Annealed Lower Bound Optimization

• Annealed Lower Bound Optimization is a technique that defines minimal ferromagnetic chain couplings in quantum annealing to ensure accurate logical encoding by preventing domain walls.
• It uses formal inequalities and a closed-form bound to optimize embedding energy, outperforming older methods like Choi’s bounds in precision and efficiency.
• Empirical results on D-Wave hardware show that this optimization increases ground-state success probability and reduces annealing time for combinatorial problems.

Annealed lower bound optimization refers to the process of determining and attaining the minimal feasible values of certain parameters—in particular, chain (ferromagnetic) couplings in quantum annealing minor embedding—subject to constraints ensuring proper operation of the annealer, such as the absence of domain walls in embedded chains. This approach directly impacts the physical realization of quantum annealing for problems mapped onto hardware graphs of limited connectivity, and is central to maximizing ground-state success probability by minimizing total embedding energy. Analytical lower bounds that tighten these constraints are both theoretically significant and have practical consequences for hardware implementations and algorithmic efficiency.

1. Minor Embedding in Quantum Annealing

Quantum annealing hardware often employs a physical graph topology with limited connectivity, typically much sparser than the logical connectivity graphs of combinatorial optimization problems. Minor embedding is the canonical procedure for mapping each logical qubit $i$ of the problem Hamiltonian onto a connected subtree—or most commonly, a chain—of physical qubits $\iota(i)$ on the hardware.

Each edge $(i_p, i_q)$ in $\iota(i)$ is assigned a uniform ferromagnetic coupling $F_i^{pq} = F_i < 0$, aligning all the qubits in the tree to encode a logical state. The collective strength of these couplings across all logical qubits defines the minor embedding energy: $$E_{\rm embed} = \sum_{i \in V(G)} |\iota(i)|_E\,|F_i|$$ where $|\iota(i)|_E$ indicates the number of coupler edges in $\iota(i)$. For chain embeddings, $E_{\rm embed} \approx \sum_i|F_i|$. Sufficiently large $|F_i|$ are necessary to prevent control noise or thermal fluctuations from introducing “domain walls”—boundaries in a chain where spins misalign, corrupting the logical encoding and ground-state readout (Fang et al., 2019).

2. No–Domain–Wall Constraint: Formal Inequalities

A domain wall occurs when a nonempty proper subset $W_i \subsetneq \iota(i)$ of a chain flips relative to its complement, resulting in ambiguous logical state assignment. Formal analysis yields that to suppress such configurations, the following must be enforced for all $W_i$: $$|F_i| > \frac{1}{|\partial W_i|}\,\min\left\{ |h(W_i) - J(W_i)|\,,\; |h(W_i)-h_i-J(\iota(i)\setminus W_i)| \right\}$$ with $h(W_i)$ the sum of local fields over $W_i$, $J(W_i)$ the total magnitude of external couplings from $W_i$, and $|\partial W_i|$ the number of ferromagnetic chain edges connecting $W_i$ to its complement. The global constraint is

$$|F_i| \ge \max_{\;\emptyset\neq W_i\subsetneq\iota(i)}\; \frac{1}{|\partial W_i|} \; \min\left\{ |h(W_i) - J(W_i)|,\; |h(W_i)-h_i-J(\iota(i)\setminus W_i)| \right\}$$

This constraint applies simultaneously to every embedded logical qubit and all possible nontrivial chain partitions (Fang et al., 2019).

3. Analytical Construction of the Tightest Lower Bounds

The derivation begins by comparing the energy of uniform chain alignment to the energy of states containing putative domain walls. For any subset $W_i$, the energy difference upon flipping $W_i$ is

$$\Delta(W_i) = 2\left[ h(W_i) - J(W_i) - |\partial W_i| F_i \right]$$

Enforcing $\Delta(W_i) > 0$ for both $W_i$ and its complement yields the two inequalities, from which the tightest lower bound is constructed by taking the minimum on the right-hand sides and the maximum over all subsets. The main theorem (Theorem 2.2 in (Fang et al., 2019)) gives the closed-form optimal bound: $$|F_i| \ge \max_{\emptyset\neq W\subsetneq\iota(i)}\; \frac{1}{|\partial W|} \; \min\left\{ |h(W)-J(W)|,\; |h(W)-h_i-J(\iota(i)\setminus W)| \right\}$$ This result is provably tighter than previous sufficient conditions such as Choi’s 2008 bounds, which rely on cruder estimates involving total field and coupling magnitudes (Fang et al., 2019).

4. Optimization Procedure for Minimal Embedding Energy

Given node-wise lower bounds $M_i$ for each logical qubit $i$, the global embedding energy optimization reduces to

$$\min \sum_{i} |F_i| \quad \text{subject to} \quad |F_i| \ge M_i$$

with trivial solution $|F_i| = M_i$. However, $M_i$ itself depends on the distribution of logical field $h_i$ among the chain’s physical qubits. The tightest value is computed via the auxiliary optimization: $$M_i^{\rm opt} = \min_{\sum_k h_{i(k)}=h_i} \; \max_{\emptyset\neq W\subsetneq\iota(i)} \frac{1}{|\partial W|}\min\left\{|h(W)-J(W)|,\;|h(W)-h_i-J(\iota(i)\setminus W)|\right\}$$ Enumerative or dynamic programming approaches are feasible for practical chain lengths. Uniform splitting $h_{i(k)}=h_i/|\iota(i)|$ often yields values strictly below Choi’s bounds in relevant cases. In most applications, this procedure efficiently determines the minimal necessary coupling strengths for stable minor embedding (Fang et al., 2019).

5. Empirical Validation and Superiority Relative to Previous Bounds

Experimental data from D-Wave hardware (DW2X, 2000Q) confirms that optimizing $|F|$ subject to the derived lower bound maximizes ground-state success probability. In Sherrington–Kirkpatrick spin glass experiments, the new bound closely tracks the measured optimum, whereas both Choi bounds substantially overestimate the necessary chain coupling (see Figure 6 in (Fang et al., 2019)). In job-shop scheduling, the true optimum for $|F|$ is around $0.6\,E_{\rm problem}$; Choi 2 prescribes the unnecessarily conservative $|F| \ge 0.75\,E_{\rm problem}$ (Figure 7). These results demonstrate that annealed lower bound optimization directly improves both physical problem normalization and ground-state yield, reducing resources for annealing and time-to-solution compared to prior practice.

6. Mathematical Summary and Theoretical Impact

The existence of analytically constructible, provably tight lower bounds for chain couplings in minor embedding is central to the efficiency of quantum annealing implementations. The closed-form bound: $$|F_i|\ge \max_{W\subset\iota(i)}\; \frac{1}{|\partial W|} \min\left\{|h(W)-J(W)|,\;|h(W)-h_i-J(\iota(i)\setminus W)|\right\}$$ is always no larger—and frequently strictly smaller—than previous bounds. This minimizes total embedding energy, reduces downward rescaling of the problem Hamiltonian, and yields higher success probabilities on current hardware (Fang et al., 2019). A plausible implication is that careful annealed lower bound optimization is necessary for hardware-efficient quantum annealing on nonnative graphs and for achieving practical speed-up in combinatorial optimization tasks.