We consider the problem of characterizing the set of input-output correlations that can be generated by an arbitrarily given quantum measurement. Our main result is to provide a closed-form, full characterization of such a set for any qubit measurement, and to discuss its geometrical interpretation. As applications, we further specify our results to the cases of real and complex symmetric, informationally complete measurements and mutually unbiased bases of a qubit, in the presence of isotropic noise. Our results provide the optimal device-independent tests of quantum measurements. DOI: 10.1103/PhysRevLett.118.250501 In operational quantum theory, it is a natural question to ask whether a given data sample, provided in the form of a conditional probability distribution representing the measured input-output correlation, is compatible with a particular hypothesis about the theoretical model underlying the experiment. A theoretical model can be more or less specific: for example, it could consist only of a general hypothesis about the theory describing the physics, as in the case in Bell tests [1][2][3], or it could be extremely detailed, as in the case of a tomographic reconstruction [4][5][6]. More generally, a hypothesis could be specific about a portion of the underlying model, while leaving the remaining elements completely uncharacterized.Here, we address hypotheses about the measurement producing the final outcomes of the experiment. This is the problem of characterizing the set SðπÞ of all input-output correlations p yjx ¼ Tr½ρ x π y compatible with an arbitrarily given quantum measurement π ≔ fπ y g (i.e., the hypothesis) and any family of input quantum states fρ x g, namely,We first note that a correlation p is compatible with measurement π if and only if for any fixed x there exists a state ρ x such that p yjx ¼ Tr½ρ x π y . Hence, SðπÞ is fully characterized by the range of π, namely, the set S 1 ðπÞ of output distributions q y ¼ Tr½ρπ y generated by π for varying input state ρ. This is in stark contrast with the analogous problem of characterizing the set of correlations compatible with a given quantum channel, which in general requires more than one input [7].Our main result is a closed-form characterization of the range S 1 ðπÞ, and hence of the set SðπÞ of all compatible correlations, when the hypothesis π is a qubit measurement. It turns out that an output distribution q y belongs to S 1 ðπÞ if and only ifwhere 1 is the identity matrix, t is the vector t y ≔ . Imposing the system of equalities in Eq. (2) causes linear dependencies-if any-among measurement elements π y to emerge as linear constraints on the probabilities. Such linear dependencies are present, for example, in any overcomplete measurement. Provided that these constraints are satisfied, the inequality in Eq. (2) recasts-through the transformation Q þ -the set of distributions compatible with π as an ellipsoid centered on distribution t. This in particular provides a simple and clear geometrical representation for the range of any q...