CRAB — the Chopped RAndom Basis method — finds good controls with very few parameters. Instead of optimizing the value of the field at every time step, it expands the correction to a pulse in a small basis of smooth functions with randomized frequencies, and optimizes the handful of expansion coefficients, typically with a gradient-free optimizer. CRAB was introduced by Doria, Calarco and Montangero to control many-body dynamics simulated with tensor networks [1]. Its refinement dCRAB (“dressed” CRAB) is now one of the standard methods of the field [2].

The idea

Many control problems do not need a fully flexible pulse: the set of physically relevant controls is often well described by a few smooth components. CRAB writes the pulse as a guess \(u_0(t)\) times a correction expanded in a truncated basis,

$$u(t) = u_0(t)\left[1 + \frac{1}{\lambda(t)}\sum_{n=1}^{N_c} \Big(A_n \sin(\omega_n t) + B_n \cos(\omega_n t)\Big)\right],$$

where \(\lambda(t)\) enforces boundary conditions — for example that the correction vanishes at \(t=0\) and \(t=T\) [3]. The frequencies are the Fourier harmonics with a random detuning,

$$\omega_n = \frac{2\pi n\,(1 + r_n)}{T}, \qquad r_n \in [-\tfrac12, \tfrac12] \text{ random},$$

and the \(2N_c\) coefficients \(\{A_n, B_n\}\) (sometimes also the \(\omega_n\)) are optimized. With \(N_c\) of order ten, the search space is small enough for direct-search methods such as Nelder–Mead [4]. Those methods only need the value of the figure of merit, not its gradient.

Why so few parameters can suffice has an information-theoretic answer. The number of independent control parameters needed is set by the dimension of the manifold of states the dynamics can actually reach in the available time, and for physical many-body dynamics this can be far smaller than the dimension of the Hilbert space [5]. See Control complexity & resources.

dCRAB: escaping the traps of a truncated basis

Truncating the basis has a price: the restricted optimization can get stuck in “false traps” that are not local optima of the full problem. dCRAB removes them by working in super-iterations. After optimizing the coefficients in one random basis, it keeps the resulting pulse and adds a correction in a new random basis,

$$u^{(j)}(t) = u^{(j-1)}(t) + \sum_{n=1}^{N_c} c_n^{(j)} f_n^{(j)}(t),$$

so that, over many super-iterations, the pulse explores an unbounded function space while each individual optimization stays low-dimensional. Rach et al. showed that this gives access to the trap-free landscape of the full problem while the bandwidth stays limited [6]; see Control landscapes.

Why it is useful

  • Few parameters, smooth pulses. A small Fourier-like basis builds in a bandwidth limit, so pulses are naturally smooth and experimentally friendly; see Pulse parametrization & constraints.
  • No gradients needed. The figure of merit can come from any simulation — tensor networks, mean-field or stochastic methods — or from the experiment itself. This makes CRAB a natural choice for closed-loop optimization.
  • Many-body problems. Combined with tensor-network simulations, it was designed for interacting systems from the start; see Many-body optimal control.
  • Remote and citizen-science optimization. In its remote form (RedCRAB), dCRAB has been used to optimize an ultracold-atom experiment over the internet, alongside citizen scientists [7].

Strengths and limitations

Strengths

  • Small, smooth, bandwidth-limited search space with interpretable pulses.
  • Works with any figure of merit, including noisy experimental ones.
  • dCRAB avoids the false traps of a fixed truncated basis.

Limitations

  • Gradient-free optimizers scale poorly with the number of parameters, so each super-iteration must stay small.
  • When gradients are cheap, gradient-based methods (GRAPE, GOAT) usually converge in fewer evaluations. CRAB’s basis can also be used with gradients, and modern suites support both [8].
  • Performance depends on choosing the guess pulse, basis type and number of components sensibly.