

If a quantum circuit produces only low-entanglement states, an MPS with modest bond dimension reproduces it efficiently on a classical computer.
This is a primary tool of dequantization: measure the entanglement, and if it stays small, a tensor network imitates the circuit.
Genuine quantum advantage therefore requires states whose entanglement outgrows any feasible bond dimension.

python

# Bond dimension as a simulability gauge (schematic)
# entanglement entropy S across a cut sets required bond dim ~ exp(S)
S = bipartite_entropy(circuit_state, cut=n // 2)
required_bond_dim = int(round(2 ** S))
print('MPS bond dimension needed to simulate:', required_bond_dim)

Higher-dimensional networks

Beyond MPS lie richer geometries: projected entangled pair states (PEPS) for two-dimensional systems, and the multiscale entanglement renormalization ansatz (MERA) for scale-invariant systems. Each captures more entanglement structure at higher contraction cost. The choice of geometry encodes an assumption about how correlations are organized, an inductive bias analogous to network architecture.

Why they matter to QML

Tensor networks play a double role. As models they show that much of what quantum circuits do at low entanglement is achievable classically. As analysis tools they set the entanglement bar a quantum method must clear to claim advantage. Understanding them is essential to reading QML claims honestly: high entanglement that resists tensor-network compression is a necessary, though not sufficient, condition for a real quantum edge.