Bounds on adiabatic path geometry from the width class of the gap profile

Technical Summary: Bounds on Adiabatic Path Geometry from the Width Class of the Gap Profile

Problem Statement
In adiabatic quantum evolution, a system initialized in the ground state of a Hamiltonian $H(0)$ evolves to the ground state of $H(1)$ as a parameter $s$ is tuned from 0 to 1. The efficiency of this process is governed by the evolution time $T$, which is fundamentally constrained by the geometry of the "adiabatic path"—the curve traced by the ground state (or ground subspace) in Hilbert space. Two critical geometric quantities are the path length $L$ and the total curvature $K$.

Traditional analysis bounds these quantities using only the minimum energy gap $\Delta_*$ over the interval $s \in [0, 1]$. This approach yields $L = O(\Delta_*^{-1/2})$ and $K = O(\Delta_*^{-1})$. However, the authors observe that these bounds are often overly pessimistic in practice. The conventional method treats the gap function $\Delta(s)$ as a constant equal to its minimum, ignoring the functional form of the gap profile. Consequently, the integral determining the path geometry is overestimated, leading to loose bounds that do not reflect the actual behavior of many physical systems, such as avoided crossings.

Methodology
To address this discrepancy, the authors introduce the concept of the width class of the gap profile. Instead of relying solely on the depth of the gap minimum ($\Delta_*$), they characterize the "narrowness" of the region where the gap is small.

1. Width Class Definition: Let $\mu_<(\gamma)$ be the measure (width) of the set of parameters $s$ where the gap $\Delta(s) < \gamma$. A gap profile belongs to width class $p \ge 1$ if $\mu_<(\gamma)$ vanishes at least as fast as $\gamma^{1/p}$ as $\gamma \to 0$. Specifically, $\mu_<(\gamma) \le C_W (\gamma/\Gamma)^{1/p}$.

   • $p=1$: Corresponds to a linear vanishing of the width, typical of avoided crossings where the gap behaves like $\sqrt{\Delta_*^2 + \omega^2(s-s_*)^2}$.
   • $p>1$: Corresponds to slower vanishing widths, yielding power-law behaviors.

2. Theoretical Framework:

   • The authors analyze affine Hamiltonians of the form $H(s) = H_0 + sH'$ and generalize to non-affine, twice continuously differentiable Hamiltonians.
   • They define the path length $L_P$ and total curvature $K_P$ for the ground subspace of rank $g$.
   • Using the spectral projector $P(s)$ and its derivatives, they derive bounds for $L_P$ and $K_P$ in terms of the integral of $1/\Delta(s)$.
   • By integrating the width class condition, they evaluate the gap integral $\int_0^1 ds/\Delta(s)$ more precisely than the conventional $1/\Delta_*$ bound.

3. Evolution Time Analysis:

   • The derived geometric bounds are applied to two scheduling strategies: the standard linear schedule and the constant geometric speed (CGS) schedule.
   • The analysis incorporates the gap width class to refine the runtime bounds of the adiabatic theorem.

Key Contributions and Results

• Improved Bounds on Path Length ($L$) and Curvature ($K$):

  • For width class $p=1$ (typical avoided crossings):
    • $L = O(\sqrt{\log \Delta_*^{-1}})$
    • $K = O(\log \Delta_*^{-1})$
    • This replaces the conventional power-law scaling with logarithmic scaling.
  • For width class $p > 1$:
    • $L = O(\Delta_*^{-(p-1)/(2p)})$
    • $K = O(\Delta_*^{-(p-1)/p})$
    • These exponents are strictly smaller than the conventional $1/2$ and $1$, respectively.
  • The authors prove the tightness of these scaling laws for $L$ for all integers $p \ge 1$ and for $K$ at $p=1$ by constructing specific Hamiltonian families that attain these scalings.

• Bounds for Non-Affine Hamiltonians:

  • The results are extended to non-affine Hamiltonians by bounding the second derivative of the energy. The scaling behavior with respect to $\Delta_*$ remains unchanged, provided the family of Hamiltonians satisfies specific boundedness conditions on their derivatives.

• Boundedness in Fixed Subspaces:

  • The paper proves that if the ground state remains within a fixed two-dimensional subspace (independent of $s$), both $L$ and $K$ are bounded by constants ($L \le \pi/2$, $K=0$), regardless of the gap size. This explains why certain systems (like the Grover search) exhibit bounded geometry despite vanishing gaps.

• Runtime Improvements:

  • Linear Schedule: For width class $p$, the evolution time scales as $T = O(\Delta_*^{-(3-1/p)})$. At $p=1$, this improves the worst-case $O(\Delta_*^{-3})$ to $O(\Delta_*^{-2})$.
  • Constant Geometric Speed (CGS) Schedule:
    • If $L$ is bounded independently of $\Delta_*$, the CGS schedule achieves $T = O(\Delta_*^{-(2-1/p)})$. At $p=1$, this yields $T = O(\Delta_*^{-1})$ (up to polylogarithmic corrections), improving the linear schedule by one order of magnitude.
    • Even when $L$ is unbounded, the CGS schedule provides improved scaling compared to the linear schedule for all $p$.

Applications Demonstrated
The authors validate their theoretical bounds on three specific systems, all of which fall into width class $p=1$:

1. Adiabatic Grover Search: The ground state stays in a fixed 2D subspace. The computed path length approaches $\pi/2$, consistent with the constant bound, while the new theoretical bound ($O(\sqrt{\log \Delta_*^{-1}})$) is significantly tighter than the conventional $O(\Delta_*^{-1/2})$.
2. XXZ Spin Chain: The ground state remains largely within a plane spanned by two Néel orderings. The computed $L$ stays bounded (near $\pi/4$), and the new bounds are orders of magnitude tighter than conventional estimates.
3. Molecular Electronic Hamiltonians: Applied to the Nitrogen molecule ($N_2$) and a [2Fe-2S] cluster. The computed path lengths and curvatures are bounded by factors of 2–7 and 5–95 times the new bounds, respectively, whereas the conventional bounds exceed the computed values by factors of $10^3$ to $10^4$.

Significance
The paper argues that the conventional bounds on adiabatic path geometry are too conservative because they reduce the entire gap profile to a single number ($\Delta_*$). By introducing the width class, the authors provide a more nuanced characterization of the gap profile that captures how the small-gap region behaves.

The significance lies in:

1. Tighter Theoretical Guarantees: The new bounds are provably tighter and often logarithmic rather than polynomial, offering a more accurate prediction of the resources required for adiabatic evolution.
2. Practical Relevance: The results explain why numerical experiments often show much better performance than worst-case theoretical predictions suggest.
3. Algorithm Design: The findings support the use of the constant geometric speed schedule, which requires only a lower bound on the gap (rather than the full gap function) to achieve optimal scaling, thereby reducing the prior spectral knowledge needed to design efficient adiabatic algorithms.

The authors conclude that while the width class improves bounds to polylogarithmic levels for typical avoided crossings, the actual path geometry in many physical systems appears to be bounded by constants, a phenomenon they attribute to geometric constraints (fixed subspaces) rather than spectral ones. They identify the mechanism for this constant boundedness and the tightness of curvature scaling for $p>1$ as open directions for future research.