Chapter III: Low-Rank Approximation of Electronic Structure Hamiltonian . 45
3.5 Quantum Resource Estimates
is given by Eq. (3.25), whereπ = π»βπ»0andΞ¨(π0) =π0
0. Ifπ0
0is accurate enough, then the correlation energy (defined as πΈ0
π = πΈ0β πΈ0
π» πΉ, where πΈ0
π» πΉ is the ground state energy obtained using Hartree-Fock) with the first order correction is accurate toO (π2).
Next, let us consider the implementation of the basis rotations. As written, the single-particle basis changesπ(β) can be implemented using π2
β πβπβ
2
Givens rotations [48]. (Details are given later at the end of this section.) These rotations can be implemented efficiently using two-qubit gates on a linearly connected architec- ture [1, 2]. If one takesππ spin symmetry into account, then one can perform basis rotations separately for spin-up and spin-down orbitals, requiring 2 π/22
- 2 (πβπ2β)/2 Givens rotations and a circuit depth (on a linear architecture) of (π+πβ)/2.
Lastly, using a fermionic swap network, the component corresponding to evolution of the pairwise operatorπ(β) can be implemented in π2β
linear nearest-neighbor two-qubit gates, with a two-qubit gate depth of exactly πβ [1].
Altogether, the gate count for implementing the full Trotterized exponentiated op- erator is π2
+Γ
β π
π 2
β πβπβ , π
2 + πβ , π
2
for both the Hamiltonianπ»and the uCC cluster operatorπ. Note that the sum overβstarts from 1 and the initial basis rotation is written explicitly. Forπ», the indicesπare dummy indices and can be ignored.
To realize this algorithm on a near-term device, where the critical cost model is the number of two-qubit gates, one can implement the gates directly in hardware [49], which requiresΓπΏ
β=1
π πβ 4 + π2β
4 βπβ
gates on a linear nearest neighbor architecture, with circuit depthΓπΏ
β=1 π 2 + 3πβ
2 . If decomposing into a standard two-qubit gate set (e.g. CZ or CNOT), the gate count would be three times the above count.
From our analysis of the decomposition steps, we expect the Trotterized Hamiltonian exponential to have a gate count of ππππ‘ π π βΌ O (π3) for fixed molecular size and increasing basis size. (We expect to see O (π2) scaling for asymptotically large molecular sizes, but none of the systems studied here reach that limit.) Furthermore, the divergence from linear scaling we saw earlier with increasing alkane chain length in hβiis no longer as visible due to tails in the distribution in πβ). We expect the Trotterized uCC exponential to have a gate count of ππππ‘ π π βΌ O (π4) for increasing molecular size andO (π3) with increasing basis size. These scalings are verified in Fig. 3.4, which plots the two-body gate counts for Trotterizedπ»0andπ0expontentials with different truncation thresholds.
For a system size of about 50 qubits (believed to roughly be the quantum computer size at which quantum supremacy can be demonstrated [50]), a single Trotter step can be implemented in roughly 4,000 layers of parallel gates, assuming a linear architecture.
Partial basis rotation
Here we will walk through implementing a partial basis rotation, which can be implemented using π2
β πβπβ
2
Givens rotations.
Recall that the single particle basis rotationπ(β) is defined such that
Λ πβ
π =
π
Γ
π=1
ππ π(β)πβ
π πΛπ =
π
Γ
π=1
π(β)π π
β
ππ . (3.27)
According to the Thouless theorem [51], one can perform basis rotations of fermionic operators on quantum computers by applying a series of rotations that act on two rows (π,π) ofπ(β) at a time. These rotations are determined by the angles of the Givens rotations ππ π(ππ π) used to perform a QR decomposition ofπ(β). Because π(β) is unitary, the rotations will yield a diagonal π where the diagonal elements are of magnitude 1, and the QR decomposition only requires performing rotations on the π2
elements below the diagonal. The number of two-body gates used to build the unitary operator is the same as the number of Givens rotations, and can be performed in linear depth on a device with linear connectivity when performed in a particular order [1]. (It is straightforward to extend this to basis rotations of bosonic systems. In fact, the necessary rotations to build the unitary operator would more closely reflect the Givenβs rotations.)
Here, we are interested in performing an approximation basis transformation us- ing O (πβπ) Givens rotations, where πβ β€ π and is the number of eigenvalues retained after making the eigenvalue truncation low-rank approximation. Consider the eigenvalue decomposition of the auxiliary matrix used to obtain Eq. (3.17),
π
Γ
π=1
Lπ π(β)π(β)
ππ =π(β)
π π(β)
ππ , (3.28)
whereL(β) is theβ-th auxiliary vector reshaped into an πΓ π square matrix, and π(β) is the diagonal matrix of the corresponding eigenvalues ordered by decreasing magnitude.
In the eigenvalue truncation of L(β), we only use the πβ largest eigenvalues or, equivalently, the eigenvectors associated with the first πβ columns of π(β). This means the sizes of the rotation matrices are reduced, and the eigenvalue equation becomes
Β― π(β)
β
L(β)πΒ―(β) =πΒ―(β)
π
, (3.29)
where Β―π and Β―π are matrices comprising the first πβ columns of π and diagonal elements ofπ, respectively. If we perform a QR decomposition of Β―π(β), the top left πβ Γ πβ block is diagonal and the lower(π βπβ) rows are zero. Since Β―π(β) is only anπ Γπβ matrix, fewer Givens rotations are needed to perform the decomposition than in the full case. Specifically, it at worst requires π2
β πβπβ
2
rotations. Thus, as stated in the earlier discussion, the quantum circuit needed to implement rotation also contains that same number of gates inπ(β) layers.
In the quantum algorithm proposed in the main text, one rotates from the basis of one Cholesky vector to another. This rotation, explicitly given byπβ (β)π(β+1), can be reduced into a single operation, corresponding to a π(β) Γ π(β+1) matrix. Using the same Givens rotation decomposition as above, we find that the costs are found by making the substitutionsπ(β) β π(β+
1) and π β πβ.