10.1. Quantum States as Tensor Networks
Consider a quantum many body system which consists of physical spins with a local Hilbert
space of dimension , which we will call the physical
dimension, are located at every site of some lattice . This gives rise to
a total Hilbert space of the system where is the total number of sites
in the lattice. A general quantum state in this many-body Hilbert space can be
represented in terms a set of complex coefficients , where , with respect to the computational basis as
The exponential increase in the number of coefficients with the system size means that it is
entirely impossible to store the full state vector of a quantum system of any reasonable
size in this way. For example, a system of spins with has coefficients, which is far more than the number of atoms in the universe.
Instead of directly storing this full state vector, we can alternatively parametrize it as a
tensor network. Consider for example the case . We can then represent the state vector
as a tensor with four indices, where each index corresponds to a
physical spin. The full state is then recovered as
We can now split the full tensor into separate components by consecutively applying the
SVD between pairs of physical indices. For example, splitting out the first index we can
rewrite as
In this expression we can interpret a a matrix, as a
matrix and as a by matrix. The horizontal edge in this
diagram is called a virtual bond and the dimension of this bond is called the bond
dimension. The bond dimension is a measure of the entanglement in the state, and in this
case encodes the amount of entanglement between the first site and the rest of the system.
So far we have not actually done anything significant, since this decomposition of
actually increased the total number of required coefficients, instead of reducing it. The
key point is that we can reduce the number of parameters by truncating to
only keep the largest singular values. This results in a low rank approximation of the
original state, where the quality of the approximation is controlled by the chosen final
bond dimension .
By repeatedly applying this procedure, grouping and splitting indices in the resulting
diagrams and absorbing the bond tensors into the site tensors we can
decompose into a tensor network of any geometry. For example, we can approximate as
the contraction of a square network to end up with a tensor network state of the form
In words, this expression means that for every basis state its
corresponding coefficient in the superposition is obtained by indexing all of the physical
legs pointing downward according to the corresponding physical basis state and contracting
the resulting network.
We can therefore parametrize an arbitrary quantum state in terms of a set of local tensors
, where each of these tensors encodes a number of parameters that is polynomial in
its physical dimension and bond dimensions (which can in principle be different for
every virtual bond). For a general quantum state however, a good tensor network state
approximation requires a bond dimension which scales exponentially with the system size,
meaning that we have not actually gained anything in terms of efficiency. However, it turns
out that for many physically relevant states the bond dimension can be bounded by a constant
independent of the system size, in which case the tensor network representation leads to an
exponential reduction in the number of variational parameters.
10.2. Area Laws and Tensor Network States
To see why this is the case, let us study the entanglement entropy of a tensor network
state. Consider the following two-dimensional network, where all physical indices have a
dimension and we assume all virtual bonds have the same dimension ,
We now want to quantify the entanglement between the shaded region and the
rest of the system for this specific state. To this end, we first recall the formula for the
bipartite entanglement entropy Eq. (6.1), and note that the number of
terms in this expression is determined by the number of non-zero Schmidt coefficients, the
latter of which is referred to as the Schmidt rank. Looking back now at our initial
decomposition of the full state tensor by splitting out its first index above, we see
that the Shchmidt rank is precisely given by the bond dimension across this cut. From
this, you should be able to convince yourself that the maximal entanglement entropy across
this cut is determined by the bond dimension as . Extending this line of
reasoning to our question of the entanglement between the region and the rest
of the system, we see that each virtual leg connecting to the rest of the
system can contribute a term to the entanglement entropy. Therefore we arrive at
where is the size of the boundary of (which in this
two-dimensional case is its perimeter).
Clearly, this tensor network state then naturally obeys an area law for its entanglement
entropy. In our discussion of the
low temparature properties of quantum many body systems however, we have
already seen that low-energy states of locally interacting Hamiltonians obey exactly such an
area law. It is this fact that tensor network states inherently encode area law entanglement
that makes them so well suited for representing low-energy states of quantum systems. They
can only target a tiny corner of the full exponentially large Hilbert space, but this corner
is precisely where the most relevant physics happens. This observation has given rise to a
large family of tensor network states which allow for an efficient parametrization of states
with varying geometries.
Note
An equally important feature of tensor networks is that they, aside from providing an
efficient parametrization of states, also allow for efficient manipulations of these
states. This means that they can be used to compute interesting features of quantum systems,
and can be optimized to target states of specific interest such as ground states and
low-lying excitations. For all of the network geometries depicted above there exist
corresponding algorithms that put them to efficient use, some of which will be highlighted
in future sections of this tutorial.