The reduced basis workflow

In this first example we want to provide an introduction to the RB framework as applied to quantum spin systems. We want to see, from start to finish, how to set up a physical model, how to generate a surrogate basis and how to finally compute observable quantities. For that purpose, we cover the three basic steps of the RB workflow:

1. Model setup: We first need to initialize the model Hamiltonian and the associated physical parameters.
2. Offline phase: An assembly strategy and a truth solving method is chosen, with which we generate the RB surrogate and we prepare observables for later measurement.
3. Online phase: Using the surrogate, we measure observables with reduced computational cost.

Let us see, how to perform these steps using ReducedBasis.jl. As a first application, we will explore a canonical model from quantum spin physics, the one-dimensional spin-1/2 XXZ chain

\[H = \sum_{i=1}^{L-1} \big[ S_i^x S_{i+1}^x + S_i^y S_{i+1}^y + \Delta S_i^z S_{i+1}^z \big] - \frac{h}{J} \sum_{i=1}^L S_i^z .\]

To assemble the basis, a greedy algorithm will be used that tries to select as few snapshots as possible to generate a good surrogate. And to keep things simple, we will utilize exact diagonalization techniques to perform the eigenvalue solves to obtain snapshots at the desired parameter points. This means, $H$ will be represented by a (sparse) matrix and the snapshots by vectors of Hilbert space dimension. Alternatively, one could, e.g., provide $H$ and its ground states in a tensor-based format allowing for low-rank approximations, which is reserved for a later example.

Model setup

Let us first set up the parametrized Hamiltonian matrix. In this specific example, we will need some utility functions to generate many-body spin Hamiltonians, but in other applications, possibly not connected to physics, the Hamiltonian setup will of course differ. However, all parametrized Hamiltonians will need to be cast into the form of an affine decomposition

\[H (\bm{\mu}) = \sum_{q=1}^Q \theta_q(\bm{\mu})\, H_q .\]

The corresponding type in ReducedBasis is the AffineDecomposition which, as we will see, will account for both Hamiltonians and other observables that one would want to measure.

Now coming to the XXZ chain, we want to implement the parametrized Hamiltonian matrix, for which we first need a way to create global many-body operators

\[S_i^\gamma = (\otimes^{i-1} I) \otimes \frac{1}{2}\sigma^\gamma \otimes (\otimes^{N-i} I),\]

which in this case are $S=1/2$ operators featuring the Pauli matrices $\sigma^\gamma$. So, let us first define the Pauli matrices as sparse matrices

using LinearAlgebra
using SparseArrays
using ReducedBasis

σx = sparse([0.0 1.0; 1.0 0.0])
σy = sparse([0.0 -im; im 0.0])
σz = sparse([1.0 0.0; 0.0 -1.0]);

and create a function to promote single-site operators to global operators at site i for a many-body system of length L:

function to_global(op::M, L::Int, i::Int) where {M<:AbstractMatrix}
    d = size(op, 1)
    if i == 1
        kron(op, M(I, d^(L - 1), d^(L - 1)))
    elseif i == L
        kron(M(I, d^(L - 1), d^(L - 1)), op)
    else
        kron(kron(M(I, d^(i - 1), d^(i - 1)), op), M(I, d^(L - i), d^(L - i)))
    end
end;

To be able to create an AffineDecomposition, we first need to identify the terms $H_q$ and the coefficient functions $\theta_q(\bm{\mu})$. In our specific case, we can identify the parameter vector $\bm{\mu} = (\Delta, h/J)$ and the associated coefficient function as $\bm{\theta}(\bm{\mu}) = (1, \mu_1, -\mu_2)$. Note that the first entry $\theta_1 = 1$ since the first two summands of $H$ do not have any non-trivial coefficients. Hence we arrive at the following Hamiltonian implementation:

function xxz_chain(L)
    H1 = 0.25 * sum(1:L-1) do i
        to_global(σx, L, i) * to_global(σx, L, i + 1) +
        to_global(σy, L, i) * to_global(σy, L, i + 1)
    end
    H2 = 0.25 * sum(1:L-1) do i
        to_global(σz, L, i) * to_global(σz, L, i + 1)
    end
    H3 = 0.5  * sum(1:L) do i
        to_global(σz, L, i)
    end
    AffineDecomposition([H1, H2, H3], μ -> [1.0, μ[1], -μ[2]])
end;

Using these functions, we initialize a small system of length $L=6$ with a matrix dimension of $2^6 = 64$:

L = 6
H = xxz_chain(L);

Offline phase

Now we can proceed by assembling the reduced basis. To that end we first choose a solver to find the lowest eigenvectors of $H$, for which in this case we use the LOBPCG solver. Since the XXZ model as defined above harbors degenerate ground states at some parameter points, we need to choose the right solver settings to account for that. To obtain the ground-state subspace, with different eigenvalues are being distinguished up to some tolerance tol_degeneracy, we set:

lobpcg = LOBPCG(; tol_degeneracy=1e-4);

Next, we need to restrict our surrogate to a certain domain in the $(\Delta, h/J)$ parameter space and define a discrete grid of points on that domain. This is achieved, for example, by defining a 2-dimensional regular grid of parameter points using RegularGrid:

Δ = range(-1.0, 2.5; length=40)
hJ = range(0.0, 3.5; length=40)
grid_train = RegularGrid(Δ, hJ);

For reasons of numerical stability, it is important orthogonalize the RB during assembly (or use similar methods to keep the problem well-conditioned). Hence there are different protocols to extend a RB by a new snapshot. A numerically efficient way to realize this is to use QR decomposition methods as implemented in QRCompress. Note that we choose a tolerance tol to discard snapshot vectors that do not significantly contribute to the basis:

qrcomp = QRCompress(; tol=1e-9);

We lastly need to set the parameters for the greedy basis assembly by creating a Greedy object. This includes choosing an error estimate, as well as an error tolerance below which we use to stop the basis assembly:

greedy = Greedy(; estimator=Residual(), tol=1e-3);

With that, we gathered all elements to be able generate the reduced basis:

rbres = assemble(H, grid_train, greedy, lobpcg, qrcomp);

n      max. err    ‖BᵀB-I‖     time      μ
------------------------------------------------------------
1      NaN         5.68e-15     200ms    [-1.0, 0.0]
2      2.33        5.7e-15      226ms    [2.231, 0.538]
3      0.946       5.8e-15     83.0ms    [0.795, 0.897]
4      0.508       5.91e-15    59.5ms    [1.154, 1.795]
5      0.472       6.21e-15    91.2ms    [2.5, 1.077]
6      0.385       6.51e-15     102ms    [0.346, 0.179]
7      0.0727      1.23e-14    75.6ms    [-0.551, 0.0]
8      0.0554      1.93e-14    85.2ms    [-0.462, 0.269]
9      0.0163      9.5e-14     92.9ms    [1.423, 0.538]
10     0.00522     1.61e-13     102ms    [-0.372, 0.449]
11     0.004       3.05e-13     121ms    [1.692, 1.346]
12     0.00151     4.61e-13     135ms    [-0.821, 0.0]
13     0.00039     1.08e-12     150ms    [-0.821, 0.09]
reached residual target

To finish up the offline phase, we want to define an observable, again as an AffineDecomposition and then compress it, to be able to measure it efficiently in the online stage. We will use the magnetization $M = 2L^{-1} \sum_{i=1}^L S_i^z$ that serves as a so-called order parameter to distinguish different phases of the system in the parameter space. Conveniently, the magnetization already is contained in the third term of $H$:

M    = AffineDecomposition([H.terms[3]], [2 / L])
m, _ = compress(M, rbres.basis);

Note that the compression again produces an AffineDecomposition which now contains only the low-dimensional matrices that operate in RB space. In addition to the compressed observable, compress also returns a second decomposition for analysis purposes, which we will not cover in this example and hence did not assign. Since the coefficient $2L^{-1}$ is actually parameter-independent, we can just construct m without passing any parameter point to obtain the constant reduced magnetization matrix:

m_reduced = m();

Online phase

Having assembled a RB surrogate, we now want to scan the parameter domain by measuring observables, in particular the magnetization from above. Fortunately, we have finished all Hilbert-space-dimension dependent steps and only operate in the low dimensional RB space. This allows us to now compute observables on a much finer grid:

Δ_online = range(first(Δ), last(Δ); length=100)
hJ_online = range(first(hJ), last(hJ); length=100)
grid_online = RegularGrid(Δ_online, hJ_online);

Instead of solving for the ground states of $H$, we solve for the lowest eigenvectors of the reduced Hamiltonian in the online phase. Again, we need a solver, which in this case is FullDiagonalization, i.e., a wrapper around LinearAlgebra.eigen. To use degeneracy settings that match lobpcg from above, we can use the matching constructor:

fulldiag = FullDiagonalization(lobpcg);

To compute expectation values on all online grid points — which we mean by "scanning" the parameter domain — it is convenient to use a map:

using Statistics
magnetization = map(grid_online) do μ
    _, φ_rb = solve(rbres.h, rbres.basis.metric, μ, fulldiag)
    mean(u -> abs(dot(u, m_reduced, u)), eachcol(φ_rb))
end;

Finally, we can take a look at the results. Note that, to plot a magnetization heatmap, we need to pass the transposed matrix magnetization', in order to use the rows as the x-axis. In addition to the magnetization, let us also plot the parameter points at which we performed truth solves in the offline stage:

using Plots
hm = heatmap(grid_online.ranges[1], grid_online.ranges[2], magnetization';
             xlabel=raw"$\Delta$", ylabel=raw"$h/J$", title="magnetization",
             colorbar=true, clims=(0.0, 1.0), leg=false)
params = unique(rbres.basis.parameters)
scatter!(hm, [μ[1] for μ in params], [μ[2] for μ in params];
         markershape=:xcross, color=:springgreen, ms=3.0, msw=2.0)

As expected from theory, we observe $L/2+1$ discrete magnetization plateaus, as well as a ferromagnetic ($M=1$) and an antiferromagnetic phase ($M=0$) in the ground state phase diagram.