OBSTACLES TO STATE PREPARATION AND VARIATIONAL OPTIMIZATION FROM SYMMETRY PROTECTION
3.3 The Recursive Quantum Approximate Optimization Algorithm
dist(๐ , ๐) > 2๐ , one infers that๐โ ๐๐๐ and๐โ ๐๐๐ have disjoint support. Thus h+๐|๐โ ๐๐๐๐๐|+๐i =h+๐| (๐โ ๐๐๐) (๐โ ๐๐๐) |+๐i
= h+๐|๐โ ๐๐๐|+๐i ยท h+๐|๐โ ๐๐๐|+๐i=0.
This proves Eq. (3.9). Suppose one prepares the state๐|+๐iand measures a pair of qubits ๐ < ๐ in the standard basis. Then ๐๐ , ๐ is the probability that the measured values on qubits ๐ and๐ disagree. By the union bound,
๐๐ , ๐ โค
๐โ1
โ๏ธ
๐=๐
๐๐,๐+
1.
Indeed, if qubits ๐ and ๐ disagree, at least one pair of consecutive qubits located in the interval [๐ , ๐] must disagree. Set๐ = ๐ +2๐ +1. Then๐๐ , ๐ =1/2 by Eq. (3.9).
Take the expected value of Eq. (3.10) with respect to random uniform ๐ โZ๐. This gives
1 2
โค 2๐ +1 ๐
โ๏ธ
๐โZ๐
๐๐,๐+
1= 2๐ +1 ๐
h+๐|๐โ ๐ป๐๐|+๐i
proving Eq. (3.8). In Appendix 3.B, we construct aZ2-symmetric range-๐ circuit๐ such that๐|+๐iis a tensor product of GHZ-like states on consecutive segments of 2๐ +1 qubits. We show that such circuit saturates the upper bound Eq. (3.6). This completes the proof of Theorem 3.2.3.
First, run the standard QAOA to maximize the expected value of ๐ป๐ on the state
|๐i = ๐(๐ฝ, ๐พ) |+๐i. For every edge (๐ , ๐) โ ๐ธ, compute ๐๐ , ๐ = h๐โ|๐๐๐๐|๐โi, where๐โ is the optimal variational state.
Next, find a pair of qubits (๐, ๐) โ ๐ธ with the largest magnitude of ๐๐, ๐ (breaking ties arbitrarily). The corresponding variables ๐๐ and ๐๐ are correlated if ๐๐, ๐ > 0 and anti-correlated if๐๐, ๐ < 0. Impose the constraint
๐๐ =sgn(๐๐, ๐)๐๐
and substitute it into the Hamiltonian๐ป๐to eliminate the variable๐๐. For example, a term๐๐๐๐ with๐ โ{๐, ๐}gets mapped to sgn(๐๐, ๐)๐๐๐๐. The term ๐ฝ๐, ๐๐๐๐๐ gets mapped to a constant energy shift๐ฝ๐, ๐sgn(๐๐, ๐). All other terms remain unchanged.
This yields a new Ising Hamiltonian ๐ป๐โ
1 that depends on ๐ โ1 variables. By construction, the maximum energy of๐ป๐โ
1coincides with the maximum energy of ๐ป๐over the subset of assignments satisfying the constraint Eq. (3.11).
Finally, call RQAOA recursively to maximize the expected value of ๐ป๐โ
1. Each recursion step eliminates one variable from the cost function. The recursion stops when the number of variables reaches some specified threshold value๐๐ ๐. The remaining instance of the problem with ๐๐ variables is then solved by a purely classical algorithm (for example, by a brute force method). Thus the value of ๐๐ controls how the workload is distributed between quantum and classical computers.
We describe a generalization of RQAOA applicable to Ising-like cost functions with multi-spin interactions in Appendix 3.C.
Imposing a constraint of the form (3.11) can be viewed as rounding correlations among the variables ๐๐ and ๐๐. Indeed, the constraint demands that these vari- ables must be perfectly correlated or anti-correlated. This is analogous to rounding fractional solutions obtained by solving linear programming relaxations of combi- natorial optimization problems. We note that reducing the size of a problem to the point that it can be solved optimally by brute force is a widely used and effective approach in combinatorial optimization.
We compare the performance of the standard QAOA, RQAOA, and local classical algorithms by considering the Ising Hamiltonians in Eq. (3.3) with couplings๐ฝ๐,๐ =
ยฑ1 defined on the cycle graph. In Appendix 3.D, we prove:
Theorem 3.3.1. For each integer๐divisible by6, there exists a family of2๐/3Ising Hamiltonians of the form ๐ป๐ = ร
๐โZ๐๐ฝ๐๐๐๐๐+
1 with ๐ฝ๐ โ {1,โ1} such that the
(a) (b)
Figure 3.1: Comparisons of level-1 RQAOA and the Goemans-Williamson Al- gorithm. (a) Approximation ratios achieved by level-1 RQAOA (blue) and the Goemans-Williamson (GW) algorithm [21] (red) for 15 instances of the Ising cost function with random ยฑ1 couplings defined on the 2D toric grid of size 16ร 16.
In case (b) the Ising Hamiltonian also includes random ยฑ1 external fields. The RQAOA threshold value is ๐๐ = 20. We found that the standard level-1 QAOA achieves approximation ratios below 1/2 for all considered instances (not shown).
The GW algorithm was implemented with๐ =256 rounding attempts and the best found solution was selected. The exact maximum energy was computed using integer linear programming.
following holds for all Hamiltonians in the family:
(i) There is a local classical algorithm which achieves the approximation ratio1. (ii) Level-๐QAOA achieves an approximation ratio of at most ๐/(๐+1).
(iii) Level-1RQAOA achieves the approximation ratio1.
Our definition of local classical algorithms follows [20]. We also show that the level-1 RQAOA achieves the optimal approximation ratio for any 1D Ising model with coupling coefficients๐ฝ๐ โ {1,โ1}. This proves that, in certain cases, RQAOA is strictly more powerful than QAOA.
Finally, we report a numerical comparison between the level-1 RQAOA and the Goemans-Williamson algorithm [21] for the Ising cost function Eq. (3.3) with ran- dom coefficients๐ฝ๐ , ๐ =ยฑ1. Two graphs are considered: (a) the 2D grid, and (b) the 2D grid with one extra vertex connected to all grid points. The latter is equivalent to the 2D Ising model with random ยฑ1 external fields. As shown in [23], the problem of maximizing the energy ๐ถ(๐ฅ) admits an efficient algorithm in case (a) while case (b) is NP-hard. To compute the mean values h๐(๐ฝ, ๐พ) |๐๐๐๐|๐(๐ฝ, ๐พ)i,
we used a version of the algorithm by Wang et al [24], as detailed in Appendix 3.C.
Figure 3.1 shows approximation ratios achieved by each algorithm for 15 problem instances with the grid size 16ร16. It can be seen that RQAOA outperforms the Goemans-Williamson algorithm for all except for one instance. We found that the standard level-1 QAOA achieves approximation ratio below 1/2 for all considered instances.
Note added: After submission of this work, analogous limitations were estab- lished for random regular graphs by exploiting the locality and spatial uniformity of QAOA [25], [26]. We focus onZ2-symmetry and locality, and our statements also apply to non-uniform local algorithms.
Acknowledgements
The authors thank Giacomo Nannicini and Kristan Temme for helpful discussions.
SB was partially supported by the IBM Research Frontiers Institute and by the Army Research Office (ARO) under Grant Number W911NF-20-1-0014. ET ac- knowledges the support of the Natural Sciences and Engineering Research Council of Canada (NSERC) and funding provided by the Institute for Quantum Information and Matter, an NSF Physics Frontiers Center (NSF Grant No. PHYโ1733907).
RK and AK gratefully acknowledge support by the DFG cluster of excellence 2111 (Munich Center for Quantum Science and Technology) and by IBM.
3.A Proof of Corollary 3.2.2
In this appendix, we give a proof of Corollary 3.2.2 in the main text. Here and below, we will denote the expected approximation ratio achieved by the QAOA with Hamiltonian๐ปas
QAOA๐(๐ป) =
max
๐ฝ,๐พโR๐
hฮจ๐ป(๐ฝ, ๐พ) |๐ป|ฮจ๐ป(๐ฝ, ๐พ)i
ยท
max
๐ฅโ{0,1}๐
h๐ฅ|๐ป|๐ฅi โ1
, where
|ฮจ๐ป(๐ฝ, ๐พ)i =๐๐ป(๐ฝ, ๐พ) |+๐i and ๐๐ป(๐ฝ, ๐พ) =
๐
ร
๐=1
๐๐ ๐ฝ๐๐ต๐๐ ๐พ๐๐ป
(3.16) for ๐ฝ, ๐พ โ R๐ and where๐ต =ร๐
๐=1๐๐. Let us first record a few general features of the QAOA for later use.
Let ๐บ = (๐ , ๐ธ) be a graph, ๐ = |๐|, ๐ = |๐ธ|, and let ๐ฝ = (๐ฝ๐)๐โ๐ธ โ R๐ธ be an assignment of edge weights on๐บ. Let us define the Hamiltonian๐ป๐บ(๐ฝ) as
๐ป๐บ(๐ฝ) = โ๏ธ
{๐ข,๐ฃ}โ๐ธ
๐ฝ{๐ข,๐ฃ}๐๐ข๐๐ฃ . (3.17) It will be useful for later to also define
๐ป๐บ = โ๏ธ
{๐ข,๐ฃ}โ๐ธ
๐๐ข๐๐ฃ, and ๐ปMaxCut
๐บ = 1
2
(๐ ๐ผ โ๐ป๐บ),
where ๐ปMaxCut
๐บ is the Hamiltonian used in QAOA for the Maximum Cut problem on the graph๐บ. We will use the following bound on the circuit depth of a QAOA unitary.
Lemma 3.A.1. Let ๐ = ๐๐ป(๐ฝ, ๐พ) with ๐ฝ, ๐พ โ R๐ be a level-๐ QAOA unitary (cf. Eq.(3.A)) for a Hamiltonian๐ป =๐ป๐บ(๐ฝ)on a graph๐บ(cf.(3.A)). Let๐ทbe the maximum degree of๐บ. Then๐ can be realized by a circuit of depth๐ โค ๐(๐ท+1) consisting of2-qubit gates.
If ๐บ is ๐ท-regular and bipartite, then the circuit depth of ๐ can be bounded by ๐ โค ๐ ๐ท.
Proof. By Vizingโs theorem [27], there is an edge coloring of๐บ with at most๐ท+1 colors. Taking such a coloring ๐ธ = ๐ธ
1โช ยท ยท ยท โช ๐ธ๐ท+
1, we may apply each level ๐๐ ๐ฝ ๐ต๐๐ ๐พ ๐ป of๐ in depth ๐ท +1 by applying (ร
๐ฃโ๐๐๐ ๐ฝ ๐๐ฃ)ร๐ท+1
๐=1 ๐๐(๐พ), where each ๐๐(๐พ) =
ร
{๐ข,๐ฃ}โ๐ธ๐
๐๐ ๐พ ๐ฝ{๐ข , ๐ฃ}๐๐ข๐๐ฃ
is a depth-1-circuit of two-local gates.
If๐บ is๐ท-regular and bipartite, we may reduce the chromatic number upper bound from ๐ท +1 to ๐ท since all bipartite graphs are ๐ท-edge-colorable by Kลnigโs line coloring theorem. We illustrate the construction of the circuit on Figure 3.2 for the case๐ท =3 and ๐=1.
The expected QAOA approximation ratios of suitably related instances are identical:
Lemma 3.A.2. LetL โ ๐ be an arbitrary subset of vertices and ๐L be the set of edges that have exactly one endpoint inL. Let๐ฝ =(๐ฝ๐)๐โ๐ธ โR๐ธ be arbitrary edge weights. Define๐ฝห= (๐ฝห๐)๐โ๐ธ โR๐ธ by
ห ๐ฝ๐ =
๏ฃฑ๏ฃด
๏ฃด
๏ฃฒ
๏ฃด๏ฃด
๏ฃณ
โ๐ฝ๐ if๐ โ๐L ๐ฝ๐ otherwise.
Figure 3.2: Example for the construction of the circuit given in Lemma 3.A.1: a 4-colorable graph with maximum degree 3 alongside its associated depth-5 quantum circuit for the level-1 QAOA unitary.
Then expected QAOA approximation ratios satisfy
QAOA๐(๐ป๐บ(๐ฝ)) =QAOA๐(๐ป๐บ(๐ฝห)).
Proof. Let us write ๐ป = ๐ป๐บ(๐ฝ) and ห๐ป = ๐ป๐บ(๐ฝห) for brevity. Let ๐ = ๐[L]
be a tensor product of Pauli-๐ operators acting on every qubit in L โ ๐. Then
ห
๐ป = ๐ ๐ป ๐, which implies that
๐ฅโ{max0,1}๐h๐ฅ|๐ป|๐ฅi= max
๐ฅโ{0,1}๐h๐ฅ|๐ปห|๐ฅi. (3.18) Let๐ฝ, ๐พ โR๐be arbitrary. Then we also have
๐|ฮจ๐ปห(๐ฝ, ๐พ)i =
๐
ร
๐=1
(๐ ๐๐ ๐ฝ๐๐ต๐๐ ๐พ๐๐ปห๐) |+๐i=
๐
ร
๐=1
(๐๐ ๐ฝ๐๐ต๐๐ ๐พ๐๐ป) |+๐i=|ฮจ๐ป(๐ฝ, ๐พ)i,
where identities in the middle follow since|+๐iis stabilized by๐, and since[๐ , ๐ต] = 0. Therefore we have
hฮจ๐ปห(๐ฝ, ๐พ) |๐ปห|ฮจ๐ปห(๐ฝ, ๐พ)i =hฮจ๐ปห(๐ฝ, ๐พ) |๐ ๐ป ๐|ฮจ๐ปห(๐ฝ, ๐พ)i =hฮจ๐ป(๐ฝ, ๐พ) |๐ป|ฮจ๐ป(๐ฝ, ๐พ)i.
Combined with (3.A), this implies the claim.
In particular, if๐บ = (๐ , ๐ธ)is a bipartite graph, then Lemma 3.A.2 implies that QAOA๐(๐ป๐บ) =QAOA๐(โ๐ป๐บ)
and
QAOA๐(๐ปMaxCut
๐บ ) = 1
2
(1+QAOA๐(๐ป๐บ)). (3.19) We now prove Corollary 1. It is a direct consequence of Theorem 1, which we restate here for convenience in the notation of this appendix:
Theorem 3.A.1. Consider a family {๐บ๐ = ( [๐], ๐ธ๐)}๐โI of graphs with Cheeger constant lower bounded asโ(๐บ๐) โฅ โ > 0for all๐โ I. Then
h๐|๐โ ๐ป๐บ
๐๐|๐i< |๐ธ๐| โ โ๐ 3
for any Z2-symmetric depth-๐ circuit ๐ composed of two-qubit gates, any Z2- symmetric product state๐, and any๐ >4828๐,๐ โ I.
Then we have the following:
Corollary 3.A.1. For every integer๐ท โฅ 3, there exists an infinite family of bipartite ๐ท-regular graphs{๐บ๐}๐โI such that
QAOA๐(๐ปMaxCut
๐บ๐
) โค 5 6
+
โ ๐ทโ1
3๐ท as long as
๐ < (1/3 log2๐โ4)๐ทโ1 .
Proof. Fix some๐ท โฅ 3. By the results of [28], [29], there exists an infinite family {๐บ๐}๐โI of bipartite ๐ท-regular Ramanujan graph with๐ vertices for every๐ โ I. Consider a fixed ๐ โ I and let ๐ = ๐(๐) be the associated QAOA level. Let ๐๐ = ๐๐ป
๐บ๐(๐ฝโ, ๐พโ) be a level-๐ QAOA unitary for the Hamiltonian ๐ป๐บ
๐ on ๐บ๐, and assume that ๐ฝโ, ๐พโ โ R๐ are such that the expectation of ๐ป๐บ
๐ is maximized.
Because๐บ๐is๐ท-regular, the circuit depth of๐๐can be bounded from above by ๐ ๐ท according to Lemma 3.A.1. Condition (3.12) implies that๐ >4828๐ ๐ท, thus
QAOA๐(๐ป๐บ
๐) = 1
|๐ธ๐|h+๐|๐โ
๐๐ป๐บ
๐๐๐|+๐i < 1โ โ
3|๐ธ๐|๐ =1โ 2โ 3๐ท
by Theorem 1, where we have used that |๐ธ๐| =๐ ๐ท/2. With (3.A) (using that๐บ๐is bipartite), we conclude that
QAOA๐(๐ปMaxCut
๐บ๐
) < 1โ โ 3๐ท
. The claim then follows from the boundโ/๐ท โฅ (๐ทโ2
โ
๐ทโ1)/(2๐ท), valid for all Ramanujan graphs.
3.B Optimal Variational Circuit for the Ring of Disagrees
In this section we prove that the upper bound of Theorem 2 in the main text is tight whenever๐is a multiple of 2๐ +1. Let
|GHZni=2โ1/2(|0๐i + |1๐i) be the GHZ state of๐qubits.
Lemma 3.B.1. Suppose๐=2๐+1for some integer๐. There exists aZ2-symmetric range-๐quantum circuit๐ such that
|GHZni=๐|+๐i.
Proof. We shall writeCX๐,๐ก for the CNOT gate with a control qubit๐ and a target qubit๐ก. Let๐= ๐+1 be the central qubit. One can easily check that
|GHZni=ยฉ
ยญ
ยซ
๐
ร
๐=1
CX๐,๐โ๐CX๐,๐+๐
ยช
ยฎ
ยฌ
๐ป๐|0๐i.
AllCXgates in the product pairwise commute, so the order does not matter. Inserting a pair of Hadamards on every qubit ๐ โ [๐] \ {๐}before and after the respectiveCX gate and using the identity(๐ผ โ๐ป)CX(๐ผโ ๐ป) =CZ, one gets
|GHZni=ยฉ
ยญ
ยซ ร
๐โ[๐]\{๐}
๐ป๐ยช
ยฎ
ยฌ
ยฉ
ยญ
ยซ
๐
ร
๐=1
CZ๐,๐โ๐CZ๐,๐+๐ยช
ยฎ
ยฌ
|+๐i.
Let๐=exp[๐(๐/4)๐] be the phase-shift gate. Define the two-qubit Clifford gate RZ=(๐โ ๐)โ1CZ=exp(โ๐ ๐/4)exp[โ๐(๐/4) (๐ โ๐)].
ExpressingCZin terms ofRZand๐ in Eq. (3.14), one gets
|GHZni =๐2
๐ ๐
ยฉ
ยญ
ยซ ร
๐โ[๐]\{๐}
๐ป๐๐๐ยช
ยฎ
ยฌ
ยฉ
ยญ
ยซ
๐
ร
๐=1
RZ๐,๐โ๐RZ๐,๐+๐
ยช
ยฎ
ยฌ
|+๐i.
Multiply both sides of Eq. (3.15) on the left by a product of ๐ gates over qubits ๐ โ [๐] \ {๐}. Noting that
๐ ๐ป ๐ =๐exp[โ๐(๐/4)๐], one gets (ignoring an overall phase factor)
ร
๐โ[๐]\{๐}
๐๐|GHZni=๐2
๐ ๐
ยฉ
ยญ
ยซ ร
๐โ[๐]\{๐}
exp[โ๐(๐/4)๐๐]ยช
ยฎ
ยฌ
ยฉ
ยญ
ยซ
๐
ร
๐=1
RZ๐,๐โ๐RZ๐,๐+๐ยช
ยฎ
ยฌ
|+๐i.
Using the identity
ร
๐โ[๐]\{๐}
๐๐|GHZni =๐2
๐
๐ |GHZni, one can cancel๐2
๐
๐ that appears in both sides of Eq. (3.16). We arrive at Eq. (3.13) with
๐ =ยฉ
ยญ
ยซ ร
๐โ[๐]\{๐}
exp[โ๐(๐/4)๐๐]ยช
ยฎ
ยฌ
ยฉ
ยญ
ยซ
๐
ร
๐=1
RZ๐,๐โ๐RZ๐,๐+๐ยช
ยฎ
ยฌ .
The circuit diagram of๐ in the case ๐ = 7 is shown in Figure 3.3. Obviously,๐ isZ2-symmetric since any individual gate commutes with ๐โ๐. Let us check that ๐ has range-๐. Consider any single-qubit observable๐๐ acting on the ๐-th qubit.
Consider three cases:
Case 1: ๐ = ๐. Then ๐โ ๐๐๐ may be supported on all ๐ qubits. However, [๐โ ๐, ๐+๐] = [1, ๐], so the ๐-range condition is satisfied trivially.
Case 2: 1 โค ๐ < ๐. Then all gates RZ๐,๐+๐ in๐ cancel the corresponding gates in๐โ , so that๐โ ๐๐๐ has support in the interval[1, ๐] โ [๐โ ๐, ๐+ ๐]. Thus the ๐-range condition is satisfied.
Case 3: ๐ < ๐ โค ๐. This case is equivalent to Case 2 by symmetry.
Recall that we consider the ring of disagrees Hamiltonian ๐ป๐ = 1
2
โ๏ธ
๐โZ๐
(๐ผโ๐๐๐๐+
1).
Lemma 3.B.2. Consider any integers๐, ๐such that๐is even and๐is a multiple of 2๐+1. Then there exists aZ2-symmetric range-๐circuit๐such that
h+๐|๐โ ๐ป๐๐|+๐i= 2๐+1/2 2๐+1
.
Proof. Let๐be theZ2-symmetric range-๐unitary operator preparing the GHZ state on 2๐+1 qubits starting from |+2๐+1i, see Lemma 3.B.1. Suppose๐ =๐(2๐+1) for some integer๐. Define
๐ =๐
1๐
2,
Figure 3.3: The Z2-symmetric range-3 quantum circuit to prepare the GHZ state
|GHZ2p+1iof 2๐+1=7 qubits (๐=3). Here,๐ ๐(๐) =exp(โ๐ ๐ ๐). where
๐1= (๐ โ ๐ผ)โ๐/2 and ๐
2 =๐โ๐.
Since each copy of๐ acts on a consecutive interval of qubits and has range ๐, one infers that๐ has range๐. We have
๐โ
1๐ป๐๐
1 = โ๏ธ
๐โZ๐
๐บ๐, ๐บ๐ = 1 2
(๐ผ+๐๐๐๐+
1). The state๐
2|+๐iis a tensor product of GHZ states supported on consecutive tuples of 2๐+1 qubits. The expected value of ๐บ๐ on the state๐
2|+๐iequals 1 if ๐บ๐ is supported on one of the GHZ states. Otherwise, if๐บ๐crosses the boundary between two GHZ states, the expected value of๐บ๐on the state๐
2|+๐iequals 1/2. Thus h+๐|๐โ ๐ป๐๐|+๐i = โ๏ธ
๐โZ๐
h+๐|๐โ
2
๐บ๐๐
2|+๐i =๐(2๐+1/2) =๐2๐+1/2 2๐+1
.
3.C Recursive QAOA
In this appendix, we outline the Recursive QAOA algorithm (RQAOA) for general cost functions.