For quantum systems with high purity, we find all observables that, when continuously monitored, maximize the instantaneous reduction in the average linear entropy. This allows us to obtain all locally optimal feedback protocols with strong feedback, and explicit expressions for the best such protocols for systems of size N ≤ 4. We also show that for a qutrit the locally optimal protocol is the optimal protocol for observables with equispaced eigenvalues, providing the first fully optimal feedback protocol for a 3-state system. PACS numbers: 03.65. Yz, 87.19.lr, 02.30.Yy, 03.65.Ta Observation and control of coherent quantum behavior has been realized in a variety of mesoscopic devices [1, 2, 3, 4]. With further refinements, such devices may well form the basis of new technologies, for example in sensing [5] and communication [6,7]. Feedback, in which a system is continuously observed and the information used to control its behavior in the presence of noise, is an important element in the quantum engineer's toolbox [6,7,8,9,10,11]. In view of this, one would like to know the limits on such control, given any relevant limitations on the measurement and/or control forces. However, except in special cases [12], the dynamics of continuously observed quantum systems is nonlinear. Further, results on the quantum-to-classical transition show that this nonlinear dynamics, described by stochastic master equations (SME's), is necessarily every bit as complex (and chaotic) as that of nonlinear classical systems [13]. Because of this, fully general and exact results regarding optimal quantum feedback are unlikely to exist; certainly no such results have been found for nonlinear classical systems [14]. Nevertheless, one would like to obtain results that give insights applicable across a range of systems.Quantum feedback control is implemented by modifying a "control" Hamiltonian, H, that is some part of the system Hamiltonian. Here we will examine feedback protocols in the regime where the controls are able to keep the system close to a pure state. This is an important regime, both because it is where many quantum control systems will need to operate, and because it allows one to simplify the problem by using a power series expansion [15]. In addition to working in the regime of good control, we make two further simplifications. The first is that the control is strong -that is, that 1) the only constraint on H is that Tr[H 2 ] ≤ µ 2 for some constant µ, and 2) that H can induce dynamics much faster than both the dynamics of the system and the rate at which the measurement extracts information. This means that H is effectively unconstrained. We thus deal strictly with a subset of the regime of good control, defined by µ ≫ k and k ≫ β. Here k is the strength of the measurement (defined precisely below), and β is the noise strength, which we define as the rate of increase of the linear entropy due to the noise. The latter inequality is essential for good control. This regime is applicable, for example, to mesoscopic supercond...