The control landscape is the figure of merit viewed as a function of the controls, \(u(\cdot) \mapsto J[u]\) — an extremely high-dimensional surface over the space of all pulses. Its shape decides how hard an optimization is. If every local optimum is also a global one, any local method such as GRAPE or Krotov will succeed from almost any starting point. If there are traps — local optima that are not global — the optimizer can get stuck. A central, and at first surprising, result of the field is that for many quantum control problems the landscape is trap-free. This explains why optimal control works so well in practice.

Kinematic and dynamical landscapes

The map from controls to the figure of merit can be split into two steps:

$$u(\cdot)\;\xrightarrow{\;\text{dynamics}\;}\;U(T)\;\xrightarrow{\;\text{kinematics}\;}\;J\big(U(T)\big).$$

The kinematic landscape \(J(U)\) is a function on the unitary group, for example the transition probability \(|\langle f|U|i\rangle|^2\) or an expectation value \(\mathrm{Tr}\big(O\,U\rho U^\dagger\big)\). Its critical points can be classified completely: they are global maxima, global minima and saddles, with no local optima [1]. The dynamical landscape \(J[u]\) inherits this structure if the first map is well behaved [2, 3].

When is the landscape trap-free?

Three assumptions guarantee that there are no traps [4]:

  1. the system is controllable;
  2. the map from controls to \(U(T)\) is locally surjective — its Jacobian has full rank at the critical points;
  3. the controls are unconstrained.

Traps: where the assumptions fail

Singular controls. Where the Jacobian loses rank — at singular controls, the singular extremals of Pontryagin's maximum principle — the kinematic argument does not apply. A three-level example shows that such points can indeed be traps [5], and the assumptions of landscape analysis have been examined critically [6]. A later geometric analysis proved that almost all closed, finite-dimensional systems are nonetheless trap-free in a broad class of situations, including qubit evolutions in quantum computing. Being singular is not by itself enough for a control to stop an optimization [7].

Constraints. The practically most important source of traps is the third assumption. Real optimizations restrict the controls:

  • few parameters — a truncated basis, as in CRAB, or a coarse time grid;
  • short durations — close to the minimal time;
  • bounded amplitudes or bandwidth.

Simulations show that such constraints can create artificial traps, severe enough to prevent successful optimization [4]. The minimum number of controls needed for a high yield depends on the number of accessible states [8]. Even a two-level system develops a rich landscape structure when the total time approaches the speed limit [9].

Near the minimal time. In many-body state preparation, the landscape changes character as the protocol duration decreases. A glass-like transition appears: an exponential number of near-optimal solutions separated by barriers, with the true optimum exponentially hard to find [10, 11]. See Quantum speed limits and Many-body optimal control.

What it means for optimization

  • Why local methods usually work. With enough control freedom and enough time, a gradient method from a random start reaches a global optimum. This is the common experience with GRAPE and Krotov.
  • When to expect trouble. Few parameters, short times, strong bounds: use restarts, continuation in time or parameters, or methods that escape false traps. dCRAB changes its basis to reach the trap-free landscape [12]; global methods find the best solution of the constrained problem directly (QCPOP, Gradient-free & global optimizers).
  • Beyond quantum control. The same questions — how controllability shapes the landscape — decide the trainability of variational quantum algorithms, where the dimension of the dynamical Lie algebra determines whether gradients vanish (barren plateaus) [13]. See Variational algorithms & QAOA as control.