We introduce a self-learning tomographic technique in which the experiment guides itself to an estimate of its own state. Self-guided quantum tomography (SGQT) uses measurements to directly test hypotheses in an iterative algorithm which converges to the true state. We demonstrate through simulation on many qubits that SGQT is a more efficient and robust alternative to the usual paradigm of taking a large amount of informationally complete data and solving the inverse problem of post-processed state estimation.The act of inferring a quantum mechanical description of a physical system-assigning it a quantum state-is referred to as tomography. Tomography is a required, and now routine, task for designing, testing and tuning qubits-the building blocks of a quantum information processing device [1]. However, in a grand irony, the exact same exponential scaling that gives a quantum information processing device its power also limits our ability to characterize it.That tomography is a problem exponentially hard in the number of qubits has lead to many proposals for efficient learning within restricted subsets of quantum states [2,3]. On the other hand, if we expect to have prepared a specific target state, efficient protocols exist to estimate the fidelity to this state [4,5]. These protocols are direct in the sense that few measurements are required to provide an estimate of the fidelity to the target rather than first reconstructing the state then calculating the fidelity.The proposal presented here is direct in the same sense as [4,5], but converges to the state itself. The algorithm is iterative-from directly estimating a distance measure to the underlying state, the experiment guides itself to a description of its own state: self-guided quantum tomography (SGQT). The distance measure m is arbitrary, in the sense that any measure will work. However, the more rapidly the experiment can provide an estimate for m, the more rapidly SGQT will converge, such that if m can be estimated efficiently, then SGQT will be efficient.Before describing exactly what SGQT is, we first state what it is not by reviewing the problem of tomography. There is some true state ρ which generates data, a list of measurement outcomes corresponding to effects D = {E 0 , E 1 , . . .}. The probability to observe this data is given by the Born ruleThe prevailing method is to solve the inverse problem of identifying an accurate estimate, σ, of ρ given a sample data set drawn from this distribution. Here the approach is quite different. We begin with a distance measure on states m(ρ, σ). The only requirement is that this measure can be estimated from experiment, such that we have access toThis quantity fluctuates from noise which can come from a variety of sources but is always present due to the fundamental statistical nature of quantum mechanics-also known as shot-noise.Here we will provide an algorithm to iteratively propose new states σ such that we converge to ρ only via estimates of f (σ). The core of the algorithm is a stochastic optimizatio...