To test the capacity of the MesoNh model to predict seeing, we used Scidar as the reference. Scidar has been tested in many ways, in various sites and compared with other techniques such as DIMM and balloons. Moreover, in Sect. 2 we proved that during the PARSCA93 campaign the correlation between Scidar and DIMMESO was good. We compared Scidar and MesoNh seeing values during each selected night and we analyzed the correlation between the measurements and simulations. During the worst seeing night we noticed that the seeing increased suddenly after about 3 hours due to the occurrence of a layer at 4 km, also seen by the Scidar. Thus, we decided to make 4 hours simulations for each night and we analyzed all the forecasts ranging between 30 min and 4 hours. The first 30 min are discarded because we verified that, for all the simulations, the flow is not yet adapted to the orography. Being aware that our statistical sample is poor (8 nights only) we tried to extract the most complete information with the available data using different techniques. We used two methods that will be named Method A and Method B. For both methods we calculated a linear regression fit and computed the following statistical estimators:
(1) 
(2) 
To estimate the reliability of the correlation coefficient we computed the probability P that two uncorrelated distributions of the analyzed data give a correlation coefficient greater than that observed.
(3) 


In this method we compare the seeing deduced from Scidar measurements with the MesoNh simulations above Paranal every 2.5 s. In order to have a better estimation of the correct adaptation time, we considered the seeing averaged over different time intervals and we computed the statistical parameters defined before for each interval. In Table 2 we report the statistical results obtained over 6 different time intervals over a complete 4 simulation time: [], [], [], [], [] and finally []. This method gives good temporal statistics but does not give any information about the vertical structure of the optical turbulence.
As we are interested not only in the seeing prediction but also in the turbulent profile prediction, for each night we compared the averaged profiles from the Scidar with those obtained from the MesoNh output every 30 min. So doing, we have direct access to the temporal evolution of profile, that is the optical turbulence evolution at all the model levels.
For each night, we splitted the atmosphere into two regions: we computed the contribution of the boundary layer (BL) defined here between ground level and 5 km and that of the free atmosphere (FA) above 5 km. The same splitting has been used for both Scidar and MesoNh. As we are not completely confident in the ability of the Scidar to measure the optical turbulence in the surface layer (first hundred of meters) nor in the MesoNh model, we used two sets of MesoNh outputs, with and without the surface layer in order to evaluate the sensitivity of the numerical model to the orographic effect. We thus defined a , a and a in the following way
(4) 
(5) 
(6) 
We underline that the ground level altitude is 2560 m and not 2640 m (the true Paranal altitude) because of an average effect due to the horizontal model resolution used. The second method B is less well statistically defined than the method A. We can average, in fact, only 4 profiles for each night related to the 1, 2, 3 and 4 outputs, but we can analyze the model sensitivity in the first 100 m. In Table 3 and Table 4 are reported the statistical results for two different configurations. We estimated that this test was necessary because we often found that a strong layer was produced by the model at this low altitude. At the moment we have no a priori reasons to reject or accept this contribution because we know that the Scidar sensitivity at this altitude is poor. Scidar is based on scintillation measurements and it is particularly sensitive to the high troposphere turbulence. The Generalized configuration is sensitive to the low levels turbulence too. During this campaign only a Classic version of Scidar was employed.

Having a small amount of data, the correlation coefficient is a poor estimator for deciding whether an observed correlation is statistically significant or not. tells how good is the fit to a straight line but ignores of the individual distributions x_{i} and y_{i}.
We therefore compute the probability P that two uncorrelated distributions x_{i} and y_{i} (belonging to the same parental distribution) give a correlation coefficient greater than that found. A large P means that is poor, while a small P means that is good. A classic test (Press et al. 1989), adapted for a small amount of data, (N < 20) gives the results reported in the Tables 24.
The standard deviation of data from the linear regression line might estimate in a complementary way, the data dispersion around the optimized regression line. We report the values in Tables 24.
In both methods A and B we made a calibration of the MesoNh outputs using the following procedure. We compute the mean value of Scidar seeing measured and the MesoNh simulated seeing All the simulated values are multiplied by the calibration coefficient , before doing any statistical analysis as described in Method A and B. In Method B, the calibration coefficient is named .
Copyright The European Southern Observatory (ESO)