Neural Network Quantum State

A machine learning approach

Roger Luo

A brief intro for Boltzmann machine

This part is in my blog

Chinese version: Boltzmann Machine

A Boltzmann machine, like a Hopefield network, is a network of units with an "energy" defined for the network. It also has binary units, but unlike Hopefield nets, Boltzmann machine units are stochastic. The global energy E, in a Boltzmann machine is identical in form to that of a Hopfield network

E = −(∑i, jωijsisj + ∑iθisi)


  • ωij is the connection strength between unit j and unit i
  • si is the state, si ∈ 0, 1, of unit i
  • θi is the bias of unit i in the global energy function. (θi is the activation threshold for the unit)

Often the weights are represented in matrix form with a symmetric matrix W, with zeros along the diagonal.

Probability of a unit's state

The difference in the global energy that results from a single unit i being 0 (off) versus 1 (on), written ΔEi, assuming a symmetric matrix of weights, is given by:

ΔEi = ∑jωijsj + θi

This can be expressed as the difference of energies of two states:

ΔEi = Ei = off − Ei = on

We then substitute the energy of each state with its relative probability

$$ \begin{aligned} \Delta E_i &= -T\ln(p_{i=off})+T\ln(p_{i=on})\\\\ \frac{\Delta E_i}{T} &= \ln(p_{i=on}) - \ln(1-p_{i=on})\\\\ p_{i=on} &= \frac{1}{1+\exp(-\frac{\Delta E_i}{T})} \end{aligned} $$

Where T is the temperature of the system.

Equilibrium state

  • choosing a unit
  • setting its state according to the above formula
  • Repeat!


Divide units into:

  • visible
  • hidden

The visible units are those which receive information from the 'environment', i.e. our training set is a set of binary vectors over the set V. The distribution over the training set is denoted P+(V).

As is discussed above, the distribution over global states converges as the Boltzmann machine reached thermal equilibrium. We denote this distribution, after we marginalize it over the hidden units, as P(V).

Our goal is to approximate the 'real' distribution P+(V) using the P(V) which will be produced (eventually) by the machine. To measure how similar the two distributions are, we use the Kullback-Leibler divergence (ususlly noted as KL-divergence) KL(+|−):

$$ KL(+|-) = \sum_{v} P^+ (v) \ln(\frac{P^{+}(v)}{P^{-}(v)}) $$

where the sum is over all the possible states of V. KL(+|−) is a function of the weights, since they determine the energy of a state, and the energy determines P(v). Therefore, a gradient decent can be simply applied over KL

Actually, the process of training a Boltzmann machine we discussed above can be considered as a thermal equilibrium process with artificial boundary condition.

(The boundary is determined by visible units) The training method we use is likely a kind of equilibrium statistical mechanics, recent research has been considering the application about nonequilibrium statistical mechanics on unsupervised learning. (see arxiv:1503.03585)

Neural-Network Quantum State

A quantum state can be perfectly mapped to a Boltzmann Machine's probability distribution, which has been widely applied into quantum machine learning algorithms.



P(v, h)|v⟩|h⟩→PModel(v, h|W)

In this work, the quantum state is represented in this form

P(v, h|W)=∑hiejajσjz + ∑ibihi + ∑ijWijhiσjz

where σjz represents the jth spin configuration on z-axis, and it is called the Neural-network Quantum State.

And the quantum state is

|Ψ⟩=∑P(h, v|W)|h⟩|v


Training is accomplished by optimizing the Energy

$$ E = \frac{\langle\Psi|H|\Psi\rangle}{\langle\Psi|\Psi\rangle} $$

Optimization Method

Current Optimization Method we re-implemented is the Stochastic Reconfiguration Method, however according to the author its performance is not good.

SR optimization


Will imaginary evolution work?


Since the performence of a neural network mainly depends on it topological structure rather than optimization method, several possible nets could be applied in calculating ground state inspired by NQS

  • Reinforcement learning methods
  • Deep Belief Nets
  • Deep Boltzmann Machine + Deep Belief Nets

Furthermore, will NQS be expended to Hamiltonian? eg. Convert a complex Hamiltonian to a simple Hamiltonian by Neural nets?

some possible approaches inspired from quantum machine learning algorithms:

  • Quantum Boltzmann machine
  • Fermion Bolzamnn machine