We propose a change of style for numerical estimations of physical quantities from measurements to inferences. We estimate the most probable quantities for all the parameter region simultaneously by using the raw data cooperatively. Estimations with higher precisions are made possible. We can obtain a physical quantity as a continuous function, which is differentiated to obtain another quantity. We applied the method to the Heisenberg spin-glass model in three dimensions. A dynamic correlation-length scaling analysis suggests that the spin-glass and the chiral-glass transitions occur at the same temperature with a common exponent ν. The value is consistent with the experimental results. We found that a size-crossover effect explains a spin-chirality separation problem.Introduction-Estimations of physical quantities in numerical simulations are based on equilibrium statistical physics [1]. We virtualize a model system in a computer and perform independent measurements on the system using a definition of a physical quantity. When an evaluation process is complex, both systematic and statistical errors are accumulated in the obtained data. We sometimes encounter numerical instabilities, which may affect a final physical conclusion. In what follows, we explain the situation of interest using a correlation-length estimation.An estimation formula for a correlation length, ξ, is given by the second-moment method:. Here, χ 0 denotes the susceptibility and χ k its Fourier transform with k as the lowest wave number of the system. This expression itself is problematic. Both numerator and denominator of this expression approach zero as the system size increases (k → 0), where this formula becomes exact. We encounter the numerical instability caused by the expression 0/0. In order to avoid this problem, Bellettiet al. [3] proposed the reduction of this instability by estimating ξ through the integrals I k = 0 drr k f (r) and ξ is obtained as I 2 /I 1 . Suwa and Todo [4] proposed a generalized moment method for gap (∆ ∼ 1/ξ) estimation in quantum systems. Systematic errors and ambiguity caused by using small-L data are eliminated.Recently, big-data handling has become possible due to rapid increase in computational power. Data science is now one of the most promising fields in science and technology. As regards its application to physics, the topic of Bayesian inference has attracted considerable interest [5,6]. In this context, Harada [7] introduced Bayesian inference into a parameter estimation of the finite-size scaling analysis.In this paper, we extend its application to estimations of physical quantities. For example, we can obtain an analytic expression for an energy out of the discrete raw data as the most-probable model function. Then, we obtain the specific heat by analytically differentiating it. A critical temperture is estimated automatically within this procedure. Since directly-observed (raw) data are