Quantum Ml

Data Re-uploading

Data re-uploading feeds the input into a circuit multiple times, interleaved with trainable layers, letting even a single qubit approximate rich functions.

Encoding more than once

In a plain variational classifier the data enters the circuit once through a feature map, then trainable layers process it. Data re-uploading repeats this: encode the input, apply a trainable block, encode the input again, apply another trainable block, and so on. Each re-upload injects the data afresh, letting later layers build compositions of it.

The single-qubit universality result

Kronos motion — data assimilation

A striking result shows that a single qubit with enough re-upload layers can approximate any continuous function on a bounded domain, acting as a universal approximator. The intuition is that alternating encoding and rotation on the Bloch sphere composes to arbitrarily complex maps, echoing how alternating linear and nonlinear layers make classical networks universal. Entanglement across multiple qubits then extends this to functions of several variables.

The Fourier view

Re-uploading has a clean interpretation as a truncated Fourier series in the input. A model with L encoding repetitions can represent frequencies up to order L; the encoding gates set the available frequencies and the trainable gates set their coefficients. Adding re-uploads adds harmonics, directly enlarging the fittable function space. This is developed in encoding and Fourier expressivity.

python

# Data re-uploading on one qubit (schematic)
def reupload(x, theta, layers):
    for l in range(layers):
        qml.RX(x, wires=0)          # re-encode the input
        qml.RY(theta[l, 0], wires=0) # trainable
        qml.RZ(theta[l, 1], wires=0) # trainable
    return qml.expval(qml.PauliZ(0))

Practical value

It squeezes more capacity from few qubits, easing the qubit-count pressure of NISQ hardware.
It lets shallow-width circuits reach high-frequency functions by trading width for repetition depth.
It pairs naturally with cheap angle encoding, which is easy to repeat.

Costs and limits

Each re-upload adds depth, raising noise exposure, and more trainable layers can push the model toward a barren plateau. There is a sweet spot: enough re-uploads to represent the target frequencies, few enough to keep the circuit trainable and faithful on noisy hardware. Re-uploading is one of the most useful tools in near-term QML, but like every tool it adds to the depth budget that hardware tightly constrains.