Emergence of spacetime geometry from entanglement in quantum many body systems

Can Spacetime Geometries Emerge From a Many-Body Quantum System If We Let
Entanglement Define Distance?
Bobak Hashemi1
Department of Physics, University of California at San Diego, La Jolla, CA 92093
Following work of Brian Swingle [1], we sumarize the basics of applying Entanglement Renormalization repeatedly to a quantum many-body state to create a MERA network. Then we discuss a
correspondence between the entropy of subsystems and a discrete geometry.
Recent progress in real-space renormalization techniques have shown that the entanglement structure of
a many-body quantum system can be used to simplify
the description of the system’s ground state. Much like
knowledge of oscillitory solutions in a classical Hamiltionian allows a significant simplification by considering the
problem under a change to a Fourier decomposition in
the canonical variables.
If a many body lattice quantum system has entanglement on every length scale, then a MERA network description of that state is calculationally advantageous.[2]
Such a structure is common amoungst the ground states
of Hamiltionians which can be written as the sum of local
operators.[5] So the study of ground states of these local
Hamiltonians is often done with the help of the MERA
A MERA network is defined by a quantum state, and
essentially represents the information in the state in a
layered structure, where each layer corresponds to information about correlations or bulk properties on some
length scale. There is a proposed connection between certain quantum field theories in n dimensions and geometries in n + 1 dimensions. MERA networks may provide
the link to study this connection by taking the length
scale as degree of freedom in the geometry and using entanglement entropy to define notions of distance.
The Multiscale Entanglement Renormalization
Ansantz, which seems to have an ambiguous part of
speech, is a name given to the procedure of applying
Entanglement Renormalization to a system repeatedly.
Entanglement Renormalization is technique that grew
out of Density Matrix Renormalization[3].
Density Matrix Renormalization is a technique for
truncating the density matrix of a quantum state in a
many body system with controlled error in calculating
expectation values.[4] Entanglement Renormalization
is essentially Density Matrix Renormalization with
and extra “disentanglement step.” We will start by
first undertstanding Density Matrix Renormalization
In DMRG, we consider a state in an N -body quantum
system. NSo we have a vector, |ψi, in a Hilbert space,
H ∼
= V N , where V can be interpreted as the Hilbert
space for a single particle in this large identical body
system. The dimension of H is then Dim(V )N . We will
construct a projector from |ψi ∈ H to |ψsmall i ∈ Hsmall ,
where the dimension of Hsmall is less than the dimension
of H , this process is called course graining. It is helpful
to consider the example of a 1-dimensional ising chain.
To each lattice site, we can assign the vector space C2 ,
the Hilbert space for spin-half degrees of freedom. In
an N -particle ising chain, the total Hilbert space of the
chain with its group of linear operators is isomorphic to
multilinear maps from N copies of C2 back to N copies
of C2 .
Any vector in H has a representation |ψi =
P )
ψi1 ,i2 ,...,iN |i1 i ⊗ ... ⊗ |iN i, where ψi1 ,i2 ,...,iN form
i1 ,...,iN =1
the components of tensor.
H ∼
= V1 ⊗ V2 ⊗ ... ⊗ Vj ⊗ Vj+1 ⊗ ... ⊗ VN
we can construct an operator U = 1⊗1⊗...⊗w⊗...⊗1,
which acts as the identity on all subspaces of H except
the support of w, supp(w) = Vj ⊗ Vj+1 , which we will
call A for simplicity[7].
The trick here is finding the right operator, w, so that
the minimal amount of information about |ψi is lost while
reducing the dimensionality of Hsmall as much as possible. In DMRG, we construct w by looking at the reduced
density matrix of |ψi in the subspace Vj ⊗ Vj+1 .
We are given a state |ψi and construct the reduced
density operator for some spacially connected[8] region
ρ[A] = TrĀ (|ψihψ|),
where TrĀ is the partial trace over the closure of A.
We can then diagonalize ρ[A] so that
dim(V )2
Pk |kihk|.
Suppose we arrange the way we label the states |ki
so that Pk ≥ Pk+1 , as in we write eq. (1) in descending order. We can truncate ρ[A] by choosing some number, m < (dim(V ))2 , of these vectors to keep in our new
hilbert space. We control the error in this truncation by
choosing m subject to the condition that 1 −
Pk = .
and Ā. If there is maximal entanglement between A and
Ā, then every Pk = (dim(V
))2 , and the renormalization
becomes undefined since we can not order the eigenvalues in a unique way. Outside of this singular effect, we
generally have that more entanglement means we keep a
larger subspace, m is a function of the entanglement of
the state and .[3]
Then we let
Pk |kihk|
Now we can see what happens when we apply U to |ψi.
U |ψi = (1 ⊗ 1 ⊗ ... ⊗
Pk |kihk| ⊗ ... ⊗ 1)|ψi
dim(V )
ψi1 ,i2 ,...,iN Pk Ck,ij ,ij+1 |i1 i ⊗ ... ⊗ |ki...
k=1 i1 ,...,iN =1
... ⊗ ... ⊗ |iN i
This equation looks horrible, but it can also be written
U |ψi =
Pk |ki ⊗ |φk i
What we have done through this process is construct a
new Hilbert space, H 0 , with lower dimensionality. The
degrees of freedom in the region A had total dimension
(dim(V ))2 , but we’ve course grained that space so that
the information content is now represented in only m
dimensions with minimal loss of information. Notice also
that the new Hilbert space is factorizable; H 0 ∼
= V1 ⊗
... ⊗ A ⊗ ...VN , where A is the Hilbert space containing
the vectors |ki.
If we were to cover the entire Hilbert space, H , with
R regions like A,[9] we could do this procedure to each
region. A single iteration of that process is called a course
graining step. For each course graining step, we must
contruct dim(H
dim(A) operators like U , one for each region.
Then we will end up with a new Hilbert space Hsmall
which we can guarentee will be factorizable Hsmall ∼
A1 ⊗ ... ⊗ AR , where each Ai carries the information
content of the ith block. And for |ψsmall i ∈ Hsmall , we
have that
|ψsmall i = U1 U2 ...UR |ψi.
Notice a few key features of the DMRG method. The
reduction in size of the Hilbert space by the course graining of a single block A is determined not only by the number , but also by the amount of entanglement between A
To work around this, we can follow Vidal by introducting another set of operators that act on the state
|ψi before we perform the DMR. Rather than a single
course graining step corresponding to the application of
the operators U , we first apply a set of operators which
are called disentanglers to |ψi. That is, for each of the
R regions Ai , we apply an operator which can be represented as Di = 1 ⊗ ... ⊗ di ⊗ 1, where the support of di
is the region Ai union its boundary in Ā (which I will
call the outer boundary). Additionally di is composed of
unitaries whose support is one site on the boundary of Ai
and the closest site on the outer boundary. The purpose
of d is to reduce the number m in the DMRG step.
Again consider the example of a 1D ising chain with
a block size of 2. A block Ai would then be two sites
somewhere on the chain. We want to consider the block
and its boundary, in this case, that corresponds to the
block Ai and the sites immediately to its left and right.
Then Di = 1 ⊗ di+ ⊗ di− , where for notational simplicity,
we have grouped together the closure of supp(d) rather
than writing the identity on each site, there is no loss of
generality. di− is a unitary operator whose support is the
two left-most sites and di+ is a unitary operator whose
support is the two right-most sites.
The simplest way to understand this process is through
physical intuition. Unitary operators in quantum mechanics correspond to time evolution of a quantum state
without a measurement. We can imagine taking a highly
entangled ising chain and, for instance, applying just the
right magnetic field locally at a pair of sites so that
the particles become disentangled. This of course corresponds to changing state of quantum system, but in
our theoretical treatment, we will know precisely what
untary operators where applied and thus there is no loss
of information from this step. For a more explicit look at
applying the MERA algorithm, check this reference. [2]
After proceding with this disentanglement procedure
for every site on the boundary of the course graining
blocks, we proceed with the density matrix renormalization. This now constitutes a single step of the entanglement renormalization procedure, so that we replace eq.
(2) with
|ψsmall i = U1 U2 ...UR D1 ...DR |ψi.
We can now consider a system where we have repeadly applied the Entanglement Renormalization prescription. For reasons which seem to contradict all English style and rules about parts of speech, this procedure
is called (the) ”Multiscale Entanglement Renormalization Ansantz.” [10]
Again we can drop into the exmple of a 1D Ising chain.
We are free to call the distance between lattice sites 1
unit. When computing expectation values of operators
that act on the lattice, we say that we are computing
properties at a length scale of 1 unit. If we choose a block
size of 2, then each ER iteration combines 2 sites into a
single site. We then say we are working on a length scale
of 2. We will attempt to treat length scale as a sort of
emergent dimension of our quantum many body system.
In this way we have a natural connection between the
number of iterations of ER we have performed and the
length scale which is our extra emergent dimension. If
we call the length scale r, and the number of iterations
m, then r = 2m or more generally, r = a2m for a lattice
spacing of a in arbitrary units. Then we can define a
discrete geometry by taking points to be lattice sites at
various scales.
We then identify the size of a cell at any level with
the entanglement entropy of the sites within that cell.
That is to say that given some cell, we can look at the
entanglement entropy of every site in that cell at each
graining level m. As an upper bound to the entropy of a
block, we can consider the entropy of all traced out sites
along the causal cone of the block. Where the causal
cone is the set of all operators and sites that have been
course grained into the block in question.
If we restrict ourselves to considering the ground state
of a critical Hamiltonian, the scale invariance of the state
means that for each MERA iteration, the disentanglers
will be the same, which is not genrally true for an arbitrary state. In such a system, the entropy of a block in
the bulk is proportional to the length of a minimal curve
that stretches from the original lattice out the block in
question. Following this paradigmn, it can be shown that
there is a discrete geometry generated by the quantum
Transverse Field Ising Model which can be identified with
anti-de Sitter space.
[1] B. Swingle, Phys. Rev. D 86, 065007 (2012), URL http:
[2] G. Evenbly and G. Vidal, Phys. Rev. B 79, 144108
(2009), 0707.1454.
[3] G. Vidal, Physical Review Letters 99, 220405 (2007),
[4] S. R. White, Phys. Rev. Lett. 69, 2863 (1992), URL
[5] Operators with compact support. Intuitevly, operators
which act on a finite region of the lattice sites
[6] Again, notice the ambiguous part of speech
[7] There is no mathematical reason why we should choose
A to span only two component vector spaces, but the
generalization is straightforeward
[8] The spacial connectedness of the region is not truely required in theory. We could equally choose any set of spin
degrees of freedom and do the same process. However, in
practice this would cause a loss of information about the
close range entanglement within subsystems of A.
[9] As in, if we have an N particle ising chain, and course
grain with regions of block size 2, considering reduced
density operators for 2 quibit subsystems, we would have
R = N2 blocks that cover H .
[10] It is lost on the author what the Ansantz is...