Why Tensor Networks

A register of $n$ qubits lives in a Hilbert space of dimension $2^n$ . To store a general state we must keep one complex amplitude per basis vector, so the state vector has $2^n$ entries. At $n = 50$ that is over $10^{15}$ complex numbers — more than a petabyte of memory at double precision. Simulating a general quantum computer by brute force is therefore exponentially expensive, and that exponential wall is exactly what makes quantum hardware interesting.

The states we actually care about are special

The crucial observation behind this module is that the exponential is a worst case, not a typical case. The full Hilbert space is enormous, but the states that arise in physics and in many algorithms occupy a vanishingly small, highly structured corner of it. The structure that matters is entanglement.

Consider a product state of $n$ qubits,

|\psi\rangle = |\phi_0\rangle \otimes |\phi_1\rangle \otimes \cdots \otimes |\phi_{n-1}\rangle .

Although it lives in the $2^n$ -dimensional space, it is fully described by $n$ single-qubit states — just $2n$ numbers. There is no need to store $2^n$ amplitudes because the state factorizes. A product state has no entanglement, and its description cost is linear in $n$ .

Real states are not products, but many are weakly entangled: each qubit is strongly correlated only with its neighbours, and the correlations decay with distance. Tensor networks are a data structure that interpolates smoothly between these two extremes. They store a state as a network of small tensors whose size is controlled by a single knob — the bond dimension — that measures how much entanglement the representation can carry.

What a tensor network buys you

The promise of the tensor-network approach is this: if a state's entanglement is bounded, you can represent and manipulate it with cost polynomial in $n$ instead of exponential. Concretely, a one-dimensional chain of $n$ qubits whose entanglement across any cut is bounded by a constant can be written as a matrix product state (MPS) using on the order of $n \chi^2$ numbers, where $\chi$ is the bond dimension. When $\chi$ is small the saving over $2^n$ is dramatic.

A different lens on the same simulator

Throughout this course you have used a state-vector simulator that stores all $2^n$ amplitudes explicitly. That is the honest, exact baseline, and we will keep using it to check small examples. Tensor networks are a compression of that same information: when the compression is lossless (or nearly so) you can push to far larger systems than the state vector allows. The rest of the module builds this idea up carefully — what an MPS is, how bond dimension tracks entanglement, how to contract a network, and where the method finally fails.

Where this leads

By the end of the module you will be able to look at a circuit or a physical state and make an informed estimate of whether a classical tensor-network simulation is feasible. That skill is central to modern quantum computing: it underlies claims of quantum advantage (which require ruling out efficient classical simulation) and it powers the classical tools used to validate quantum hardware up to dozens of qubits.