OBSTACLES TO STATE PREPARATION AND VARIATIONAL OPTIMIZATION FROM SYMMETRY PROTECTION
C.4 Classical Simulability of Level- 1 RQAOA for Ising Models
Suppose π½ is a real symmetric matrix of size π. Here we consider Ising-like cost functions such that the corresponding Hamiltonian is
π» = βοΈ
1β€π <πβ€π
π½π,πππππ .
The mean values of a Pauli operatorππππon the level-1 QAOA state
|Ξ¨π»(π½, πΎ)i =ππ π½ π΅ππ πΎ π»|+πi
can be computed in time π(π) using an explicit analytic formula. Such a formula was derived for the Max-Cut cost function by Wang et al. [24, Theorem 1]. Below we provide a generalization to general Ising Hamiltonians. Since the total number of terms in the cost function isπ(π2), simulating each step ofRQAOAtakes time at mostπ(π3). Assuming thatππ=π(1), the number of steps is roughlyπso that the full simulation cost isπ(π4). Crucially, the simulation cost of this method does not depend on the depth of the variational circuit. This is important because RQAOA may potentially increase the depth fromπ(1)toπ(π)since it adds many new terms to the cost function.
Lemma 3.C.1. Fix a pair of qubits 1 β€ π’ < π£ β€ π. Let π = cos(2π½) and
π =sin(2π½). Then
hΞ¨π»(π½,1) |ππ’ππ£|Ξ¨π»(π½,1)i = (π 2/2) Γ
πβ π’,π£
cos[2π½π’, πβ2π½π£ , π] β (π 2/2) Γ
πβ π’,π£
cos[2π½π’, π+2π½π£ , π] +π π Β·sin(2π½π’,π£)
"
Γ
πβ π’,π£
cos(2π½π’, π) + Γ
πβ π’,π£
cos(2π½π£ , π)
# .
Here we only consider the caseπΎ =1 sinceπΎcan be absorbed into the definition of π½.
Proof. Given a 2-qubit observableπ, define the mean value π(π) =hΞ¨π»(π½,1) |ππ’,π£|Ξ¨π»(π½,1)i. We are interested in the observableπ =π π β‘ π β π.
We note that all terms in π» and π΅ that act trivially on {π’, π£} do not contribute to π(π). Such terms can be set to zero. Given a 2-qubit observableπ, define a mean value
π0(π)= h+π|ππ π»
0
ππ’,π£πβπ π»
0|+πi, where π»0= βοΈ
πβ π’,π£
(π½π’, πππ’+π½π£ , πππ£)ππ. Using the identities
ππ π½(ππ’+ππ£)ππ’ππ£πβπ π½(ππ’+ππ£) = π2ππ’ππ£+π 2ππ’ππ£+π π (ππ’ππ£+ππ’ππ£), ππ π½π’ , π£ππ’ππ£ππ’ππ£πβπ π½π’ , π£ππ’ππ£ = ππ’ππ£,
ππ π½π’ , π£ππ’ππ£ππ’ππ£πβπ π½π’ , π£ππ’ππ£ = ππ’ππ£
ππ π½π’ , π£ππ’ππ£ππ’ππ£πβπ π½π’ , π£ππ’ππ£ = cos(2π½π’,π£)ππ’ππ£+sin(2π½π’,π£)ππ£, ππ π½π’ , π£ππ’ππ£ππ’ππ£πβπ π½π’ , π£ππ’ππ£ = cos(2π½π’,π£)ππ’ππ£+sin(2π½π’,π£)ππ’, and noting thatπ0(π π)=0, one easily gets
π(π π) =π 2Β·π0(π π) +π π Β·cos(2π½π’,π£) [π0(ππ) +π0(π π)] +π π Β·sin(2π½π’,π£) [π0(π πΌ) +π0(πΌ π)]. Using the explicit form ofπ»0, one gets
πβπ π»
0|+πi= 1 2
βοΈ
π,π=0,1
|π, πiπ’,π£ β |Ξ¦(π, π)ielse, where|Ξ¦(π, π)i is a tensor product state ofπβ2 qubits defined by
|Ξ¦(π, π)i = Γ
πβ π’,π£
|π½π’, π(β1)π+π½π£ , π(β1)πiπ where |πi β‘πβπ π π|+i.
Combining Eqs. (3.25) and (3.28), one gets π0(π) = (1/4) βοΈ
π,π,π0,π0=0,1
hπ0, π0|π|π, πi Β· hΞ¦(π0, π0) |Ξ¦(π, π)i.
Using the tensor product form of the states |Ξ¦(π, π)i and the identity hπ0|πi = cos(πβπ0)gives
hΞ¦(π0, π0) |Ξ¦(π, π)i = Γ
πβ π’,π£
cos[π½π’, π(β1)πβπ½π’, π(β1)π0+π½π£ , π(β1)πβπ½π£ , π(β1)π0]. From Eqs. (3.30) and (3.31), one can easily compute the mean valueπ0(π) for any 2-qubit observable.
Consider first the caseπ =π π. Then the only terms contributing to Eq. (3.30) are those withπ0=πβ1 andπ0=πβ1. The identityhπβ1|π|πi=βπ(β1)πgives
π0(π π) =β(1/4) βοΈ
π,π=0,1
(β1)π+π Γ
πβ π’,π£
cos[2π½π’, π(β1)π+2π½π£ , π(β1)π], that is,
π0(π π)= (1/2) Γ
πβ π’,π£
cos[2π½π’, πβ2π½π£ , π] β (1/2) Γ
πβ π’,π£
cos[2π½π’, π+2π½π£ , π].
Next, consider the caseπ =π π. Note that the matrix elementshπ0, π0|π|π, πihave zero real part. From Eqs. (3.30) and (3.31), one infers that π0(π π) has zero real part. This implies
π0(π π)= π0(ππ) =0.
Finally, consider the caseπ =π πΌ. Then the only terms that contribute to Eq. (3.30) are those withπ0=π β1 and π0=π. We get
π0(π πΌ) = Γ
πβ π’,π£
cos(2π½π’, π).
Here we noted that the inner product Eq. (3.31) withπ0=πβ1 andπ0=πdoes not depend onπ, π. By the same argument,
π0(πΌ π) = Γ
πβ π’,π£
cos(2π½π£ , π).
Combining Eq. (3.27) and Eqs. (3.33),(3.34),(3.35),(3.36), one arrives at Eq. (3.23).
For more general cost functions that include interactions among three or more variables, there are two complications: First, unlike in the Ising case, the variable elimination process will typically increase the degree of non-locality of interactions.
Second, mean values of Pauli operators on the QAOA stateΞ¨π»(π½, πΎ) lack a simple analytic formula (as far as we know). However, one can approximately compute the mean values using the Monte Carlo method due to Van den Nest [30]. A specialization of this method to simulation of the level-1 QAOA is described in [31].
The Monte Carlo simulator has runtime scaling polynomially with the number of qubits, number of terms in the cost function, and the inverse error tolerance, see [31]
for details. This method also requires no restrictions on the depth of the variational circuit.
An important distinction between QAOA and RQAOA lies in the measurement step.
QAOA requires few-qubit measurements to estimate the variational energy as well as the finalπ-qubit measurement that assigns a value to each individual variable. This last step is what makes QAOA hard to simulate classically and may lead to a quantum advantage [32]. In contrast, RQAOA only needs few-qubit measurements to estimate mean values of individual terms in the cost function. Theπ-qubit measurement step is replaced by the correlation rounding that eliminates variables one by one. One may ask whether the lack of multi-qubit measurements also precludes a quantum advantage. Indeed, in the special case of level-1 variational circuits and the Ising-like cost function RQAOA can be efficiently simulated classically, see above. However, level-π RQAOA with π > 1 as well as level-1 RQAOA with more general cost functions are not known to be classically simulable in polynomial time, leaving room for a quantum advantage.
3.D Comparison of QAOA, RQAOA, and Classical Algorithms D.1 QAOA versus Classical Local Algorithms
In this section, we discuss another limitation of QAOA which results from its locality and the covariance condition discussed in Lemma 3.A.2: we compare QAOA to a certain very simple classical local algorithm (see Lemma 3.D.1 below). We show that there is an exponential number of problem instances for which the classical local algorithm outperforms QAOA.
Let us briefly sketch the notion of a local classical algorithm. We envision that the tuple(π½π)πβπΈ is given as input. Here we are interested in algorithms which are local
with respect to the underlying graphπΊ. Forπ βNandπ£ βπ, define πΈπ(π£)=
π
Γ
β=1
Γ
(π
1,...,πβ) path withπ£βπ
1
{π
1, . . . , πβ}
to be the set of edges that belong to a path starting at π£ of length bounded by π. Consider a classical algorithmAwhich on input{π½π}πβπΈoutputsπ₯ = (π₯
1, . . . , π₯π) β {0,1}π. We say that A isπ-local if there is a family of functions {ππ£ : RπΈπ(π£) β {0,1}}π£βπ such that the following holds for every problem instance (π½π)πβπΈ β RπΈ: We have
π₯π£ =ππ£ {π½π}πβπΈπ(π£)
for everyπ£ βπ .
In other words, in anπ-local classical algorithm, every output bitπ₯π£only depends on edge weightsπ½πbelonging to paths of length bounded byπstarting atπ£. We note that this definition can easily be generalized to the probabilistic case (e.g., by including local random bits). For the purposes of this section, deterministic functions turn out to be sufficient.
The (choice of) family{ππ£}π£βπ can be considered as a set of variational parameters for the classical algorithm. To keep the number of variational parameters constant, we consider vertex-transitive graphs πΊ. Fix π£β β π. For every π£ β π, fix an automorphism ππ£ of πΊ such that ππ£(π£β) = π£. Then the sets πΈπ(π£) for different π£ β π can be identified via πΈπ(π£) = ππ£(πΈπ(π£β)). We say that anπ-local classical algorithm is uniformif (after this identification)ππ£ β‘ π for all π£ βπ, i.e., if there is a single functionπ : RπΈπ(π£β) β {0,1} specifying the behavior of the algorithm.
To obtain general-purpose algorithms (applicable to any instance), the function π : RπΈπ(π£β) β {0,1} should be chosen adapatively (i.e., potentially depending on the instance). The definition of local classical algorithm sketched here includes e.g., the algorithms considered in Ref. [20], though it is slightly more general as the local functions can be arbitrary.
Letπ=6π be a multiple of 6. Considerπ-qubit Hamiltonians (cf. (3.A)) of the form π»(π½) = βοΈ
πβZπ
π½πππππ+
1 where π½ = (π½
0, . . . , π½πβ
1) β {1,β1}π.
To define locality and uniformity for the cycle graphZπ, letππ£(π€) =π£+π€ (mod π) be chosen as translation moduloπforπ£ βZπ. We show the following:
Lemma 3.D.1. There is a subsetS β {1,β1}πof2π/3problem instances such that the following holds:
(i) QAOAπ(π»(π½)) β€ π/(π+1) for everyπ βNand everyπ½ β S.
(ii) There is a1-local uniform classical algorithm such that for every π½ β S, the algorithm outputsπ₯ β {0,1}πsuch thathπ₯|π»(π½) |π₯i=1.
(iii) Level-1RQAOAachieves the approximation ratio1. Proof. For everyπ = (π
0, . . . , π
2πβ1) β {0,1}2π, defineπ½ =π½(π ) β {1,β1}πby π½3π =π½
3π+1 =(β1)π π, and π½
3π+2 =1,
for allπ =0,1, . . . ,2πβ1. We claim that the setS ={π½(π ) | π β {0,1}2π}has the required properties. Consider an instanceπ»(π½(π ))withπ β S. Define
π(π ) =
2πβ1
Γ
π=0
π3π+1 . Thenπ»(π½(π )) is related toπ»
Zπ =Γ
πβZπππππ+
1by the gauge transformation π»(π½(π ))= π(π )π»
Zππ(π )β1.
Since the QAOA algorithm is invariant under such gauge transformation (see Lemma 3.A.2), we obtain
QAOAπ(π»(π½(π )))=QAOAπ(π»
Zπ) β€ π π+1 where we use the bound
QAOAπ(π»MaxCut
Zπ ) β€ 2π+1 2π+2 ,
proven in [22] for evenπ, in combination with Lemma 3.A.2. This shows (i).
For the proof of (ii), consider the classical algorithm A which on input π½ = (π½
0, . . . , π½πβ
1)outputs
π₯π£ =π(π½π£β
1, π½π£) for everyπ£ βZπ, where
π(π½ , π½0) =


ο£²

ο£³
1 if(π½ , π½0)= (β1,β1) 0 otherwise.
Clearly, the algorithm A is uniform and 1-local, and it is easy to check that the output satisfieshπ₯|π»(π½) |π₯i=1.
The proof of (iii) is given as a part of Lemma 3.D.2.