From 0819305edd2df8e4890deb94fd112c78e2e9400a Mon Sep 17 00:00:00 2001 From: mhjensen Date: Wed, 23 Dec 2020 23:08:51 +0100 Subject: [PATCH] added files --- doc/src/Bayesian/chapter11.dlog | 29 ++ doc/src/Bayesian/chapter11.do.txt | 444 +++++++++++++++++++ doc/src/Bayesian/chapter11.do.txt~ | 449 ++++++++++++++++++++ doc/src/Bayesian/ipynb-chapter11-src.tar.gz | Bin 0 -> 85179 bytes 4 files changed, 922 insertions(+) create mode 100644 doc/src/Bayesian/chapter11.dlog create mode 100644 doc/src/Bayesian/chapter11.do.txt create mode 100644 doc/src/Bayesian/chapter11.do.txt~ create mode 100644 doc/src/Bayesian/ipynb-chapter11-src.tar.gz diff --git a/doc/src/Bayesian/chapter11.dlog b/doc/src/Bayesian/chapter11.dlog new file mode 100644 index 000000000..2c8957f3d --- /dev/null +++ b/doc/src/Bayesian/chapter11.dlog @@ -0,0 +1,29 @@ +translating doconce text in chapter11.do.txt to ipynb +copying after regex "start import modules" until "end import modules" + file: src/coinflipping.py, lines 2-4 (format: pycod) +copying after regex "start simulation" until "end simulation" + file: src/coinflipping.py, lines 9-13 (format: pycod) +copying after regex "start bayesian setup" until "end bayesian setup" + file: src/coinflipping.py, lines 16-29 (format: pycod) +copying after regex "start plotting" until "end plotting" + file: src/coinflipping.py, lines 32-42 (format: pycod) +copying after regex "start generate data" until "end generate data" + file: src/singlephotoncount.py, lines 11-17 (format: pycod) +copying after regex "start visualize data" until "end visualize data" + file: src/singlephotoncount.py, lines 20-24 (format: pycod) +copying after regex "start frequentist" until "end frequentist" + file: src/singlephotoncount.py, lines 30-35 (format: pycod) +copying after regex "start bayesian setup" until "end bayesian setup" + file: src/singlephotoncount.py, lines 38-47 (format: pycod) +copying after regex "start bayesian mcmc" until "end bayesian mcmc" + file: src/singlephotoncount.py, lines 50-61 (format: pycod) +copying after regex "start visualize bayesian" until "end visualize bayesian" + file: src/singlephotoncount.py, lines 64-68 (format: pycod) +copying after regex "start bayesian CI" until "end bayesian CI" + file: src/singlephotoncount.py, lines 79-90 (format: pycod) +*** warning: footnotes are not supported for format ipynb + footnotes will be left in the doconce syntax +collected all required additional files in ipynb-chapter11-src.tar.gz which must be distributed with the notebook +Failed to remove ans_at_end environment +Failed to remove sol_at_end environment +output in chapter11.ipynb diff --git a/doc/src/Bayesian/chapter11.do.txt b/doc/src/Bayesian/chapter11.do.txt new file mode 100644 index 000000000..6f730a159 --- /dev/null +++ b/doc/src/Bayesian/chapter11.do.txt @@ -0,0 +1,444 @@ +TITLE: Data Analysis and Machine Learning: +AUTHOR: Christian Forssén at Department of Physics, Chalmers University of Technology, Sweden +AUTHOR: Morten Hjorth-Jensen {copyright, 1999-present|CC BY-NC} at Department of Physics, University of Oslo & Department of Physics and Astronomy and National Superconducting Cyclotron Laboratory, Michigan State University +DATE: today + + +======= Elements of Bayesian theory and Bayesian Neural Networks ======= + + +===== Why Bayesian Statistics? ===== +!bblock + +We have already made ourselves familiar with elements of a statistical +data analysis via quantities like the bias-variance tradeoff as well +as some central distribution functions such as the Normal +distribution, the binomial distribution and other probability +distribution functions. + +In essentially all the Machine Learning algorithms we have studied, +our focus has been on a so-called _frequentist approach_, where +knowledge of an underlying likelihood function has not been +emphasized. Our data, whether we had a classification or a regression +problem, have been our central points of departure. + +Here we wish to merge this approach with the derivation of a +likelihood function which can be used to make prediction on how our +system under study evolves. We will venture into the realm of what is +called Bayesian Neural Networks. To get an overarching view on what +this entails, the following figure conveys the essential differences +between a standard Neural network that we have met earlier and a +Bayesian Neural Network. In order to get there, we need to present +some of the basic elements of Bayesian statistics, starting with the +product rule and Bayes' theorem. + +!eblock + + +===== Inference ===== +!bblock +- Inference: ``the act of passing from one proposition, statement or judgment considered as true to another whose truth is believed to follow from that of the former'' (Webster) + Do premises $A, B, \ldots \to$ hypothesis, $H$? + +- Deductive inference: Premises allow definite determination of truth/falsity of H (syllogisms, symbolic logic, Boolean algebra) + $B(H|A,B,...) = 0$ or $1$ + +- Inductive inference: Premises bear on truth/falsity of H, but don’t allow its definite determination (weak syllogisms, analogies) + $A, B, C, D$ share properties $x, y, z$; $E$ has properties $x, y$ + $\to$ $E$ probably has property $z$. +!eblock + + +===== Statistical Inference ===== +!bblock +* Quantify the strength of inductive inferences from facts, in the form of data ($D$), and other premises, e.g. models, to hypotheses about the phenomena producing the data. +* Quantify via probabilities, or averages calculated using probabilities. Frequentists ($\mathcal{F}$) and Bayesians ($\mathcal{B}$) use probabilities very differently for this. +* To the pioneers such as Bernoulli, Bayes and Laplace, a probability represented a *degree-of-belief* or plausability: how much they thought that something as true based on the evidence at hand. This is the Bayesian approach. +* To the 19th century scholars, this seemed too vague and subjective. They redefined probability as the *long run relative frequency* with which an event occurred, given (infinitely) many repeated (experimental) trials. +!eblock + + +===== Some history ===== +Adapted from D.S. Sivia[^Sivia]: + +[^Sivia]: Sivia, Devinderjit, and John Skilling. Data Analysis : A Bayesian Tutorial, OUP Oxford, 2006 + +!bquote +Although the frequency definition appears to be more objective, its range of validity is also far more limited. For example, Laplace used (his) probability theory to estimate the mass of Saturn, given orbital data that were available to him from various astronomical observatories. In essence, he computed the posterior pdf for the mass M , given the data and all the relevant background information I (such as a knowledge of the laws of classical mechanics): prob(M|{data},I); this is shown schematically in the figure [Fig. 1.2]. +!equote + + +FIGURE: [fig/sivia_fig_1_2.png, width=700 frac=0.9] + + +!bquote +To Laplace, the (shaded) area under the posterior pdf curve between $m_1$ and $m_2$ was a measure of how much he believed that the mass of Saturn lay in the range $m_1 \le M \le m_2$. As such, the position of the maximum of the posterior pdf represents a best estimate of the mass; its width, or spread, about this optimal value gives an indication of the uncertainty in the estimate. Laplace stated that: ‘ . . . it is a bet of 11,000 to 1 that the error of this result is not 1/100th of its value.’ He would have won the bet, as another 150 years’ accumulation of data has changed the estimate by only 0.63%! +!equote + + +!bquote +According to the frequency definition, however, we are not permitted to use probability theory to tackle this problem. This is because the mass of Saturn is a constant and not a random variable; therefore, it has no frequency distribution and so probability theory cannot be used. + +If the pdf [of Fig. 1.2] had to be interpreted in terms of the frequency definition, we would have to imagine a large ensemble of universes in which everything remains constant apart from the mass of Saturn. +!equote + + +!bquote +As this scenario appears quite far-fetched, we might be inclined to think of [Fig. 1.2] in terms of the distribution of the measurements of the mass in many repetitions of the experiment. Although we are at liberty to think about a problem in any way that facilitates its solution, or our understanding of it, having to seek a frequency interpretation for every data analysis problem seems rather perverse. +For example, what do we mean by the ‘measurement of the mass’ when the data consist of orbital periods? Besides, why should we have to think about many repetitions of an experiment that never happened? What we really want to do is to make the best inference of the mass given the (few) data that we actually have; this is precisely the Bayes and Laplace view of probability. +!equote + + +!bquote +Faced with the realization that the frequency definition of probability theory did not permit most real-life scientific problems to be addressed, a new subject was invented — statistics! To estimate the mass of Saturn, for example, one has to relate the mass to the data through some function called the statistic; since the data are subject to ‘random’ noise, the statistic becomes the random variable to which the rules of probability the- ory can be applied. But now the question arises: How should we choose the statistic? The frequentist approach does not yield a natural way of doing this and has, therefore, led to the development of several alternative schools of orthodox or conventional statis- tics. The masters, such as Fisher, Neyman and Pearson, provided a variety of different principles, which has merely resulted in a plethora of tests and procedures without any clear underlying rationale. This lack of unifying principles is, perhaps, at the heart of the shortcomings of the cook-book approach to statistics that students are often taught even today. +!equote + + +===== The Bayesian recipe ===== +!bblock +Assess hypotheses by calculating their probabilities $p(H_i | \ldots)$ conditional on known and/or presumed information using the rules of probability theory. +!eblock +!bblock +Probability Theory Axioms: +- Product (AND) rule : $p(A, B | I) = p(A|I) p(B|A, I) = p(B|I)p(A|B,I)$ + Should read $p(A,B|I)$ as the probability for propositions $A$ AND $B$ being true given that $I$ is true. +- Sum (OR) rule: $p(A + B | I) = p(A | I) + p(B | I) - p(A, B | I)$ + $p(A+B|I)$ is the probability that proposition $A$ OR $B$ is true given that $I$ is true. +- Normalization: $p(A|I) + p(\bar{A}|I) = 1$ + $\bar{A}$ denotes the proposition that $A$ is false. + +!eblock + + +===== Bayes' theorem ===== +!bblock +Bayes' theorem follows directly from the product rule +!bt +$$ +p(A|B,I) = \frac{p(B|A,I) p(A|I)}{p(B|I)}. +$$ +!et +The importance of this property to data analysis becomes apparent if we replace $A$ and $B$ by hypothesis($H$) and data($D$): +!bt +\begin{align} +p(H|D,I) &= \frac{p(D|H,I) p(H|I)}{p(D|I)}. +label{eq:bayes} +\end{align} +!et +The power of Bayes’ theorem lies in the fact that it relates the quantity of interest, the probability that the hypothesis is true given the data, to the term we have a better chance of being able to assign, the probability that we would have observed the measured data if the hypothesis was true. +!eblock + + +!bblock +The various terms in Bayes’ theorem have formal names. +* The quantity on the far right, $p(H|I)$, is called the *prior* probability; it represents our state of knowledge (or ignorance) about the truth of the hypothesis before we have analysed the current data. +* This is modified by the experimental measurements through $p(D|H,I)$, the *likelihood* function, +* The denominator $p(D|I)$ is called the *evidence*. It does not depend on the hypothesis and can be regarded as a normalization constant. +* Together, these yield the *posterior* probability, $p(H|D, I )$, representing our state of knowledge about the truth of the hypothesis in the light of the data. + +In a sense, Bayes’ theorem encapsulates the process of learning. +!eblock + + +===== The friends of Bayes' theorem ===== +!bblock +- Normalization: $\sum_i p(H_i|\ldots) = 1$. +- Marginalization: $\sum_i p(A,H_i|I) = \sum_i p(H_i|A,I) p(A|I) = p(A|I)$. +- Marginalization (continuum limit): $\int dx p(A,H(x)|I) = p(A|I)$. + +In the above, $H_i$ is an exclusive and exhaustive list of hypotheses. For example,let’s imagine that there are five candidates in a presidential election; then $H_1$ could be the proposition that the first candidate will win, and so on. The probability that $A$ is true, for example that unemployment will be lower in a year’s time (given all relevant information $I$, but irrespective of whoever becomes president) is then given by $\sum_i p(A,H_i|I)$. + +In the continuum limit of propositions we must understand $p(\ldots)$ as a pdf (probability density function). + +Marginalization is a very powerful device in data analysis because it enables us to deal with nuisance parameters; that is, quantities which necessarily enter the analysis but are of no intrinsic interest. The unwanted background signal present in many experimental measurements are examples of nuisance parameters. +!eblock + + +===== Inference With Parametric Models ===== +!bblock +Inductive inference with parametric models is a very important tool in the natural sciences. +* Consider $N$ different models $M_i$ ($i = 1, \ldots, N$), each with parameters $\boldsymbol{\alpha}_i$. Each of them implies a sampling distribution (conditional predictive distribution for possible data) +!bt +$$ +p(D|\boldsymbol{\alpha}_i, M_i) +$$ +!et +* The $\boldsymbol{\alpha}_i$ dependence when we fix attention on the actual, observed data ($D_\mathrm{obs}$) is the likelihood function: +!bt +$$ +\mathcal{L}_i (\boldsymbol{\alpha}_i) \equiv p(D_\mathrm{obs}|\boldsymbol{\alpha}_i, M_i) +$$ +!et +* We may be uncertain about $i$ (model uncertainty), +* or uncertain about $\boldsymbol{\alpha}_i$ (parameter uncertainty). +!eblock + + +!bblock +- Parameter Estimation: Premise = choice of model (pick specific $i$) + $\Rightarrow$ What can we say about $\boldsymbol{\alpha}_i$? +- Model comparison: Premise = $\{M_i\}$ + $\Rightarrow$ What can we say about $i$? +- Model adequacy: Premise = $M_1$ + $\Rightarrow$ Is $M_1$ adequate? +- Hybrid Uncertainty: Models share some common params: $\boldsymbol{\alpha}_1 = \{ \boldsymbol{\varphi}, \boldsymbol{\eta}_i\}$ + $\Rightarrow$ What can we say about $\boldsymbol{\varphi}$? (Systematic error is an example) +!eblock + + +===== Illustrative examples with python code ===== +!bblock +* Is this a fair coin? (analytical) +* Flux from a star (single parameter, MCMC) +* The lighthouse problem (two parameters, MCMC) +* Linear fit with outliers (nuisance parameters) +* ... +!eblock + + +===== Example: Is this a fair coin? ===== +Let us begin with the analysis of data from a simple coin-tossing experiment. +Given that we had observed 6 heads in 8 flips, would you think it was a fair coin? By fair, we mean that we would be prepared to lay an even 1 : 1 bet on the outcome of a flip being a head or a tail. If we decide that the coin was fair, the question which follows naturally is how sure are we that this was so; if it was not fair, how unfair do we think it was? Furthermore, if we were to continue collecting data for this particular coin, observing the outcomes of additional flips, how would we update our belief on the fairness of the coin? + +A sensible way of formulating this problem is to consider a large number of hypotheses about the range in which the bias-weighting of the coin might lie. If we denote the bias-weighting by $H$, then $H = 0$ and $H = 1$ can represent a coin which produces a tail or a head on every flip, respectively. There is a continuum of possibilities for the value of H between these limits, with $H = 0.5$ indicating a fair coin. Our state of knowledge about the fairness, or the degree of unfairness, of the coin is then completely summarized by specifying how much we believe these various propositions to be true. + +Let us perform a computer simulation of a coin-tossing experiment. This provides the data that we will be analysing. + +@@@CODE src/coinflipping.py from-to: start import modules@end import modules +@@@CODE src/coinflipping.py from-to: start simulation@end simulation + +In the light of this data, our inference about the fairness of this coin is summarized by the conditional pdf: $p(H|D,I)$. This is, of course, shorthand for the limiting case of a continuum of propositions for the value of $H$; that is to say, the probability that $H$ lies in an infinitesimally narrow range is given by $p(H|D,I) dH$. + +To estimate this posterior pdf, we need to use Bayes’ theorem (ref{eq:bayes}). We will ignore the denominator $p(D|I)$ as it does not involve bias-weighting explicitly, and it will therefore not affect the shape of the desired pdf. At the end we can evaluate the missing constant subsequently from the normalization condition +!bt +\begin{equation} +\int_0^1 p(H|D,I) dH = 1. +label{eq:coin_posterior_norm} +\end{equation} +!et + +The prior pdf, $p(H|I)$, represents what we know about the coin given only the information $I$ that we are dealing with a ‘strange coin’. We could keep a very open mind about the nature of the coin; a simple probability assignment which reflects this is a uniform, or flat, prior +!bt +\begin{equation} +p(H|I) = \left\{ \begin{array}{ll} +1 & 0 \le H \le 1, \\ +0 & \mathrm{otherwise}. +\end{array} \right. +label{eq:coin_prior_uniform} +\end{equation} +!et +We will get back later to the choice of prior and its effect on the analysis. + +This prior state of knowledge, or ignorance, is modified by the data through the likelihood function $p(D|H,I)$. It is a measure of the chance that we would have obtained the data that we actually observed, if the value of the bias-weighting was given (as known). If, in the conditioning information $I$, we assume that the flips of the coin were independent events, so that the outcome of one did not influence that of another, then the probability of obtaining the data `R heads in N tosses' is given by the binomial distribution (we leave a formal definition of this to a statistics textbook) +!bt +\begin{equation} +p(D|H,I) \propto H^R (1-H)^{N-R}. +\end{equation} +!et +It seems reasonable because $H$ is the chance of obtaining a head on any flip, and there were $R$ of them, and $1-H$ is the corresponding probability for a tail, of which there were $N-R$. We note that this binomial distribution also contains a normalization factor, but we will ignore it since it does not depend explicitly on $H$, the quantity of interest. It will be absorbed by the normalization condition (ref{eq:coin_posterior_norm}). + +We perform the setup of this Bayesian framework on the computer. + +@@@CODE src/coinflipping.py from-to: start bayesian setup@end bayesian setup + +The next step is to confront this setup with the simulated data. To get a feel for the result, it is instructive to see how the posterior pdf evolves as we obtain more and more data pertaining to the coin. The results of such an analyses is shown in Fig. ref{fig:coinflipping}. + +@@@CODE src/coinflipping.py from-to: start plotting@end plotting +FIGURE:[fig/coinflipping_fig_1.png, width=500 frac=0.95] The evolution of the posterior pdf for the bias-weighting of a coin, as the number of data available increases. The figure on the top left-hand corner of each panel shows the number of data included in the analysis. label{fig:coinflipping} + +The panel in the top left-hand corner shows the posterior pdf for $H$ given no data, i.e., it is the same as the prior pdf of Eq. (ref{eq:coin_prior_uniform}). It indicates that we have no more reason to believe that the coin is fair than we have to think that it is double-headed, double-tailed, or of any other intermediate bias-weighting. + +The first flip is obviously tails. At this point we have no evidence that the coin has a side with heads, as indicated by the pdf going to zero as $H \to 1$. The second flip is obviously heads and we have now excluded both extreme options $H=0$ (double-tailed) and $H=1$ (double-headed). We can note that the posterior at this point has the simple form $p(H|D,I) = H(1-H)$ for $0 \le H \le 1$. + +The remainder of Fig. ref{fig:coinflipping} shows how the posterior pdf evolves as the number of data analysed becomes larger and larger. We see that the position of the maximum moves around, but that the amount by which it does so decreases with the increasing number of observations. The width of the posterior pdf also becomes narrower with more data, indicating that we are becoming increasingly confident in our estimate of the bias-weighting. For the coin in this example, the best estimate of $H$ eventually converges to 0.6, which, of course, was the value chosen to simulate the flips. + + +===== A few words on different priors ===== +* uniform +* Gaussian +* Jeffrey's prior +Repeat the coin flipping experiment with other priors. + + +===== Bayesian parameter estimation (single parameter) ===== +!bblock +We will now consider the very important task of model parameter estimation using statistical inference. +[CF: maybe stress that model parameters are not random variables, and the meaning of parameter estimation is therefore very different between frequentist and bayesian approaches.] + +Throughout this section we will consider a specific example that involves a model with a single parameter: ``Measured flux from a star''. +!eblock + + +=== Example: Measured flux from a star === +Adapted from the blog "Pythonic Perambulations": "http://jakevdp.github.io" by Jake VanderPlas. + +Imagine that we point our telescope to the sky, and observe the light coming from a single star. For the time being, we'll assume that the star's true flux is constant with time, i.e. that is it has a fixed value $F_\mathrm{true}$ (we'll also ignore effects like sky noise and other sources of systematic error). We'll assume that we perform a series of $N$ measurements with our telescope, where the ith measurement reports the observed photon flux $F_i$ and error $e_i$[^errors]. +The question is, given this set of measurements $D = \{F_i, e_i\}$, what is our best estimate of the true flux $F_\mathrm{true}$? + +[^errors]: We'll make the reasonable assumption that errors are Gaussian. In a Frequentist perspective, $e_i$ is the standard deviation of the results of a single measurement event in the limit of repetitions of *that event*. In the Bayesian perspective, $e_i$ is the standard deviation of the (Gaussian) probability distribution describing our knowledge of that particular measurement given its observed value. + +Because the measurements are number counts, a Poisson distribution is a good approximation to the measurement process: + +@@@CODE src/singlephotoncount.py from-to: start generate data@end generate data + +Now let's make a simple visualization of the ``observed'' data, see Fig. ref{fig:flux}. + +@@@CODE src/singlephotoncount.py from-to: start visualize data@end visualize data + +FIGURE:[fig/singlephotoncount_fig_1.png, width=400 frac=0.8] Single photon counts (flux measurements). label{fig:flux} + +These measurements each have a different error $e_i$ which is estimated from Poisson statistics using the standard square-root rule. In this toy example we already know the true flux $F_\mathrm{true}$, but the question is this: given our measurements and errors, what is our best estimate of the true flux? + +Let's take a look at the frequentist and Bayesian approaches to solving this. + +=== Simple Photon Counts: Frequentist Approach === +We'll start with the classical frequentist maximum likelihood approach. Given a single observation $D_i = (F_i, e_i)$, we can compute the probability distribution of the measurement given the true flux Ftrue given our assumption of Gaussian errors +!bt +\begin{equation} +p(D_i | F_\mathrm{true}, I) = \frac{1}{\sqrt{2\pi e_i^2}} \exp \left( \frac{-(F_i-F_\mathrm{true})^2}{2e_i^2} \right). +\end{equation} +!et +This should be read ``the probability of $D_i$ given $F_\mathrm{true}$ +equals ...''. You should recognize this as a normal distribution with mean $F_\mathrm{true}$ and standard deviation $e_i$. + +We construct the *likelihood function* by computing the product of the probabilities for each data point +!bt +\begin{equation} +\mathcal{L}(D | F_\mathrm{true}, I) = \prod_{i=1}^N p(D_i | F_\mathrm{true}, I), +\end{equation} +!et +here $D = \{D_i\}$ represents the entire set of measurements. Because the value of the likelihood can become very small, it is often more convenient to instead compute the log-likelihood. Combining the previous two equations and computing the log, we have +!bt +\begin{equation} +\log\mathcal{L} = -\frac{1}{2} \sum_{i=1}^N \left[ \log(2\pi e_i^2) + \frac{(F_i-F_\mathrm{true})^2}{e_i^2} \right]. +\end{equation} +!et + +What we'd like to do is determine $F_\mathrm{true}$ such that the likelihood is maximized. For this simple problem, the maximization can be computed analytically (i.e. by setting $d\log\mathcal{L}/d F_\mathrm{true} = 0$). This results in the following observed estimate of $F_\mathrm{true}$ +!bt +\begin{equation} +F_\mathrm{est} = \sum_{i=1}^N w_i F_i; \quad w_i = 1/e_i^2. +\end{equation} +!et +Notice that in the special case of all errors $e_i$ being equal, this reduces to +!bt +\begin{equation} +F_\mathrm{est} = \frac{1}{N} \sum_{i=1} F_i. +\end{equation} +!et +That is, in agreement with intuition, $F_\mathrm{est}$ is simply the mean of the observed data when errors are equal. + +We can go further and ask what the error of our estimate is. In the frequentist approach, this can be accomplished by fitting a Gaussian approximation to the likelihood curve at maximum; in this simple case this can also be solved analytically (the sum of Gaussians is also a Gaussian). It can be shown that the standard deviation of this Gaussian approximation is +!bt +\begin{equation} +\sigma_\mathrm{est} = \sum_{i=1}^N w_i. +\end{equation} +!et +These results are fairly simple calculations; let's evaluate them for our toy dataset: + +@@@CODE src/singlephotoncount.py from-to: start frequentist@end frequentist +`F_true = 1000` +`F_est = 998 +/- 4 (based on 50 measurements)` + +We find that for 50 measurements of the flux, our estimate has an error of about 0.4% and is consistent with the input value. + + +=== Simple Photon Counts: Bayesian Approach === +The Bayesian approach, as you might expect, begins and ends with probabilities. Our hypothesis is that the star has a constant flux $F_\mathrm{true}$. It recognizes that what we fundamentally want to compute is our knowledge of the parameters in question given the data and other information (such as our knowledge of uncertainties for the observed values), i.e. in this case, $p(F_\mathrm{true} | D,I)$. +Note that this formulation of the problem is fundamentally contrary to the frequentist philosophy, which says that probabilities have no meaning for model parameters like $F_\mathrm{true}$. Nevertheless, within the Bayesian philosophy this is perfectly acceptable. + +To compute this result, Bayesians next apply Bayes' Theorem (ref{eq:bayes}). +If we set the prior $p(F_\mathrm{true}|I) \propto 1$ (a flat prior), we find +$p(F_\mathrm{true}|D,I) \propto p(D | F_\mathrm{true},I) \equiv \mathcal{L}(D | F_\mathrm{true},I)$ +and the Bayesian probability is maximized at precisely the same value as the frequentist result! So despite the philosophical differences, we see that (for this simple problem at least) the Bayesian and frequentist point estimates are equivalent. + +=== A note about priors === +The prior allows inclusion of other information into the computation, which becomes very useful in cases where multiple measurement strategies are being combined to constrain a single model. The necessity to specify a prior, however, is one of the more controversial pieces of Bayesian analysis. +A frequentist will point out that the prior is problematic when no true prior information is available. Though it might seem straightforward to use a noninformative prior like the flat prior mentioned above, there are some "surprisingly subtleties": "http://normaldeviate.wordpress.com/2013/07/13/lost-causes-in-statistics-ii-noninformative- priors/comment-page-1/" involved. It turns out that in many situations, a truly noninformative prior does not exist! Frequentists point out that the subjective choice of a prior which necessarily biases your result has no place in statistical data analysis. +A Bayesian would counter that frequentism doesn't solve this problem, but simply skirts the question. Frequentism can often be viewed as simply a special case of the Bayesian approach for some (implicit) choice of the prior: a Bayesian would say that it's better to make this implicit choice explicit, even if the choice might include some subjectivity. + +=== Simple Photon Counts: Bayesian approach in practice === +Leaving these philosophical debates aside for the time being, let's address how Bayesian results are generally computed in practice. For a one parameter problem like the one considered here, it's as simple as computing the posterior probability $p(F_\mathrm{true} | D,I)$ as a function of $F_\mathrm{true}$: this is the distribution reflecting our knowledge of the parameter $F_\mathrm{true}$. +But as the dimension of the model grows, this direct approach becomes increasingly intractable. For this reason, Bayesian calculations often depend on sampling methods such as Markov Chain Monte Carlo (MCMC). For this practical example, let us apply an MCMC approach using Dan Foreman-Mackey's "emcee": "http://dan.iel.fm/emcee/current/" package. Keep in mind here that the goal is to generate a set of points drawn from the posterior probability distribution, and to use those points to determine the answer we seek. +To perform this MCMC, we start by defining Python functions for the prior $p(F_\mathrm{true} | I)$, the likelihood $p(D | F_\mathrm{true},I)$, and the posterior $p(F_\mathrm{true} | D,I)$, noting that none of these need be properly normalized. Our model here is one-dimensional, but to handle multi-dimensional models we'll define the model in terms of an array of parameters $\boldsymbol{\alpha}$, which in this case is $\boldsymbol{\alpha} = [F_\mathrm{true}]$ + +@@@CODE src/singlephotoncount.py from-to: start bayesian setup@end bayesian setup + +Now we set up the problem, including generating some random starting guesses for the multiple chains of points. + +@@@CODE src/singlephotoncount.py from-to: start bayesian mcmc@end bayesian mcmc + +If this all worked correctly, the array sample should contain a series of 50,000 points drawn from the posterior. Let's plot them and check. See results in Fig. ref{fig:flux-bayesian}. + +@@@CODE src/singlephotoncount.py from-to: start visualize bayesian@end visualize bayesian +FIGURE:[fig/singlephotoncount_fig_2.png, width=400 frac=0.8] Bayesian posterior pdf (represented by a histogram of MCMC samples) from flux measurements. label{fig:flux-bayesian} + +=== Best estimates and confidence intervals === +The posterior distribution from our Bayesian data analysis is the key quantity that encodes our inference about the values of the model parameters, given the data and the relevant background information. Often, however, we wish to summarize this result with just a few numbers: the best estimate and a measure of its reliability. + +There are a few different options for this. The choice of the most appropriate one depends mainly on the shape of the posterior distribution: + +*Symmetric posterior pdfs*: Since the probability (density) associated with any particular value of the parameter is a measure of how much we believe that it lies in the neighbourhood of that point, our best estimate is given by the maximum of the posterior pdf. If we denote the quantity of interest by $X$, with a posterior pdf $P =p(X|D,I)$, then the best estimate of its value $X_0$ is given by the condition $dP/dX|_{X=X_0}=0$. Strictly speaking, we should also check the sign of the second derivative to ensure that $X_0$ represents a maximum. + +To obtain a measure of the reliability of this best estimate, we need to look at the width or spread of the posterior pdf about $X_0$. When considering the behaviour of any function in the neighbourhood of a particular point, it is often helpful to carry out a Taylor series expansion; this is simply a standard tool for (locally) approximating a complicated function by a low-order polynomial. The linear term is zero at the maximum and the quadratic term is often the dominating one determining the width of the posterior pdf. Ignoring all the higher-order terms we arrive at the Gaussian approximation +!bt +\begin{equation} +p(X|D,I) \approx \frac{1}{\sigma\sqrt{2\pi}} \exp \left[ -\frac{(x-\mu)^2}{2\sigma^2} \right], +\end{equation} +!et +where the mean $\mu = X_0$ and the variance $\sigma = \left( - \left. \frac{d^2L}{dX^2} \right|_{X_0} \right)^{-1/2}$, where $L$ is the logarithm of the posterior $P$. Our inference about the quantity of interest is conveyed very concisely, therefore, by the statement $X = X_0 \pm \sigma$, and +!bt +$$ +p(X_0-\sigma < X < X_0+\sigma | D,I) = \int_{X_0-\sigma}^{X_0+\sigma} p(X|D,I) dX \approx 0.67. +$$ +!et + +*Asymmetric posterior pdfs*: While the maximum of the posterior ($X_0$) can still be regarded as giving the best estimate, the true value is now more likely to be on one side of this rather than the other. Alternatively one can compute the mean value, $\langle X \rangle = \int X p(X|D,I) dX$, although this tends to overemphasise very long tails. The best option is probably a compromise that can be employed when having access to a large sample from the posterior (as provided by an MCMC), namely to give the median of this ensamble. + +Furthermore, the concept of an error-bar does not seem appropriate in this case, as it implicitly entails the idea of symmetry. A good way of expressing the reliability with which a parameter can be inferred, for an asymmetric posterior pdf, is rather through a *confidence interval*. Since the area under the posterior pdf between $X_1$ and $X_2$ is proportional to how much we believe that $X$ lies in that range, the shortest interval that encloses 67% of the area represents a sensible measure of the uncertainty of the estimate. Obviously we can choose to provide some other degree-of-belief that we think is relevant for the case at hand. Assuming that the posterior pdf has been normalized, to have unit area, we need to find $X_1$ and $X_2$ such that: +!bt +$$ +p(X_1 < X < X_2 | D,I) = \int_{X_1}^{X_2} p(X|D,I) dX \approx 0.67, +$$ +!et +where the difference $X_2 - X_1$ is as small as possible. The region $X_1 < X < X_2$ is then called the shortest 67% confidence interval. + +*Multimodal posterior pdfs*: We can sometimes obtain posteriors which are multimodal; i.e. contains several disconnected regions with large probabilities. There is no difficulty when one of the maxima is very much larger than the others: we can simply ignore the subsidiary solutions, to a good approximation, and concentrate on the global maximum. The problem arises when there are several maxima of comparable magnitude. What do we now mean by a best estimate, and how should we quantify its reliability? The idea of a best estimate and an error-bar, or even a confidence interval, is merely an attempt to summarize the posterior with just two or three numbers; sometimes this just can’t be done, and so these concepts are not valid. For the bimodal case we might be able to characterize the posterior in terms of a few numbers: two best estimates and their associated error-bars, or disjoint confidence intervals. For a general multimodal pdf, the most honest thing we can do is just display the posterior itself. + +=== Simple Photon Counts: Best estimates and confidence intervals === +To compute these numbers for our example, you would run: + +@@@CODE src/singlephotoncount.py from-to: start bayesian CI@end bayesian CI +`F_true = 1000` +`Based on 50 measurements the posterior point estimates are:` +`...F_est = 998 +/- 4` +`or using credible intervals:` +`...F_est = 998 (posterior median)` +`...F_est in [993, 1002] (67% credible interval)` +`...F_est in [989, 1006] (95% credible interval)` + +In this particular example, the posterior pdf is actually a Gaussian (since it is constructed as a product of Gaussians), and the mean and variance from the quadratic approximation will agree exactly with the frequentist approach. + +From this final result you might come away with the impression that the Bayesian method is unnecessarily complicated, and in this case it certainly is. Using an MCMC sampler to characterize a one-dimensional normal distribution is a bit like using the Death Star to destroy a beach ball, but we did this here because it demonstrates an approach that can scale to complicated posteriors in many, many dimensions, and can provide nice results in more complicated situations where an analytic likelihood approach is not possible. + +Furthermore, as data and models grow in complexity, the two approaches can diverge greatly. + + +===== Bayesian parameter estimation (multiple parameters, covariance) ===== +!bblock +* multidimensional posterior pdf:s +* nuisance parameters (e.g. background subtraction?) +* corner plots, covariance, correlations +* best example? +!eblock + + + +===== Bayesian model selection ===== +!bblock +* Bayesian evidence +* Occam's razor +* Best example? How many spectral lines are there? +!eblock + + diff --git a/doc/src/Bayesian/chapter11.do.txt~ b/doc/src/Bayesian/chapter11.do.txt~ new file mode 100644 index 000000000..a23db26de --- /dev/null +++ b/doc/src/Bayesian/chapter11.do.txt~ @@ -0,0 +1,449 @@ +TITLE: Data Analysis and Machine Learning: Elements of Bayesian theory and Bayesian Neural Networks +AUTHOR: Christian Forssén at Department of Physics, Chalmers University of Technology, Sweden +AUTHOR: Morten Hjorth-Jensen {copyright, 1999-present|CC BY-NC} at Department of Physics, University of Oslo & Department of Physics and Astronomy and National Superconducting Cyclotron Laboratory, Michigan State University +DATE: today + + +!split +===== To do ===== + +* Develop connection between Bayesian statistics and linear regression +* add bayesian basic stat to the intro to statistics slides and add more math about bayesian approah and frequentist approach +* add examples on bayesian statistics and linear regression +* Include also bayesian logistic regression +* Think of Bayesian NN +!split +===== Why Bayesian Statistics? ===== +!bblock + +We have already made ourselves familiar with elements of a statistical +data analysis via quantities like the bias-variance tradeoff as well +as some central distribution functions such as the Normal +distribution, the binomial distribution and other probability +distribution functions. + +In essentially all the Machine Learning algorithms we have studied, +our focus has been on a so-called _frequentist approach_, where +knowledge of an underlying likelihood function has not been +emphasized. Our data, whether we had a classification or a regression +problem, have been our central points of departure. + +Here we wish to merge this approach with the derivation of a +likelihood function which can be used to make prediction on how our +system under study evolves. We will venture into the realm of what is +called Bayesian Neural Networks. To get an overarching view on what +this entails, the following figure conveys the essential differences +between a standard Neural network that we have met earlier and a +Bayesian Neural Network. In order to get there, we need to present +some of the basic elements of Bayesian statistics, starting with the +product rule and Bayes' theorem. + +!eblock + +!split +===== Inference ===== +!bblock +- Inference: ``the act of passing from one proposition, statement or judgment considered as true to another whose truth is believed to follow from that of the former'' (Webster) + Do premises $A, B, \ldots \to$ hypothesis, $H$? + +- Deductive inference: Premises allow definite determination of truth/falsity of H (syllogisms, symbolic logic, Boolean algebra) + $B(H|A,B,...) = 0$ or $1$ + +- Inductive inference: Premises bear on truth/falsity of H, but don’t allow its definite determination (weak syllogisms, analogies) + $A, B, C, D$ share properties $x, y, z$; $E$ has properties $x, y$ + $\to$ $E$ probably has property $z$. +!eblock + +!split +===== Statistical Inference ===== +!bblock +* Quantify the strength of inductive inferences from facts, in the form of data ($D$), and other premises, e.g. models, to hypotheses about the phenomena producing the data. +* Quantify via probabilities, or averages calculated using probabilities. Frequentists ($\mathcal{F}$) and Bayesians ($\mathcal{B}$) use probabilities very differently for this. +* To the pioneers such as Bernoulli, Bayes and Laplace, a probability represented a *degree-of-belief* or plausability: how much they thought that something as true based on the evidence at hand. This is the Bayesian approach. +* To the 19th century scholars, this seemed too vague and subjective. They redefined probability as the *long run relative frequency* with which an event occurred, given (infinitely) many repeated (experimental) trials. +!eblock + +!split +===== Some history ===== +Adapted from D.S. Sivia[^Sivia]: + +[^Sivia]: Sivia, Devinderjit, and John Skilling. Data Analysis : A Bayesian Tutorial, OUP Oxford, 2006 + +!bquote +Although the frequency definition appears to be more objective, its range of validity is also far more limited. For example, Laplace used (his) probability theory to estimate the mass of Saturn, given orbital data that were available to him from various astronomical observatories. In essence, he computed the posterior pdf for the mass M , given the data and all the relevant background information I (such as a knowledge of the laws of classical mechanics): prob(M|{data},I); this is shown schematically in the figure [Fig. 1.2]. +!equote + +!split +FIGURE: [fig/sivia_fig_1_2.png, width=700 frac=0.9] + +!split +!bquote +To Laplace, the (shaded) area under the posterior pdf curve between $m_1$ and $m_2$ was a measure of how much he believed that the mass of Saturn lay in the range $m_1 \le M \le m_2$. As such, the position of the maximum of the posterior pdf represents a best estimate of the mass; its width, or spread, about this optimal value gives an indication of the uncertainty in the estimate. Laplace stated that: ‘ . . . it is a bet of 11,000 to 1 that the error of this result is not 1/100th of its value.’ He would have won the bet, as another 150 years’ accumulation of data has changed the estimate by only 0.63%! +!equote + +!split +!bquote +According to the frequency definition, however, we are not permitted to use probability theory to tackle this problem. This is because the mass of Saturn is a constant and not a random variable; therefore, it has no frequency distribution and so probability theory cannot be used. + +If the pdf [of Fig. 1.2] had to be interpreted in terms of the frequency definition, we would have to imagine a large ensemble of universes in which everything remains constant apart from the mass of Saturn. +!equote + +!split +!bquote +As this scenario appears quite far-fetched, we might be inclined to think of [Fig. 1.2] in terms of the distribution of the measurements of the mass in many repetitions of the experiment. Although we are at liberty to think about a problem in any way that facilitates its solution, or our understanding of it, having to seek a frequency interpretation for every data analysis problem seems rather perverse. +For example, what do we mean by the ‘measurement of the mass’ when the data consist of orbital periods? Besides, why should we have to think about many repetitions of an experiment that never happened? What we really want to do is to make the best inference of the mass given the (few) data that we actually have; this is precisely the Bayes and Laplace view of probability. +!equote + +!split +!bquote +Faced with the realization that the frequency definition of probability theory did not permit most real-life scientific problems to be addressed, a new subject was invented — statistics! To estimate the mass of Saturn, for example, one has to relate the mass to the data through some function called the statistic; since the data are subject to ‘random’ noise, the statistic becomes the random variable to which the rules of probability the- ory can be applied. But now the question arises: How should we choose the statistic? The frequentist approach does not yield a natural way of doing this and has, therefore, led to the development of several alternative schools of orthodox or conventional statis- tics. The masters, such as Fisher, Neyman and Pearson, provided a variety of different principles, which has merely resulted in a plethora of tests and procedures without any clear underlying rationale. This lack of unifying principles is, perhaps, at the heart of the shortcomings of the cook-book approach to statistics that students are often taught even today. +!equote + +!split +===== The Bayesian recipe ===== +!bblock +Assess hypotheses by calculating their probabilities $p(H_i | \ldots)$ conditional on known and/or presumed information using the rules of probability theory. +!eblock +!bblock +Probability Theory Axioms: +- Product (AND) rule : $p(A, B | I) = p(A|I) p(B|A, I) = p(B|I)p(A|B,I)$ + Should read $p(A,B|I)$ as the probability for propositions $A$ AND $B$ being true given that $I$ is true. +- Sum (OR) rule: $p(A + B | I) = p(A | I) + p(B | I) - p(A, B | I)$ + $p(A+B|I)$ is the probability that proposition $A$ OR $B$ is true given that $I$ is true. +- Normalization: $p(A|I) + p(\bar{A}|I) = 1$ + $\bar{A}$ denotes the proposition that $A$ is false. + +!eblock + +!split +===== Bayes' theorem ===== +!bblock +Bayes' theorem follows directly from the product rule +!bt +$$ +p(A|B,I) = \frac{p(B|A,I) p(A|I)}{p(B|I)}. +$$ +!et +The importance of this property to data analysis becomes apparent if we replace $A$ and $B$ by hypothesis($H$) and data($D$): +!bt +\begin{align} +p(H|D,I) &= \frac{p(D|H,I) p(H|I)}{p(D|I)}. +label{eq:bayes} +\end{align} +!et +The power of Bayes’ theorem lies in the fact that it relates the quantity of interest, the probability that the hypothesis is true given the data, to the term we have a better chance of being able to assign, the probability that we would have observed the measured data if the hypothesis was true. +!eblock + +!split +!bblock +The various terms in Bayes’ theorem have formal names. +* The quantity on the far right, $p(H|I)$, is called the *prior* probability; it represents our state of knowledge (or ignorance) about the truth of the hypothesis before we have analysed the current data. +* This is modified by the experimental measurements through $p(D|H,I)$, the *likelihood* function, +* The denominator $p(D|I)$ is called the *evidence*. It does not depend on the hypothesis and can be regarded as a normalization constant. +* Together, these yield the *posterior* probability, $p(H|D, I )$, representing our state of knowledge about the truth of the hypothesis in the light of the data. + +In a sense, Bayes’ theorem encapsulates the process of learning. +!eblock + +!split +===== The friends of Bayes' theorem ===== +!bblock +- Normalization: $\sum_i p(H_i|\ldots) = 1$. +- Marginalization: $\sum_i p(A,H_i|I) = \sum_i p(H_i|A,I) p(A|I) = p(A|I)$. +- Marginalization (continuum limit): $\int dx p(A,H(x)|I) = p(A|I)$. + +In the above, $H_i$ is an exclusive and exhaustive list of hypotheses. For example,let’s imagine that there are five candidates in a presidential election; then $H_1$ could be the proposition that the first candidate will win, and so on. The probability that $A$ is true, for example that unemployment will be lower in a year’s time (given all relevant information $I$, but irrespective of whoever becomes president) is then given by $\sum_i p(A,H_i|I)$. + +In the continuum limit of propositions we must understand $p(\ldots)$ as a pdf (probability density function). + +Marginalization is a very powerful device in data analysis because it enables us to deal with nuisance parameters; that is, quantities which necessarily enter the analysis but are of no intrinsic interest. The unwanted background signal present in many experimental measurements are examples of nuisance parameters. +!eblock + +!split +===== Inference With Parametric Models ===== +!bblock +Inductive inference with parametric models is a very important tool in the natural sciences. +* Consider $N$ different models $M_i$ ($i = 1, \ldots, N$), each with parameters $\boldsymbol{\alpha}_i$. Each of them implies a sampling distribution (conditional predictive distribution for possible data) +!bt +$$ +p(D|\boldsymbol{\alpha}_i, M_i) +$$ +!et +* The $\boldsymbol{\alpha}_i$ dependence when we fix attention on the actual, observed data ($D_\mathrm{obs}$) is the likelihood function: +!bt +$$ +\mathcal{L}_i (\boldsymbol{\alpha}_i) \equiv p(D_\mathrm{obs}|\boldsymbol{\alpha}_i, M_i) +$$ +!et +* We may be uncertain about $i$ (model uncertainty), +* or uncertain about $\boldsymbol{\alpha}_i$ (parameter uncertainty). +!eblock + +!split +!bblock +- Parameter Estimation: Premise = choice of model (pick specific $i$) + $\Rightarrow$ What can we say about $\boldsymbol{\alpha}_i$? +- Model comparison: Premise = $\{M_i\}$ + $\Rightarrow$ What can we say about $i$? +- Model adequacy: Premise = $M_1$ + $\Rightarrow$ Is $M_1$ adequate? +- Hybrid Uncertainty: Models share some common params: $\boldsymbol{\alpha}_1 = \{ \boldsymbol{\varphi}, \boldsymbol{\eta}_i\}$ + $\Rightarrow$ What can we say about $\boldsymbol{\varphi}$? (Systematic error is an example) +!eblock + +!split +===== Illustrative examples with python code ===== +!bblock +* Is this a fair coin? (analytical) +* Flux from a star (single parameter, MCMC) +* The lighthouse problem (two parameters, MCMC) +* Linear fit with outliers (nuisance parameters) +* ... +!eblock + +!split +===== Example: Is this a fair coin? ===== +Let us begin with the analysis of data from a simple coin-tossing experiment. +Given that we had observed 6 heads in 8 flips, would you think it was a fair coin? By fair, we mean that we would be prepared to lay an even 1 : 1 bet on the outcome of a flip being a head or a tail. If we decide that the coin was fair, the question which follows naturally is how sure are we that this was so; if it was not fair, how unfair do we think it was? Furthermore, if we were to continue collecting data for this particular coin, observing the outcomes of additional flips, how would we update our belief on the fairness of the coin? + +A sensible way of formulating this problem is to consider a large number of hypotheses about the range in which the bias-weighting of the coin might lie. If we denote the bias-weighting by $H$, then $H = 0$ and $H = 1$ can represent a coin which produces a tail or a head on every flip, respectively. There is a continuum of possibilities for the value of H between these limits, with $H = 0.5$ indicating a fair coin. Our state of knowledge about the fairness, or the degree of unfairness, of the coin is then completely summarized by specifying how much we believe these various propositions to be true. + +Let us perform a computer simulation of a coin-tossing experiment. This provides the data that we will be analysing. + +@@@CODE src/coinflipping.py from-to: start import modules@end import modules +@@@CODE src/coinflipping.py from-to: start simulation@end simulation + +In the light of this data, our inference about the fairness of this coin is summarized by the conditional pdf: $p(H|D,I)$. This is, of course, shorthand for the limiting case of a continuum of propositions for the value of $H$; that is to say, the probability that $H$ lies in an infinitesimally narrow range is given by $p(H|D,I) dH$. + +To estimate this posterior pdf, we need to use Bayes’ theorem (ref{eq:bayes}). We will ignore the denominator $p(D|I)$ as it does not involve bias-weighting explicitly, and it will therefore not affect the shape of the desired pdf. At the end we can evaluate the missing constant subsequently from the normalization condition +!bt +\begin{equation} +\int_0^1 p(H|D,I) dH = 1. +label{eq:coin_posterior_norm} +\end{equation} +!et + +The prior pdf, $p(H|I)$, represents what we know about the coin given only the information $I$ that we are dealing with a ‘strange coin’. We could keep a very open mind about the nature of the coin; a simple probability assignment which reflects this is a uniform, or flat, prior +!bt +\begin{equation} +p(H|I) = \left\{ \begin{array}{ll} +1 & 0 \le H \le 1, \\ +0 & \mathrm{otherwise}. +\end{array} \right. +label{eq:coin_prior_uniform} +\end{equation} +!et +We will get back later to the choice of prior and its effect on the analysis. + +This prior state of knowledge, or ignorance, is modified by the data through the likelihood function $p(D|H,I)$. It is a measure of the chance that we would have obtained the data that we actually observed, if the value of the bias-weighting was given (as known). If, in the conditioning information $I$, we assume that the flips of the coin were independent events, so that the outcome of one did not influence that of another, then the probability of obtaining the data `R heads in N tosses' is given by the binomial distribution (we leave a formal definition of this to a statistics textbook) +!bt +\begin{equation} +p(D|H,I) \propto H^R (1-H)^{N-R}. +\end{equation} +!et +It seems reasonable because $H$ is the chance of obtaining a head on any flip, and there were $R$ of them, and $1-H$ is the corresponding probability for a tail, of which there were $N-R$. We note that this binomial distribution also contains a normalization factor, but we will ignore it since it does not depend explicitly on $H$, the quantity of interest. It will be absorbed by the normalization condition (ref{eq:coin_posterior_norm}). + +We perform the setup of this Bayesian framework on the computer. + +@@@CODE src/coinflipping.py from-to: start bayesian setup@end bayesian setup + +The next step is to confront this setup with the simulated data. To get a feel for the result, it is instructive to see how the posterior pdf evolves as we obtain more and more data pertaining to the coin. The results of such an analyses is shown in Fig. ref{fig:coinflipping}. + +@@@CODE src/coinflipping.py from-to: start plotting@end plotting +FIGURE:[fig/coinflipping_fig_1.png, width=500 frac=0.95] The evolution of the posterior pdf for the bias-weighting of a coin, as the number of data available increases. The figure on the top left-hand corner of each panel shows the number of data included in the analysis. label{fig:coinflipping} + +The panel in the top left-hand corner shows the posterior pdf for $H$ given no data, i.e., it is the same as the prior pdf of Eq. (ref{eq:coin_prior_uniform}). It indicates that we have no more reason to believe that the coin is fair than we have to think that it is double-headed, double-tailed, or of any other intermediate bias-weighting. + +The first flip is obviously tails. At this point we have no evidence that the coin has a side with heads, as indicated by the pdf going to zero as $H \to 1$. The second flip is obviously heads and we have now excluded both extreme options $H=0$ (double-tailed) and $H=1$ (double-headed). We can note that the posterior at this point has the simple form $p(H|D,I) = H(1-H)$ for $0 \le H \le 1$. + +The remainder of Fig. ref{fig:coinflipping} shows how the posterior pdf evolves as the number of data analysed becomes larger and larger. We see that the position of the maximum moves around, but that the amount by which it does so decreases with the increasing number of observations. The width of the posterior pdf also becomes narrower with more data, indicating that we are becoming increasingly confident in our estimate of the bias-weighting. For the coin in this example, the best estimate of $H$ eventually converges to 0.6, which, of course, was the value chosen to simulate the flips. + +!split +===== A few words on different priors ===== +* uniform +* Gaussian +* Jeffrey's prior +Repeat the coin flipping experiment with other priors. + +!split +===== Bayesian parameter estimation (single parameter) ===== +!bblock +We will now consider the very important task of model parameter estimation using statistical inference. +[CF: maybe stress that model parameters are not random variables, and the meaning of parameter estimation is therefore very different between frequentist and bayesian approaches.] + +Throughout this section we will consider a specific example that involves a model with a single parameter: ``Measured flux from a star''. +!eblock + +!split +=== Example: Measured flux from a star === +Adapted from the blog "Pythonic Perambulations": "http://jakevdp.github.io" by Jake VanderPlas. + +Imagine that we point our telescope to the sky, and observe the light coming from a single star. For the time being, we'll assume that the star's true flux is constant with time, i.e. that is it has a fixed value $F_\mathrm{true}$ (we'll also ignore effects like sky noise and other sources of systematic error). We'll assume that we perform a series of $N$ measurements with our telescope, where the ith measurement reports the observed photon flux $F_i$ and error $e_i$[^errors]. +The question is, given this set of measurements $D = \{F_i, e_i\}$, what is our best estimate of the true flux $F_\mathrm{true}$? + +[^errors]: We'll make the reasonable assumption that errors are Gaussian. In a Frequentist perspective, $e_i$ is the standard deviation of the results of a single measurement event in the limit of repetitions of *that event*. In the Bayesian perspective, $e_i$ is the standard deviation of the (Gaussian) probability distribution describing our knowledge of that particular measurement given its observed value. + +Because the measurements are number counts, a Poisson distribution is a good approximation to the measurement process: + +@@@CODE src/singlephotoncount.py from-to: start generate data@end generate data + +Now let's make a simple visualization of the ``observed'' data, see Fig. ref{fig:flux}. + +@@@CODE src/singlephotoncount.py from-to: start visualize data@end visualize data + +FIGURE:[fig/singlephotoncount_fig_1.png, width=400 frac=0.8] Single photon counts (flux measurements). label{fig:flux} + +These measurements each have a different error $e_i$ which is estimated from Poisson statistics using the standard square-root rule. In this toy example we already know the true flux $F_\mathrm{true}$, but the question is this: given our measurements and errors, what is our best estimate of the true flux? + +Let's take a look at the frequentist and Bayesian approaches to solving this. + +=== Simple Photon Counts: Frequentist Approach === +We'll start with the classical frequentist maximum likelihood approach. Given a single observation $D_i = (F_i, e_i)$, we can compute the probability distribution of the measurement given the true flux Ftrue given our assumption of Gaussian errors +!bt +\begin{equation} +p(D_i | F_\mathrm{true}, I) = \frac{1}{\sqrt{2\pi e_i^2}} \exp \left( \frac{-(F_i-F_\mathrm{true})^2}{2e_i^2} \right). +\end{equation} +!et +This should be read ``the probability of $D_i$ given $F_\mathrm{true}$ +equals ...''. You should recognize this as a normal distribution with mean $F_\mathrm{true}$ and standard deviation $e_i$. + +We construct the *likelihood function* by computing the product of the probabilities for each data point +!bt +\begin{equation} +\mathcal{L}(D | F_\mathrm{true}, I) = \prod_{i=1}^N p(D_i | F_\mathrm{true}, I), +\end{equation} +!et +here $D = \{D_i\}$ represents the entire set of measurements. Because the value of the likelihood can become very small, it is often more convenient to instead compute the log-likelihood. Combining the previous two equations and computing the log, we have +!bt +\begin{equation} +\log\mathcal{L} = -\frac{1}{2} \sum_{i=1}^N \left[ \log(2\pi e_i^2) + \frac{(F_i-F_\mathrm{true})^2}{e_i^2} \right]. +\end{equation} +!et + +What we'd like to do is determine $F_\mathrm{true}$ such that the likelihood is maximized. For this simple problem, the maximization can be computed analytically (i.e. by setting $d\log\mathcal{L}/d F_\mathrm{true} = 0$). This results in the following observed estimate of $F_\mathrm{true}$ +!bt +\begin{equation} +F_\mathrm{est} = \sum_{i=1}^N w_i F_i; \quad w_i = 1/e_i^2. +\end{equation} +!et +Notice that in the special case of all errors $e_i$ being equal, this reduces to +!bt +\begin{equation} +F_\mathrm{est} = \frac{1}{N} \sum_{i=1} F_i. +\end{equation} +!et +That is, in agreement with intuition, $F_\mathrm{est}$ is simply the mean of the observed data when errors are equal. + +We can go further and ask what the error of our estimate is. In the frequentist approach, this can be accomplished by fitting a Gaussian approximation to the likelihood curve at maximum; in this simple case this can also be solved analytically (the sum of Gaussians is also a Gaussian). It can be shown that the standard deviation of this Gaussian approximation is +!bt +\begin{equation} +\sigma_\mathrm{est} = \sum_{i=1}^N w_i. +\end{equation} +!et +These results are fairly simple calculations; let's evaluate them for our toy dataset: + +@@@CODE src/singlephotoncount.py from-to: start frequentist@end frequentist +`F_true = 1000` +`F_est = 998 +/- 4 (based on 50 measurements)` + +We find that for 50 measurements of the flux, our estimate has an error of about 0.4% and is consistent with the input value. + + +=== Simple Photon Counts: Bayesian Approach === +The Bayesian approach, as you might expect, begins and ends with probabilities. Our hypothesis is that the star has a constant flux $F_\mathrm{true}$. It recognizes that what we fundamentally want to compute is our knowledge of the parameters in question given the data and other information (such as our knowledge of uncertainties for the observed values), i.e. in this case, $p(F_\mathrm{true} | D,I)$. +Note that this formulation of the problem is fundamentally contrary to the frequentist philosophy, which says that probabilities have no meaning for model parameters like $F_\mathrm{true}$. Nevertheless, within the Bayesian philosophy this is perfectly acceptable. + +To compute this result, Bayesians next apply Bayes' Theorem (ref{eq:bayes}). +If we set the prior $p(F_\mathrm{true}|I) \propto 1$ (a flat prior), we find +$p(F_\mathrm{true}|D,I) \propto p(D | F_\mathrm{true},I) \equiv \mathcal{L}(D | F_\mathrm{true},I)$ +and the Bayesian probability is maximized at precisely the same value as the frequentist result! So despite the philosophical differences, we see that (for this simple problem at least) the Bayesian and frequentist point estimates are equivalent. + +=== A note about priors === +The prior allows inclusion of other information into the computation, which becomes very useful in cases where multiple measurement strategies are being combined to constrain a single model. The necessity to specify a prior, however, is one of the more controversial pieces of Bayesian analysis. +A frequentist will point out that the prior is problematic when no true prior information is available. Though it might seem straightforward to use a noninformative prior like the flat prior mentioned above, there are some "surprisingly subtleties": "http://normaldeviate.wordpress.com/2013/07/13/lost-causes-in-statistics-ii-noninformative- priors/comment-page-1/" involved. It turns out that in many situations, a truly noninformative prior does not exist! Frequentists point out that the subjective choice of a prior which necessarily biases your result has no place in statistical data analysis. +A Bayesian would counter that frequentism doesn't solve this problem, but simply skirts the question. Frequentism can often be viewed as simply a special case of the Bayesian approach for some (implicit) choice of the prior: a Bayesian would say that it's better to make this implicit choice explicit, even if the choice might include some subjectivity. + +=== Simple Photon Counts: Bayesian approach in practice === +Leaving these philosophical debates aside for the time being, let's address how Bayesian results are generally computed in practice. For a one parameter problem like the one considered here, it's as simple as computing the posterior probability $p(F_\mathrm{true} | D,I)$ as a function of $F_\mathrm{true}$: this is the distribution reflecting our knowledge of the parameter $F_\mathrm{true}$. +But as the dimension of the model grows, this direct approach becomes increasingly intractable. For this reason, Bayesian calculations often depend on sampling methods such as Markov Chain Monte Carlo (MCMC). For this practical example, let us apply an MCMC approach using Dan Foreman-Mackey's "emcee": "http://dan.iel.fm/emcee/current/" package. Keep in mind here that the goal is to generate a set of points drawn from the posterior probability distribution, and to use those points to determine the answer we seek. +To perform this MCMC, we start by defining Python functions for the prior $p(F_\mathrm{true} | I)$, the likelihood $p(D | F_\mathrm{true},I)$, and the posterior $p(F_\mathrm{true} | D,I)$, noting that none of these need be properly normalized. Our model here is one-dimensional, but to handle multi-dimensional models we'll define the model in terms of an array of parameters $\boldsymbol{\alpha}$, which in this case is $\boldsymbol{\alpha} = [F_\mathrm{true}]$ + +@@@CODE src/singlephotoncount.py from-to: start bayesian setup@end bayesian setup + +Now we set up the problem, including generating some random starting guesses for the multiple chains of points. + +@@@CODE src/singlephotoncount.py from-to: start bayesian mcmc@end bayesian mcmc + +If this all worked correctly, the array sample should contain a series of 50,000 points drawn from the posterior. Let's plot them and check. See results in Fig. ref{fig:flux-bayesian}. + +@@@CODE src/singlephotoncount.py from-to: start visualize bayesian@end visualize bayesian +FIGURE:[fig/singlephotoncount_fig_2.png, width=400 frac=0.8] Bayesian posterior pdf (represented by a histogram of MCMC samples) from flux measurements. label{fig:flux-bayesian} + +=== Best estimates and confidence intervals === +The posterior distribution from our Bayesian data analysis is the key quantity that encodes our inference about the values of the model parameters, given the data and the relevant background information. Often, however, we wish to summarize this result with just a few numbers: the best estimate and a measure of its reliability. + +There are a few different options for this. The choice of the most appropriate one depends mainly on the shape of the posterior distribution: + +*Symmetric posterior pdfs*: Since the probability (density) associated with any particular value of the parameter is a measure of how much we believe that it lies in the neighbourhood of that point, our best estimate is given by the maximum of the posterior pdf. If we denote the quantity of interest by $X$, with a posterior pdf $P =p(X|D,I)$, then the best estimate of its value $X_0$ is given by the condition $dP/dX|_{X=X_0}=0$. Strictly speaking, we should also check the sign of the second derivative to ensure that $X_0$ represents a maximum. + +To obtain a measure of the reliability of this best estimate, we need to look at the width or spread of the posterior pdf about $X_0$. When considering the behaviour of any function in the neighbourhood of a particular point, it is often helpful to carry out a Taylor series expansion; this is simply a standard tool for (locally) approximating a complicated function by a low-order polynomial. The linear term is zero at the maximum and the quadratic term is often the dominating one determining the width of the posterior pdf. Ignoring all the higher-order terms we arrive at the Gaussian approximation +!bt +\begin{equation} +p(X|D,I) \approx \frac{1}{\sigma\sqrt{2\pi}} \exp \left[ -\frac{(x-\mu)^2}{2\sigma^2} \right], +\end{equation} +!et +where the mean $\mu = X_0$ and the variance $\sigma = \left( - \left. \frac{d^2L}{dX^2} \right|_{X_0} \right)^{-1/2}$, where $L$ is the logarithm of the posterior $P$. Our inference about the quantity of interest is conveyed very concisely, therefore, by the statement $X = X_0 \pm \sigma$, and +!bt +$$ +p(X_0-\sigma < X < X_0+\sigma | D,I) = \int_{X_0-\sigma}^{X_0+\sigma} p(X|D,I) dX \approx 0.67. +$$ +!et + +*Asymmetric posterior pdfs*: While the maximum of the posterior ($X_0$) can still be regarded as giving the best estimate, the true value is now more likely to be on one side of this rather than the other. Alternatively one can compute the mean value, $\langle X \rangle = \int X p(X|D,I) dX$, although this tends to overemphasise very long tails. The best option is probably a compromise that can be employed when having access to a large sample from the posterior (as provided by an MCMC), namely to give the median of this ensamble. + +Furthermore, the concept of an error-bar does not seem appropriate in this case, as it implicitly entails the idea of symmetry. A good way of expressing the reliability with which a parameter can be inferred, for an asymmetric posterior pdf, is rather through a *confidence interval*. Since the area under the posterior pdf between $X_1$ and $X_2$ is proportional to how much we believe that $X$ lies in that range, the shortest interval that encloses 67% of the area represents a sensible measure of the uncertainty of the estimate. Obviously we can choose to provide some other degree-of-belief that we think is relevant for the case at hand. Assuming that the posterior pdf has been normalized, to have unit area, we need to find $X_1$ and $X_2$ such that: +!bt +$$ +p(X_1 < X < X_2 | D,I) = \int_{X_1}^{X_2} p(X|D,I) dX \approx 0.67, +$$ +!et +where the difference $X_2 - X_1$ is as small as possible. The region $X_1 < X < X_2$ is then called the shortest 67% confidence interval. + +*Multimodal posterior pdfs*: We can sometimes obtain posteriors which are multimodal; i.e. contains several disconnected regions with large probabilities. There is no difficulty when one of the maxima is very much larger than the others: we can simply ignore the subsidiary solutions, to a good approximation, and concentrate on the global maximum. The problem arises when there are several maxima of comparable magnitude. What do we now mean by a best estimate, and how should we quantify its reliability? The idea of a best estimate and an error-bar, or even a confidence interval, is merely an attempt to summarize the posterior with just two or three numbers; sometimes this just can’t be done, and so these concepts are not valid. For the bimodal case we might be able to characterize the posterior in terms of a few numbers: two best estimates and their associated error-bars, or disjoint confidence intervals. For a general multimodal pdf, the most honest thing we can do is just display the posterior itself. + +=== Simple Photon Counts: Best estimates and confidence intervals === +To compute these numbers for our example, you would run: + +@@@CODE src/singlephotoncount.py from-to: start bayesian CI@end bayesian CI +`F_true = 1000` +`Based on 50 measurements the posterior point estimates are:` +`...F_est = 998 +/- 4` +`or using credible intervals:` +`...F_est = 998 (posterior median)` +`...F_est in [993, 1002] (67% credible interval)` +`...F_est in [989, 1006] (95% credible interval)` + +In this particular example, the posterior pdf is actually a Gaussian (since it is constructed as a product of Gaussians), and the mean and variance from the quadratic approximation will agree exactly with the frequentist approach. + +From this final result you might come away with the impression that the Bayesian method is unnecessarily complicated, and in this case it certainly is. Using an MCMC sampler to characterize a one-dimensional normal distribution is a bit like using the Death Star to destroy a beach ball, but we did this here because it demonstrates an approach that can scale to complicated posteriors in many, many dimensions, and can provide nice results in more complicated situations where an analytic likelihood approach is not possible. + +Furthermore, as data and models grow in complexity, the two approaches can diverge greatly. + +!split +===== Bayesian parameter estimation (multiple parameters, covariance) ===== +!bblock +* multidimensional posterior pdf:s +* nuisance parameters (e.g. background subtraction?) +* corner plots, covariance, correlations +* best example? +!eblock + +!split +===== Bayesian model selection ===== +!bblock +* Bayesian evidence +* Occam's razor +* Best example? How many spectral lines are there? +!eblock + + diff --git a/doc/src/Bayesian/ipynb-chapter11-src.tar.gz b/doc/src/Bayesian/ipynb-chapter11-src.tar.gz new file mode 100644 index 0000000000000000000000000000000000000000..be049d64f77830163c718a5c4c902ac1da090e14 GIT binary patch literal 85179 zcmV)2K+L}%iwFSJzvEv31MIv9P*hDDE(qfw0|*Q`!;mEBAUWrplc3}zl9S{l3_%bC zL?oynNRSMIAPh;6AWD)b8Ic?$NRaK}|NQ=~t-H5&t8Q)WR_zEi2y^=M>3-iQyxr$G zx;XLqxcIx++;wzux+`#3kk`}Q3FCk5i=UrgR9F~+e!nFua`nzH_@DoSm$2Y1K~aQ& zkf4arEfEnx@Iih75fKq#3{ZRrX_q6XNXJEIGF~9}tMOthD^} zQjGl%u`gf0et0uJ3<1+H;L?z*^9hr8(02?{iH~CDFEk*NI!HV8B~7|+C&muyLM?5laA;px<{>5pP02!79|C$GG z4IMtuvafBP++=z7OzQY}deDCA!?EnAhw@*YRt_!bDU4Z5Ip?n{A`lQmlRNWBFuyX-;wPWlvj*Nw$A_;R(g;u~0F2Yr%TF zGxv0*|Ec7nLN&=hZO$7(lF67gpYj@iOS0`MVX!(%ylNH20i~7HGAvXhntbPxZ{MG( zz1F*)p2Bq||DVH@Ajx$}IZ_V{aEL6cgq%M2ZqGL725wFK?@p?kkm}H%`M{EvU@UNc zp5~uFRFQ%|5Ov<|LTC%vKAz|LXDrN8B9=l__zlH1Gr7GNI+T6e(a^j9j5EptO1mQDeRgzTJzwU-%!lAxxGOB`Y3yyOqa}u_lOM_c z4UProdG`L^Y>|5*xJ2{H&L?AKrg164_ip~}*{(WDe?(Kz=`%O8YO)bscX>9i)%)H! zhA?&x!y=;%{{Gm9f9DK+QW34C>X)G0i!G+>hYH&(io zK;Yh#m{IvRUVK5k$g2+E{(F=y69`@sI(O#39fC<tczi90+2#NW`%#Dgq5-A{1QUPc{Ni|Sf6^&m{d>y; zhu`}v!)2y?AOnz=ErI$vc~TrdhVq{p7VCWKVF>5OCZK(M{`;VRzCAi)tg9I_MEBp} zoP-iP2gTz>`TY2-{vqgad3Er)p!c#`iFvJKd7Q^oo%&S6{j#oJa5SrKo>VnG$s8TM zObB~CF>BgLv94x=+vH-f1pR;8V}W?pBuNGC%DIk}vMMK#roO%X?vz#O%Qh|LnkTo!`#jLJt}S1(X`w7kF4pT=!2*w&nAhp&lyu3`4? z#He9F7p1hKqxqp}Ov&(ecGzkj^;Ug@Ks-kxjY%L*k~e@SJb zV)--(!>yct`9QoCJr%+Ug*Qi z-dDkYuqLUn}QBYE%F5&G#+f#f3x45s9q0!^XID|G3Ri395hE0qxSk-;1d3r zJTmobfvVSOYeyd~m9qSm4wU%*52J>^Qdg!;NoTHiDt2zK2TkAYBB|=~6|F zS=cB?`t#e}aI5P#C||_pA8$<;JCZ28K-381fliAIxx4h|;*9m)czNayHjNs@9|89js0luHZwkA3Z&Yq3 zP$(s)MPe{ho&=>;p6`k$;cQ&$%iww)@_S$ZbiB^RP(6b)Lo17)u}`!D#5hti@9Q%fV%F)NAx*@@)7pRImUf%9s1n z4NLp+TK&mQov#)b9$~+N4U&|GVsC@#e#KT5uyY4MQWC99h}g##lPvWUkQ==%4-Q_* zQoREqbE1s6V;F2#hN)AoX7E%p!2S21vmLwX+-WpqA~vD$f$Fe}QvqBe;r-lz@2Rgf zpO(6<*mwxj$U#Ke;6?&bE+41_?H?a)O>6cf(>YAlxrp|3uL9Z=%S8cDH5QASIE~+2 zQiDj?eN3A0U-1X9V+MnUW_^4U4i1RoXah}(nqVDfbPCIN_dj_O!Hi?Qss*yr=sQpN zA~K0eNG0rK0h_E3l`1#EK@1~`d>o$&KkPBa??&QbQGkm)XZccLC7A&uZ67@6=)IHt z(z%P_(^I~D**H%IFmnvhMG>uN_A-@1;~W}s=kMW{p-sWZ6=kLOzEyn6;4-z`X}XGX z6Bdkqo>)t`z|>rnjq~zy#}@(+4s!EQ2r)Ats#H$nQnk}xKUuBb`|gh#>V5*~kkLde z|9RJ_+O`|Fg zXML2_nE|vt&j87ydm?(@B4b8~R|_smPfo0L5PH7Xv*gWw%!)~T5)O%9Qh-2~J{*5H zBqv~bF}eLAPHMybi$_0~l1eb?c__?G=Xu650w>Qs*_P)~wd zQ#}*&vdy_25myCZ4JWynGG0iAUJjBVm&3#Czo62Ow5lMTR);LGRXGjObC_1-pcqBW zYwW$;zw9n_W~>CXM`3x%WMYFLLqeB@6g;oA1PeqBX!btwPt!|=-Kt?Q`9Q&d|-OSb*2eGm;|zSNFne(o|2V}%f#qe$RocHFwwgy5gfM5$8+IN z-}`Np_C8eOeTXcBB{cwY{gld#U3+*Mz|%+pVh@R9+2|=xKr8m*mGX0a5yxfgi8vrq z5bjVdB4)~a-RGwVKZb%w33m2DZ$FK{y)8X{fPW>|B2hqk8MC+)*7Fx2(xpCsQUIKH z3sQtk1z_0MUk#mNj5^RbnBU3E* z7Q0>Oa0XZGcnFxR``oCbu^p@C)$IUZcyEmJaF~>5A-<~(7pgz9QaI|H6lPQ$Oi)8* zLP5&~^w|KN*nnCPdP@ib$uQNDC^#NmIZ77drIAiC%|UWg>kY9AqYEkC)#W}z0m>dT zvJ`pnB&)9+q1WKW4!hjl~c0kQ4-cDjCP6xFu?TF zC$U$yMH^ve3vKlQsic40?RlJf_$0y#2IW?4K` zwL2?Fp)^ioPAaLuDyzV%T##u*a&jZWE50dludPXS_VBM8=Z$=VLdcmZNt7*#SbCyU z)cVpnGT2?dw=`LQ4r}sWf!bs)^aic;k+BKc&w*~ZsaOx?NxQ_4T+yNGwI=~e9) z65NDWQ3_;PpSlBa0cC=c6Pc+45GhrHc}pYasb1CAT#bS;?EG}2(x5fufqKq_5kr{i zF+xSueTKaXRfU$fI;eCH+l$k~6~LZoIa6X-oP4S{wo4=5J`+rSm9n^z+RsJRAL5cl zjnj5fHvw^pdyd2wB}a=^oUUoXs1dsvGI6RIJjW64+YMSk#cZ`^_g{%Ie~9OM!Xirf zWL~Ab*tM%v&jl&cWhzpY;JY!djn-(lJih;`$NPp>O>(_B^IQO+ zH3`3R(hUX~%NnPl3c*7r3j7$D;|NOq#-7@Bd}`8Q%|`4i{~3W%ljH>K(rj&%IUP4& zT@v&O|Iii-r066mo#(bmXf>Nw7UCp2o>&)4g~jA^k=L+lY$zjUFiC_Qtic!>gHZW% zvP6hxBV0lc4_7y79OzTfRVaZ4n@)~i{N65W8?bu&6aoe=?|}1Z{4hU>re%~ar2U9O zAUl%(!1McDYiO|Eg<;a^i3Y^~rgb0xAh3tf@zjg6(Xg04L5O`-@)wUTc?NoaR%bK}WOZZ$N=Z?%CC3j@rxtZ%FHlW6n)rC@ zYG4T*+y(xXKq*q$3_Iq=n{jvuF&?3p;iB4Zw3bk*^g=yv(MZEgWSu%YAj2 zI|P1RIguRSycg3#r6aW4yv+)hwJ3zv!w*)UITeCvj$)))ek0HH0sQz0AQ>NE=6XEB zh9}@s+o1|8e6|QBizLN3f=_@A$E3fS4WV&e7gA)VLVQz;9gl)HorqcGO)awkeaxpU z0lVc9U0LIWMd`8?$t!Cl9jSMk_#z;XolEZp#)XPPc!_zzAdKaU4_j8A)WTwSA}6W$ z>cBJB2m8x5@=+%@|IztnSr8i+C>OxksvDRy>%+ymd}XGUnbxem2V;|H`G}K-dF}ei z)PNtjPKn3}9%d3Y9a5I(r)N8D1rY=398ajT9~@0OIdmtG={NXbU4=|Gsx|b?h~TYl zcf!LuSFvhu6>YWDT0t6OnA(+tu&8%s&EPa9Hj@R?xF|$%zQjEp#2oC;ouHqrz@I#- z&b~6aaB>J7+RIOn=)#VdYo@6F+;1{Nu7%d@>~N#FoY`k}Aob@+v1Sw$r_iXxfb$+7 zt;O^^4?du{o&Yu2_^Nt}jC5KXkctQ7B?u*F8#Ne*U^1;Z)9-^Hw@lEE0rf~uE;AK~ zR=kh0XU990OmFYJ_f_X-i58f=Ki%-@F{7eUlvEcrm58g+#JS4M=_h57n=m%``ZM9{ z+W<34;)DstD3w4|>f=CJR6MgIuVLdV;ouf+E5sm9IbT&V5gF&pgtE8{7owr-=UCa( zsg%nr?LiBLZ#aXlGJaBDDht06wwj3hEfv*uQzR&<(bkRjl4QJj4XTa#Gk1H|dC3hVYLG^B|b*_SVI;XDSt25AN0m3ZJy(}RN|>I!-5mq62M zWOADuQ7#9}g&xmQHz}d@C*w`Hdq2#Pnn_p+OlZi0dG38!7Ih?9=3_AJC0};mxz7?q zDb}az%MdV&H+aldvDso9%$WCx<&@S`8A6SuDS?vab$g`59`l4v7Z;*}#^OL?q-hS( zkYPMR(W5_;XDt6%wDn3>dBP`Z91^|D35((KZ3|rqaUy7JeFN2^DvA%XsJJrT_u!~6 zuvoPYD})L>jh@_|{3`3b$eDpqFl`@{(p|cgv_C$~82WfAS8}-j{zue-6J+Y`Hj^j`A`quxt%6mjFr>ZI(e|mn3BY%S?W3DPme- znW~d3fin(pZCWh-Ke28Efis&_CB1k8S9T^udo;F4vENB7>S;>aV{grCi|qR|2g&5t zPWF~G_LyJj9j=eDI(^RJ6sOOF?5_@%d5`Cac`1meGGb77bk$y!HL7qBx5j}Ce%(=x z-OL;Qd5B8nkKUq8sv1Ja78-#NMzhsOk@l0ZGE+Rqx97)gml;45^n;wj`f6H@FXXbD z5aC3bDL*V_e+-=9Jr-}rd;rZIau|P@fJ=O*6~=9*qLM;Scb&LA4O*@ZywHa=PHqHJ zl6bqLkN=wF3anjMsMI3DmM@ztdW;&*6%49&96ukf@!8I;M0vO&dc#ghZNq_rCvlxtSq%tq`xjeuEUQDf*Jisfm zXgR-~rsZbN`F?M^esywF7iBSLaCTr%2qK+ZVIDwVMgkd^MWM*0XaPW%z7leQXK2s! zhfz+TLl*+(hZ)ChqDnq5Xq}_~kOySAL4%vQ{Mmm#1EL`cR^A)-jN7~>;_`9%`ubSe z@>MNGo;e~I==R`3gV!4JWy~^cD$VRFi;9nXBp%7-MCCC;Hj?QimqBh~`l#heH5CGh zFu(KwaaA%?!HA?fSm}3~G^2+z1MW+COEaSbAyP!orxB&Rh^-9GCOghVCdRpPxhol_ z(G7;J?;mc%q4?FRW@?CGoGzEFo5MEMm!KO*hzLeFS0a)9uJU(YGf#A{mDp8IIM9_& zQcS4ibe+qS*WIv)5sxl?21);$d1zqjj3YF~*!|&6{O$8{%O(T&iy-)5-+r3kQBfU1 z35C=oQR0^vyiRbNZqSxFU1Mfd$QE{CGA!1~AfS_Aw~0Cwavb3I|8Z45hGw9q>hHZ) ziYIP>Ghyi0yBfx5RhsJ01Yf7~T4j$DHCysG2suL`M^y4*)JgqJIn~E= zpPsN~*lo-<2fQ||yn`-`$?Ih6k)I~Tq={}YMedTwdA9&hz%DDGBv!BaVCTc}sI>1u zmVm#DAn?^iBuF23>bm)zwoCGmNspOEqI=QT325J}o*i#|;~Nicx^^`l8W^<(xTNea zwzVVkm}z_3)1qIxlM^HDA(-(*W^@Q8@8#Zw?2o|vn?G2|Foo7S1FeisN&RrMReo)j z@UDpfVi-PVGyYdfBYq<5%Rd*KNVH8&8UyHpWP(2)f)x#bSt$Ton0~T3pprTbU9|nR zHEk;4zsU}aY&xI=O&~K@;uym86h(-XGjf4uf{9}`ClKA>A>*+yf^$YbgfeBwp=6k0 zx@ib`2%&pbCimS|OLAUIc0l84T&5Lo1ZKd*;w>cwzF0)BACTb8H2J7@3w3nr|2hAy zE#>2^bGY5Sqme+$F&b^<>}t=0uJWS`iC7+tQGS$Z?M8QCDL{e2m>6w8hl^=Q*)}~a zXetLH5|fCiCD5HDh8Ps@w-vP!6-{{@d#vwy}* z#kkU00X%FkwMFliwnI-@z^A>eRel>L@&g3TndRoS8o;ft0QFLRk7B}|7)cQ+7Ed%P z3=AZ?c#=)AqdQoohgu->Qoj>=XiRm>J*Us;=TP-b+*QG!IYLZTZOX+-AWDWwk0oRq zpcXMsmwFPSpC!l8W)cOzk$J(PcZL)U+@2AS6c-uA{BrKXd<+Pkqn5w4jcD{~g30Rhd0K=Uv za6yK7TPe~KsN9mZ$q@4X*<=v;bk^!aa zkmNN|NqI;ZoPSU-YAv!?S;%dIKjW$6=o|g@4{bv3=gwRKSF(!&^F}wXS(V(KkUiyZ z63u+pD)3>ab1-QBJGj$8e=z|G+n#P^CJ%?>>q7++uM@oGe<^x@$|#v$nz~rU%Ly$8 zkdM$sn~ZP^z$>=XVqhAdV+SDkpqvfHyf9k&=0!IG*u_`UZEhAg2NaU09sOXJg4a?nXV*TeHxi#l z)aWnv z4VS4>q8cS37<}ZGG2I`3VVAFl%9TUec3rTTSo)_0d?5Q>qs;>6g($|}NKNN3#H~4W z{FJ6F(iuM&5ThKyyaCj)QFv$f^TRYbcHo{-39%jKhrs4$u;qssgnA`V@bNN)?8=iC zkL&p+_@U+AQ&c9;ZRi3dCl8|*uw@#XE{S zLz~ff?l_puTG@J9T40N@8?6X*w+sY!nfy%HGxJs-&^IgKazwsMtk*v{SlwSrk(DUb z-MoQwf(}RXI_Fe_9ktyU7ylI5KaB}32|rtrT@nD=?)v%AT?w>!jIgYY(UO9+r3E=y zD(G9^`R{`zU;5_F8v;?gFi-Si9~-*vat=_a?^EBjTQ;NObyl4!>CuOV!HU{BhpBSA zIC4qz*wk1G%Y(P(j25mvj_QcVs&(bJyv8Mjha>h3L^X@oVwd;9s&;-(m5lx?C7E|# ztmxC9Xkv7>73<_0#P1=Lz6WjaL|*pxWpG&~m;cG@*EzF!^ITEH|Gll-Jp|gXoz$HF z8quxsT2e_7Mc|8`tMU5+Jq8G2k8yW4V6LQ588G|PcXMI4yM?{q!!avO?>ta`Om+bb z$X=az;hdA1 z((V^K(vy^yy^l2LW%$Xq@V4qHuS&FolL(qb>9* zYF|7M7daK*+nFGYkVK?m@zHyUGQRU4vX}ZenD6dSuc9k{^Ds0X1qR!R;oMyBJ^ee zYpr)GkEl7%gMn7)_fl_gDIxryRQ| z<=HV6k(-mXJ@vDn@W6$VoFvoAgqXqMj~+sE6i}oBQGrY)hQ*nFb)UTdN22@D&&Z>4*At{{{@0173g?{^5>yp&U6iZT5u=J8p zK+Q}Ig#*>vqw<2ljsTXSB%pszH~2(+3Ile&m+A}P5Gk=`iQdolqh?snB;~aQ!WcU91=%BXJA59n8hE1Y_2WUzZL_t^{_4)1YGiHjA+lI)QH1ny!@H(&Ub01 z(|#R^{0I&iW+ZZ=%0^M7o`S$c1-&9pJ>B&PB7RT7+>=DH!CJ){pVe<%8M4O~5sVeW zsFF-uB6x60%=89J!5&RLs)Oy(*v1;TY^v6WHxth`0HE;#(?DwW%3KZg8LCyCG#~q= z&c)D+f+-@Z>5}JmgVCqAn>XU`1Ahnfvi*Y9hoJf!v|?X87}4_ei|}XL`uRV&`Pt~| zMY$3;i?tIvGH=AsGzSc^VVA4Mm|RtvkntFT2MzAC7T0f3BBXc z#ejqwN(ypfqntjrHT_PNONn1N+6G5!%a4L7^)ysBll$5EZYSYz0wP((X1 zMZ!t=S=Xp&Kcg4?0i_+nCPP*SsL<=yzH49F$9R^rA~?ujJ~IE$)`PG=znu={4rY9Y zlxY^}$WxUQ$vGHOfwl{dcrG2>&~v61bTDeDtNyLRYNh_$ohO6OZ{b04@o**{MQ{+5 zqxW^pg`EvnwkjqWm6;4yhA!2TMCWzK(WkIRejyTjuwUry%V$ydCPy~>JK5sw2duys z=2g~_x*+RuYbHCKc+zOiW`CorqnxG{++s0zI2h3cm_mYL+HJc$Yje$lJaA(&PQvK>TA+&c zpdGaU2<6v&*DV%nKXmop?86f`29Tg4IP7O*t z6t!6&3vW;1@0W!>7N^ThMX$9*Uc*K{gt{rBNU=?(e`VsME6&hJLYWgenagvZPnkT= z(bdWXxcue~mfmQ!J^F04S4^+v;$dp#bC_j_7SB9vUnpWy?vRIo#eo<9#zP9O9Ly1E_=Smjm4HFTvNH9imyjQFFsfk1| zTKXy!0&mQz*y!W@>>J5R3X0hHox7FD#bk@nffl+X)ssNvxFvEN-BFt=_X5hVS!Dky zEee4|Z#p>K_@-Ko^!+)k43x|3Ske|ACK$e~-?~x#;4~A2%}fbFCt$1eV$kQ@3($_i zvqu8HBP|bn$dJQ(xsKTPRZ-QpUH+y~_%|`*h~#b%As+@EuHz^v0afHMm@Ojnuy^2@ zP=<2X^t+c%LwSM@eWdkoUF{ejcWy9KXn}}c@&0kPXr`U`As-bWCI~qJ+lOT+(Z>;rFvJM2tyRn6 z_-)P)p}6Z1X$nu*k8_X{O<}@Ng>GuoelYmkJMQ6?lbbQM)3s zU7p(LMbQZ+&3zzD6-F?|6(C~#L6x=qaVJzZXX)dk0{Gqw;^SN5-ph}FjuiLTIP{Nxp^EK6 z%_3uIj)nGN#7GA8G=8K*&S6KZK9O^SJ{~2g6|ugEV1=4Dr_hSOf?g|HV~XwN40Zna z6#h)WE_P~_?lG4rqoDt$8LN_Bad70CO}r!~%XqX|%mIHn>+Y7zaA7>Rd5zNJxx|N1 z1r@o=um1L5Yt39Gl(v5T97*0}kW_{&AYME_5SdL;9Er0Po?Wi|sV^fE+V{2yd%|Gw zYpuh*pjIkv2AQwFCWZ8~-`rNJvuyHqi6dl4uFXnb_;%gppEvm>-JfW_L=0%M8bsP>0)%0w!%shue5C}|KyF^7#!$QjH z?$@@=JUSR1^7#^FKNa~ch6~k2LbEHz-}~LuzI7N}H)Dl734(oz&U6v?SsnewNSm^> ztCH_m5QY(olPVv5U71U%^>m$!$LhWBlqy4Lp?;E5JM}KEv|Mh3=YqnP?B(y2BOq$n zZC+uQ8*aGIz9-v#@Bo3Mq#9w%A|Yt03n9{wJEr;K8xGMP@_O}ChDIX-1=;VWxZB80 zeLAYAO*ppAM9RSUO82X=g8f`;XvF~)-cH+P>(N?4ifePgj;Kk~9el=`2M3P#yA#Mr znRx|Qc~R@H38$R_*L?eBBf|4we>q!5Q3FRVlwD_p&SvoKW2Zzl2q`nsgULx3;xEFp zY;uxe=Z!@_BES9$q<8IA&i_-ws+FY|y#I+!M4G8o)Z@F<*2hOQ`kqaRsS`0sdT9}h zM5K(e>4?*$b8Ad5djB3AFqdii>Fh(8s6mpFJZT+Skf`Vo8 z_|mQ?xnlLFqnvsgOMJ-jY(RzISDv*ZyTg~?=zhM^;z`=3tYC}uByJ4Y$phKN+Q;C& z8T&=9OY5Am{MJ!-nM^s;d%HIqvB?fIU%YB>xt}}}bP%k+w=w?KkPsU#d_`7Ym37CQ zKK5n_*o}I9SAb?dTj-4AVl)8eDn)du=LXZqr^@zgUtaESy_+j#kV}a^`Dg&$rHW(4 zVxo$g4D9|vlYbB9qO_r%D^X9(0{@Dn#d{egizJ6SE8etq(|j)aHmJL1+gUSF6IpS@ z5Pr^h5_VG~y|<2atL2*VfG!eIwKL@%`Ez4wYBTorjL_FS)0q zaO_y|B-v%#NV&$K!%}N!=mya6CExAn8ti5PMOj5%+jwMGMTe!xvVF@UgaZCy^rTap zY7TEnw#fBVTzZ3Ej*NoT5<-dl4a1q}}fo5IfbdeGZm zIT#IE0=-hUyI7Rq!BED}`V@H`=9}7b$G7yOu-dMjxT9*EkYp`PZA4x5d(s^Nu;btm zY=qiOo8(8qR>Sv8Z;8dt;x2aS!_8vf00sG|P%X8i{`&;|`c5WY1^u>oW6#gzEza3% zQFK8+G7|2|a*J!#-k%mdwEw}Y^rae@sl;v1;&dVRF9KxA>K~PQ-cKhnJYb$Db{9Os z>fol!oW#GSU`kdv@Lys@mHhdWHr8!?G{PRNjemnl1HRw+)g|zmke&ccPw4@ZM_xA)&IV5oT7DBboR83{a zY3{0VgYTvOr28K!h$;g;+&a0AR3n{tTX7m?85hKsRu@AFjt@XLo2|N{qu^cP0w5!b z*ZP1t${6cBi@UDKja*JB3}V$t-*vXp`dk_dmuC+OiDOrh`l38QbFS7|8Yz47s_Wf6 zy~7N}Ueo@cM3?8)mDPl?IZ1v&o8)7{$7ipq8K957Jh}yw+eKkFTd1c!3F?MS1-g>X zL}#9Rw(XPiZ69~99R=J?ahOr{n(JgZU9ox$frOvr74T3-J@laN!igPA3a&f$`WUQ9 zFfzXy<(3yzr(BS`PDTOK&7t%eWflo?Q#|=;d^(b)@3}X>fR8P<)YCgbo@`o&y(OAA z{77{ERnsNahFwtipFi0IvqO&@b{2=4AkwP7J<6s|T4#1NiGHwoby5=z`&H%0-f`v% z&N#D~wHdsDt=-wE+VW#jsrJR-$ouf9Z7ki&T@IA!&ZP94OE5}Np8Jh#^CjdDj9WtW zW`=qkh&>^g=b4Ro#4We}w46DW-?~#hYxc1$z-w;xp5%IWkfQwq28P*SN@CZs(%0v5 z0xcV5W(G;~Mb{rD(vZ0et4!9pJQ*!B)qOUV9(C~8`y;Q(@ zdV<>3@;wVWo=LNXK02R|CWI>T1-|N_6mlG1)4zm{6OsRWm%uJb1Q)cNzADV9@{X1G z8ul<}e*%%T{FChA!{0yh%e6e?58%&tZBrhnd&&cf0{@&uT3Ki3`JHYE*{pZ%3 zh*=!Hq=^gMnv1`i*?oXG5p9cbllkW+hlyN_ZscqwVp35Z?6*rzdCt2}Td-D`6*87Xl4#B<`K9&j8Ctlj& z)>OSF>kowr8l{LCU^&u#Jh4Dm>S13Z$1J`@zh#W9A_=@W{AN7d1VBX{@SQPJk4)(4 z8w}N6{D(OLzxP+3crAVW)Np^gQ1dCT;QcA#jF&-a-VyOdujb?31E5_oA757QIcWFMIaw?OTU`mOnj- z+MKvQ-C(Q$q8!CP)J`23bG>y#UM#lz>J{d4wr z9&?$>iR9?*+$;Ml{jPMEsTephPl1&W{9Wr$T*VT}bGzXU!#Yq}-^XOk=n-0+rWNJm zR?Ys}Iyqu54Pra9Aza`Ujd+5rd-`ngu+wl!0PN$IOq+gi^AmE+T1DJN>g5zs+9b$3J2E2|rnyyz`rap-e5)|wDk31gnbyrAsl zs|rM!*8y!`0pA6*odFy$4mTV)k(!W`g}5kMfxyev=N_e&0##L6R`4S_ljjnC>y1+o zE^OQ>C(`91`AAmm%=EB}(@$VFGCH}s@VaoJqdg4aZnPbyS2xRrd|yO2bE8g}Iyqa{!FotnIsVu@*iIndv#G^>79ooM;qR~=uV$;2}fU;{T)to|1F zggFO}SiT%cma&0MleZS;HMLdq*3?3=0XM!ne2T5nXMtlvovGp@5(jwG(3_d$5bsRy zh8o}0%93MTjE~(xH?$^1Oukv34jQ$DX7b#=iBBV%UTxP?n#H={ioJo}3J)f;w~f0| z3Boe(34t>KO=g4T_67;dNgdLY;>8VMQp)HISiD%e|)}Q;-5k- zqQT$q2wdAb;iJx9-yhN3vM9}UyMd!bjVV#3GDYST{ug*`(JfTtb*%oqb*d1316Cc^ z=RaxxAW=6=FUf2(s!IIteM2r#b`_?rgxa>h&?Eq$iM$!%M(__bfwLr}sx>ce^8LMa zE(u^&H!tkD;y*VZrK$j>#Dc?wcNGmS+( zq+#XDOT{!HCmr=XsivVVfAa_DCtj1aPATcURu=T3zx3OlhyGgA$7D`#d2oPU+`f<* z5A6Y-2j>rJI9Yi_`1i`wyUX7{1gDAIGvP2UHA;7HtSz-{swo6f+$Uox4{1QW2%kq6 z85PuC1nw&YVC)|lyCzq<})D|i&V&JKO)Ux`^Y3LOTr z1T?16`h4Zn6IFo@C)bs}4E>`8T;eJ`OoEXK1;LtB9j4vg1x0k_WQg~zS+#<-{ukwV z^yQ2DuIc?g-KccB4ts3ig2d}y`~%UXVdX#C^?S>%Nlk}hC7Z|J(l^%c_~lO~0{sT~-F4T3GK7#gO!I&iTnw z%Km4`oyMciky;sArW2c}l8UUT~b?|Z@U{hZ#K26eKSR49A zCW8_KR%6}gqW70UoK@J_?4da}=n_aA*q?JEtkW+^qZL-Suiv1YAL0D9{6zQtI;sQs z)7OT@I;9L`Tqf!I1u7hPkv8&y!0j7dJ_Y18C2rdV`?lxNcF~X{6FPNBj|Y#tI6d@v zEA+k6s(HMd^uFTHiRuFM{-WyywCoQ3nHk2RZqbrwV6EY|lRr`XTYx-15oK+aQo_;W;cga=~^#G0kZMkn=@p|14npUKE< zUQx@9Z{_9%Ka=Av-nq}#HTnE-We7WYStR++6=Y4omXPBBC68({?NpUb=co*@&SM&x z+$BF{EnFF_%AHT)$Ra#yKXbVkr zb%2;%-ixRQZ&vrDdvr%$!yc`8e#>LrxBB(Z4z;}#5E53!y_r1E8(cw#6eB!c}NritO^EmYEsASODZgt?8hhaNc7ZFZ;peN7^!I_p#IReKZ`{y(OyR^4lwSvTEM6)+lv0ZOlm1p-DKL>J zboJ2@NZ1E;o&amUS?fYQLA6{vuK~_7BU> zKF#jL+L21W0@tZ$qVL)l`QP!@MOa~o!=HH-AAXfo)VPjuel3o~4Po3>G;IqD6}jD6 zZ+HU_4n5tKOPBst(p##7#1_I^EEo3pZuz?>TJgg9nskv~=mpZ;M6@igjAd9-~E5oAf= zz_GW@ouf|#-2$5C`*W%9c9P}9aIv5pBrChPi3$Mdt|Gfc7`F^Ae$pY8Z+l?>(Zeak*m6TkxQG zMZFyg5ew9$v!?QNORap+*zL#b0wk<}%_=5n-wz>CL9fn|8VN>1xkr7?!HNmx5LScaA~ki0J{ZIdGPGvmLazX%1`f`sqh!P0n)c7aL=|o z-!O>Bt7HQ91oi+~K&HRNsf*>Y6h3;FMJoXu!hRwz|Hju$_=YUT@zw;k2?>;7#3>26 zEyYpgDc3^z10}ONjWVp0X;%8}yOx~X7bhR!+wFL$NRUr-{^BP=UEdt0m(Ff9wk$jx zS0SwvHws}-a#IT)+7RysZ1TlpMJ3Aa-4*^jj}EIpbgU0Idf9a=ro1opp931&RkCP} zTDC8%0Z$XoeRD!LAl+ELoxglZW&kUnGA2!REwtrxj<{hip6nWBUWXob44s+H`^xRz z4Oga#_XmFWzfe?tZ}w;%Z(>;5*$1dSM(V&!D>F6LaR0V(u}q?dcE-5VjgpaejVMMb zBriE}m1HDzx3xZ)Z6b`giUbphaV|zJ-(vkXA{-lfJ>fXGz=P(NyU|y3qGu!o1xV>A z8`=WvEZf=leo+j$!8+buxD;+IO+SdJiW#KQ4r|b)-cV`?S0Y3$?hHZg*Hx zk9nzPyVeW8zB?^POL!M6)%B3$y=x$@!i|}hP{Z}H|pk=lVMb0eiKi0hrsLJXoGMnJPZodN@qBMc!6!2FM zy8>}r?yo#S|Bu~A@}n08{M#q3{yyCSh+CpEE24(De@X^SNvKOq7hQcL7oUn-=%=}; zE9)p#C|F1RUw{6Kql*)torjCNqnnGTr;EGOe;xz$zxE+O|I_FH2nY)E3kv=pp8q2% zDthbxfBw(^6R-cl^M4{RFfiw@p8q5Ab!Hj7L46d|jbPvv0<%YA{QIaILL2lq<{Q~! z7w}7QUu832Lk~ya02^-yjQcjeo^Bq#ZZ5V6e+O?L7Y}!YAg>@VKR3eJ*Vj{mkMF-9 z;Pvo!;v@gPZ-Rk=_|NllayDmOEYlCCPFm5?6^@0QS4>T?rKG=K{JEo{ZbjZ zmYgHyZaamUTZx%|HRYq;?`F7;bLo`GvB;F6O8)X$&wgNiGHhZb@EPL&c6_jaRXcMm z(%C)WcWIiTQ#`^L16P1C4#Ykj#)Mt{+?fW^L;ut{g8LrB=RvV@0+ z=Q7hM@a)+$G0;zh!dEyM8QH?Zg1o#u=I76!uZ#Z}^nTZ35dAh|3gHDCadd0qQo7HP z+Whfjd2_^NNmUgbgM#Ae#`nAW`tr)k%7Z)I-Q5_VsZZFcjdq8VxyI8o(!nK33;7JT z2D%>fgvTGhTgWhVy`@sz#?_}-l}bWMNw!uh7LN${tZD@vZ)^YfoLhO8_50w*(*MTV zTSdhcbnBuGgy0r}TL>21-4cRpkl^m_&^ROz+zIZ|NN{&|4-UcI-TkcI|K4Zp^KjoT zPYed#t7_J)PpW3sIy%z6q*%u6Qkmug`$h^3L#m{Oo*3hUe>JPSZ=rrCEd7y-v++8( zTZVbvo!8_?PftHMHkMyf623E^A#W6GCLAIg_Yov*iBMeB!qx|F`M#C9|=$kq-C z<)f#g>snZ7)V(aMYb>7HpZ{&20j|%tr*20Dk*Oi*5^oGIljw{-SZZhwCpoGGFB$|d zak`4Hwks!QfeXpfwUT$; z2(nS{mvC}?@LllEg2;S=6BF_DysvQp>uhxVuePVJFE%!|3viOqTUZMJ{4wnCgFz+a z2(GEAS>$VzFEIL7V^N~hoFM#i2WMw@>fu2pl7ven!})r0JbI!5z7%!SDw)A0mTe!X zP`EuG8Nwgar7L!~c5~yQgB~DLeLBVu!)5H6nfY$Vdc05_wB8-Gd2$j7v>a%G2Y(%l zLg3IcFhJP2xPiI36lG;);l#YLzkk1PCD99RVmW|41Hy5%)EGA~APGG(@fM9xG?7Wi ze6>vw5LiYRTZ0TbP0*7-XXU=nVav_n)!Fj`V5l&JpE@qKhaH@qDf#(RkB+RF^xCu) z$|Vt}HmvP>WxhV$JvX+6BRV4Jdt2i^udI$fjrP-iY8J7!W?E^k(`wA|{+ty~}vT|YaTuk`=(hwd#pv807X(aIQk z&oNsOtKWlb?zt>rTulJGTkq8tPLBDIB}zJTopx@9l2m`s6Et}VQ$gSS`J&);XH!9F zvllOehMODD)6*05)8C(f%@`57+Oe{!R-{s*RUa8cA#w0LpK~Ki3x1{L!H{|$vEnjn_5jhQWoDi`taRa9cJZ}TenK*cgw^CW z?cD^QkLH!#HSk4)=?LN>oW;^!6_sH9%bn5B*B#GB{gFELI3_DC?mRp^OIK#6E3JV( z7F{EvVqykXR>MY7f`M^y*jjbgFsO0O^_q^dh+kPAq&^fC6_se#h62#Tz{q%heN9D4 z84w+fDf0Cz5;F3inKBX)#OH5CDp!{-6iWrX6uF7j#bR1)DD*{FdqeCfZCP@2KGB}B zLqpQQC(7Ml;uM4i@6I>v07xs?XC`*~LZEjyn?*)8qyDdpsPR?^8qE9#`h# zQKWKWV(^NJiVY17(^@C;saz3vcW!`z2TSy0ZO9XNy2|a!5$XdYP?(0Dl%-kWhMLX$0DKK~6|SNOspkoZ2i+teVn!{jMZl z1E4 zBSn+hhz72cQ&Y5`J{4MsIe;D3+E^)QKM8iO{G3v$JKKHyTA+dF6iM?~x4qiBXlPcN zz@(qu;JZ@>flze^VVD8QPf1NZ+3x##C{$hsMYJR#Z@$CXQX{}CS8XM+AXDoAwDlVs z2j^F9?e%IOYXkv%Uk<$klysqKS^HhL$Y7yU+3ThsDRih9w6{4G zj)(9#F{twbmW}7#Dl<=_j3Sgo$@=vR9soHCIy&k7ZvmKd23C-rkTch#3~z1#JtN{Z zdsk`%0)3hicYPu^y|RzmC;~78r#~Qf*tOjUo}Jm7jwEF}e-a5!HJ=%lnv{n^y zc?WoS6SYr8f4>z?!0JQlI#x90jQHub)2+>C=1CRsJda9dkR_PHe+=UKRhy1}WMYc^ z8iann3|>9Ino?kQzc9WzT`hMu7YSa>QYymVTm9MWt)t=25f|l}7aOBC6elb?RV#3y0t#f9#Cq9!w~4``h^mYE`g{hp=yajy)qLR^W}tLA#O~~b1Fo@ zJ7gjZoux)bCWCA`hD*+5)y_bO?hnV2%&?J_As4^0WBI~&gshlaLJU05r8H{?3u8G7 zBF@=^1Kc*c(uCMhI^-d`#&0U}ralB>+uzgL)cf|sOswVaoCHPq>AUzj%=BrvAMTqb z2l}>T;1p4)AJtsJV6%;$kfY`1#H1vAIy$<)<{Z*NQ~?cNBkNQJxGN|yTN+pb0F{N zPL^j(w|A*&EH!wCCqpMW#zg&MO|DpQ$(t6Vu)Zzr6EeJ!*uZd9GL3q&^z+<+)cMvm zh8!O)X;2=PmXQ%!hy3Qvo4;+|o@1H93L+w~a0uue>ZokNDtrE@dF_pxyJ6}TRLRS} zOr*!=R)f>ci|j=Ocy+352~twA%@4|ui7H;!)Ys$T4@}irJFaRHM1js0gMRY1&exVd zBOpFbZPK|^C)O~FaZD9damp1_9Xv3_GXQto>RC zj0zo&py!P-kj8IYHDlhqd-$N2Q(Mi<80DTC73bWB=%*TS{9hRKwzf;8kC|4)@pkOH zWlc#4-|%j;d~~v2D`|f1zn`5tnk-Dy{-nn#gw|de@&y#s)qdY&lC6VBbB?54$~D;$ zi1GvKx3io7aU>y8qA^AtbQz;`&}6v%Tv}R6iD*)E-&%< zbH}P}v0dzE5wb0rtDAT=%R5uww_)InyoMZ)@abie)XNP_VZH`lI^y*E=%v!syK4BB zwsZUmvp>xqsl4i!nZY*#e@CbGH0S9!HD|>#H>c7qI2nJFED^ieGzhoK-sKdNZ%w4L zExZ%@@WF3>em-S&b7Ny~rM3Ad77{sX@2vEwGS=||C;v|=!PbhJZid;q4>shgQ@!>?z$@{%}6*^8nZ4#Tiy|xc=pKRr5qYlCb zb>A`TwWY1JdRALZQFGZYNoERpt2a7ACJm6t-{%8(yab_I%sd`G?B-tway70IBGASV z`b9yLFWDAYeV&8bx1U|3kCuagJ2nfuIPpPLrsp=o<$Ihqqy0%02`L zNzzL{)6tb1pO?jXRWtjC(lLU%;g;QBRwCUR=IM#kE#j71Rox+hiE}hndLz9D`>R%__fj zrhRGiy0)KttM>#mucdV{HV?v$MC<6A{FzSM82;cTyx&MA9Yk&00boWDsYh^zpeJ;b z7?(l)bGdG-+ueFlJh}{`LEaM4;(6!Y2cP8G)Rp5@2eI7k;#PtH=B2?o uq*w!Ag zWfEb%p5XCY>_iQ}>U5k{xsyKMkN=XFlcpSW1l6C76_(O%q{it@wss$R-7az_4MGlC z7MS$?06h$-RD{ofmPbNEV} zST$RB1&ibz&8)a>Uc;v5UT}^qR}#$futfvl_k8L z!L!fcww2V*#Abze;JRoe8T5(O@p!zqkCQ9k>k_c%Zo9LmTB^vg($NFo)@=Eo|Cm?j zO4Dcq4*}3taEq|HV?50L$bNU4zxn?6_luk4O7qlHAmYR_+qpQArwLfEkyg zL!;5a2=6E1>3)7tAs>qSqc2uMn(YC%;};4Z8=xVrX73*FHly4x25AVmtnmT+0X=M% zjra=_P0UGx=!L%uCC*b=2c3@8nw7k5m(dbCEzP>)UR4}!_ zH&cuUxQg9gJUtFKYwo6>(MMYAw~f zz|Ek78H*%U1|^W-ezh$;W#hCFh9rrpch!c7Z0CnG{CIhY*+S=(dJ&_Wdtv4i8Ewbj z9L%&i$Nay4D`zi%|9*>&i$lYY!R$%{pgItt>2h(TtwmrO!;GI^iRu02UcH3-*nftA%?V%hswU$wNJ= zc-IA(%Va?M(r#KA`^>$o=_!9lr3Z^t00&P3*6|*k%A+s@6n*ZLGv>7dl+*<|{!55TGbKf>o|2LsK^U z1FZdch9D>p0>Q$;5p#Ct0L--YQbWwopRo55??pa!b@yff4({@FWo2a1EA8w4y;mcD zdS;X<+a@U)>CwttE19@GV=V6I-`{(%S0w<9G_l{seF7Vs-H%OFTRi`Bs=Pv8CFSWY z`R)MdsBE=IQ|TZR+;AVBTP+TbttCM{w`9dkvJcILX|^j&45>?KTJ#!QjY8(NeEH?! z2Fg$F#5N2X&zTv{){5;Y6~ks(A$We$aez4Z8{IBCDhKfHc#PS-z66~+)r{6KyeF6G z38rw9N8=~r^cT+QCC1W1!zDZfDLy?>Pc?*Dfd%C{aIx8TJR#}J4x$9xg~iN#q>HlW z#(~==QWjWR2u{xUT`;Cd60A%uJcjeF@$C^JVDb?7={Z@=kntaTzvO$7NqN|Y02!h6 z*V09xl`SFfQVa-(>3$vy?F(uLfM&5F*)=w^!A|gr?uHugx2rO2TiFg+{Z>|1Tn_l4 z*Y+i`8i@;vQU+kw&kpp(njtu^b)IBEwUIE%sPSSY%9D=()K*?x2?}wzM3-)dPjnsA zKDj=@&L-e8xHr&~GTlfA8FLXXYF&=zv4}JMQ?NFLahmP9obdjF=8E`hGQ0`&bhBu8 zc-2D0NX!Vh->n1V;im~0Ah%NCwPNonc!Nw36sox%h+jj@`?qvg<~4>J*o-R4Z-~Rd z>2LJc=RqO@sh@~r?$0e}gT*ia*Zdk;LCzHhxSXWP`t(TyPDybc5WstQ>GCu)?U-G8 zF7{4asZs{SJjwknqy16>YCT;xv`_YX9OGz51Aq{$s!i*PG{IprlM?^|52YE7`BZ4u zY}&%rf?aOS)AzG<@UG4|$1Z3`I%sAM{~7-?V4Xd#4|>B1*zZo7PbFn#MT=+XND?qb zmg+K&Qz{)S?Sz_`g8>i1qo+Owq$KL2GnuE7Q#|^(HO_CH^+B?nPvz6^CMtQW-vX|i zvikO~j!hEg&JTG6h4OH6_IBC*XhPy#W)L2ec5r4Undj}v_IW=U4GRmVgoK30)m~?f z#guxJGcBN$q=RS!Fl!ABvJH*x>3w|5n3>UUdM9RM0axaBmr5bR3O$*jaDTAed< z;C54>&UVSXxo+fPVi55lX?Pv15N^8e2Vdf!B2GZ3`?2yt>P-d2G1HBgs4y2r~MiN%o?@aV%3iSkF zsoa*Qua)e?V6oC#)_gj0ViK5|w@$>3y*PJ(%bg7#lMbS^UElZg+MB4OHBsXjQOEi! z79Wn#qAacl5_-AY^g8dw6BQR%QdDfMjb%Wg2rwT2{O|2Uz5ZhLdJAV)W`0BoRQ&0yb0e2__d0|g46r#xZA&;2 z?k)RG4A77b(5sV$>Y%PJ5hW!h^*U<;z(H?JLS$MV4vMc%mWco<_w@W6kFJbJeh`sc zic@`ZJW!?qW@oD5q{Fce6;^n~aJ}qf?r%`>5ZlL;1I8<`K&?-yXo58>jtIl`7MM^D zsSo!}=aNxepb<@+@!nWu;JG^Dt=I#4_jh+|o15JMNZ6LP!vg~!*w~azVNl3NL{QAM za3fKS9qsIP?4(Q;ESUg_W9Txik-xH>0kyTYu_q%$8WmK@*t}F6;F|SxRv5Ol0nt{T zo+$r)JgL@iU13((caDaIPmG5hfFQE#TU<YK#S%gEKeyWNsTy3qds$sR zv)UJRaK}NZqjUoNg$AAzZmBAD=+jhNjX;~AJ#yj7?i?hhwhAVMuew(067U8k({yz- zI{H9+AYa*}0C>ybb=s*T^*u=|!JrIaA@lZ1xjejDY2jvyP8h38qi+}i)WLcA_^rd5 zXvX9^K}61lbcUWztI_mVu=~4-iSO&dJ?RHaL&*TZgVsj%VtLf;HP(Y=qZIM{3=9mQ zg|EcCj$46Iv;a#M2)| zX13Dej!MXZNKa4SQMwjPltAwg5ZWtbO{?{P%t?liY zgoL*p{un}%Dk|@x71SFFh~$3x=v*zVIqaB^QN#aJ3$U1pW)|&>Qp2b2&Jh>t9xMYZ0FdEimd+Q0_)$jlEAomPB~QN(KpPh@L^lAxAKjZ ze=k-=miU+|h4TFEV-zKLRAQ^hYh}eNEr}f-1Q5BwjvspX=|4i5>-1U}e9x1%2Cg4C z5@Ub!iv+bY(G%e_i9SWB_TMpMymNVY2MFA1XQg>YGk8Fm#MFacW?r5j&u?z}rlznJ z6cm=0mXg@b(8a{W0FN4gDPCNhh+_3R6$c<;ww$u(g*2U;g$}I0(8{o#6dvJGLyh$?lKzwm|U@id2WXC8(pz{h@P{h zrmX4k7EqQ~=5s3R!J^()rw0S!p8w#^X3P5r6MHA@27!mg(I{N32a&yd_wHwVyD*0E z6N^L?Db=S>!36B4pGYz=Meqf!WKc+tw@1qsz?a6WuIgHv_B<4szTS120BA%fud>w5 zJ{?728*jyL0A3{Qt5-a_1pWZ_`tg~~x}?uO&dZ}!?BqJWU>q&z`RQq3Z0u#xhAuvo zd(|sW&*vU8)~{1zjs?gkz#6Kp&;z zsLy+_6|}2)oDB@f$HvB7_;wV8pL^KH1-h3Sod9o_t&k~%N-A(yhnIybf~Y+Gr;>zh z39qMo%5$n(N4dB|&+P1MIH_PtR@Mg&LNwv0GdRHBvfZxwzS@vg z1M(ywAfUqSd?RJ?>FEid*OBsacf5W3^1qa}QKl5_h`)S+k(wnF3Kvb>#qP!gDU<6+ z+g##03}ZQe@#MTC_`5N6wOYTCU8!mRNL!7iwOMphfCDXEdEV<|^B`@RK9Kh#4gI_S z`UlLi9GKrct1T=n^z`&#;^9eHTQjWsJTh8XSO8uOmqEQFUoI695%I%^4}M?#-!5Oc zsg-Gy)6s?E>AJ1k)ngFzM8hGVYv1zX;NYmJs7yaGu}A+ZieAF0&rvFBXwIcy7aG+d zi64AiR302@9|^}{eX_IT17wG&IZ=0cuG->KtBS|3d#zX)xBT{~_r9hhgR`Lk;~74L z;tMUh|HS%o0_X|J{fh9&L1&oerY6VJ)?4V!c-3C_j~IXx~O+uGXBg>$x8zC7LY@bk|-xw*M5ogcpI!lVR?+5Gp4 z2EfE4#D9N?&suJb;W~aS5RGxOOl=mHFpbso_!Vgz3CQWi4n<@mc7`mf)G%Z5 zqXI2Sq#|@X&3kxgCtw)Mvy}< z;;1(L-PQqKm{e%=%15Y>@nr?gS{AxT2wmvrk~_w+D0<&VBUm4hPTMFjnLElEmLY>z zf#d>ns0EPwx8j+(nw#ydE$t~fxA=$%G%hZ#Ktu(^*4QN+mzkxBK(hz%x1*=pQJr|{ z-o5DAx51K0p>?Aq=8H>T_F2Lc7I6K5^aCISde~k(g7CZ3BftyztcIeckq))zDt*c! z{0^MHs0;Yajr3Pv9MzKOs2u0!lWbrRS)6ucspZoo0S?)p$VHGzU|1hXW*4~MimUbT zaX8Op-!>MA|Lec8vBAe|p_1tIW2@{h2DEGjHMzO3&$=I^TMG!$74KUe=-$WfSo-~0 zHwF?0M){0WttO*}h6WEGpKi5i*Ng*Qa&odMe;a7ayWzm}6v+47h6a2v_%>xb@*s8V zq;&A{o1Ej^)xhJo(Rpa&ykXS!jQuP)-O2tWPd@5PSI*M>zrPMX+B@xN-2A6fi1q(g z3O!syZz$jYKuA}sDJLrWCLzI8a87(BNcjKrTa4J={!b+mUd8_}B~CI%t1$no{)i6J zjYTb8Fr}!i9S*QJE^c5Wb=(rp!-qzC+4y&jR@Ql-7C0bQKpKA{2{iw}#?EfB+*JEz zj0zYoDd=#0-6^Y#fjvV9q6s^Lw6(RBXV1~4@}D{)#W;n)5C3~yC1KnU%3Ov2r$i|5 zXbi}s2g=4pd%&2-j^Q~x7!r(<*dUr8hnL1{?pGW{@M2Iz|EI@CGQ*qgsj-q{Ol6Iu z3uB>pZyz1?)G!zrM(WiqgGZdUa-n^;WLVDZ?6E_SOVbH&txfq5)dz5I{+S~cbD5m} z&4C+-Cg3uW1=)&lJBEg&tE#JM>FIObFLxRm8a|Q4LgP8gV-~c#N6>#l=>wX-)u60DsK@?7IhV8w*C@ z(DI)@bkLVbw6wNXR#)fREY`K$@1&a7YPP0-O~U)SeUc&Qi7({Eb8>P5d~mj0&v&^$ zogW+tU8M%#)6mXN;oCP%MMXuAuuTv2w2{r)h<;F#LlnkoI8s(+D<>$y5``4QZdl?L z-XQzucF=eQ`HgmK9|dZEAxa0p!kEm=%+T9S9ei8 zd;MVG@Hxye0nnBX(x%}{kDQ-Z*Q~Q1OyRV8an<|p5**=OKtyO%86&~Kz<3DPv5*Ch zEQ43vgP7ayevS)0hHv&qT>|#I%6?r$==n+++Bb}^JGK@Uv_F3QfEpNXuMvn$B6KZM zRaMn`q3Rc)D50+jBILF$spokqDr||}^@;qNJHtxXfHQt`+jx=}y#C`v_H-An17JsE z9PP6Dq=0p371a+og!x7%8l5KX_Q6K2dRy5Hfn;bqUY}2Wk5*cV$jHdxB;0o=Cnuo* z-}w`MBttN@s*3IC_!uTQG}Le)n%qNBPf7|2L+I`Uv=$5c!D|m|16=@V5HLugXDm@k z_}+tn$N|Iq#n?D+WsH0zh};Q$w$=$H_xvdL-E{ET*_rd;vJrNNR^agKe zy%3xz#1|N5K)hVc8%3>eZS^EDYJmz13)ki=jpkb4rE}YnqNAg8KQAZLj8@ycJ=#CN+v)t9*<263X%W!$Lk$j_ z_AV||JUmI|Kw3oVZjSbt;8t9iQ-e_Y;SKG)FK<4`)WX|64U@$qbM=`xT&ei9_EMvE z7yfry(|=YdsaRMrwY0P#UniUQ_RL{nVJqtDptr1+=(eV?8uek4ktvv%P-SOl&k5`f z(quyK#pV{YEmkcpYqZ%OiWilT;0|_(6LAO92yjo|)YXH_%b6`}2_wY27wy_d0ro4e zpQC`ryq;@vYD%NUjg{SOYzva8yD<<$F_Odz2Skq4=Wd;k5bRLz4Q`PQhRsTkwi7k%b|b8KkVNtWj%;)7#stuBo|kcqk(&DJiFj z|Ngybu6V?6^vg(5F|iqKn@B)k0ST|HsQ3kN8%#`0bH{J(9UK~uHU0wJ0zIx{z1r6L z1$VOR3^l$PB>a4Z4HaU;iOl`0tC_$A0sw$uX?fXwD@Nkv3WCUPG_X{P2H-S6e#yW% zf;u}piTPc_d&BS!TR}@P&OZjG5-Vz}pjO20uyN|~zJl+7A!zcF*MqKowl|kWNPHO{ zb*8!i83jd3O3GlN$^=>@I5Hxu>-+5TaJBDWUCo}(=c2?n!v78%TiL0W01N9!Z7t_w zoec#aU&`9rVN`XYjG8FW8(v;sfWHkbEJ7dL-7Yy{AR!_7HXlz&koC^Z zB?5_)rOLf@>=W<;GzeyBXs801Qk=z_r3ohbJvQvlC$r+^x~-B@Qum8CMMXt57E}51 z{mn5XQh1VldZH9ooKd$HS_MW=1$O*Cd>(JvWb! zp>KQnjSd;h$;koI99sCqS^vK(t}e%geV2DHrEiuy4uJ?qz+A%^`LAbbZ@$m3~ z|Alb_LcYIHO>?<3Y6Q$5VqV7pDw%`|b8E|S zK_A;A$)=7f{{lCs059L9kGZF?wtG3bx%*3v?6b44uTebzw_2+xkrVXmFT&T|lY&g2 zApm3C4~h!TE-#H{OEhGY*?vLqt1jQgLd(o{;$~xuU0YjADceqFH+x4+JW^cVrg@oE zQNc1>s?{HgL%#=nyV&X-GL#OwAE2j6N`NObvG3f# zw=g+6{0pFjhUv75Q|on6#2(xq;Uv-=J^P+wo|B)GJ*k?GRQoIaoCg&68`e5nu_!5iMQh&JkvGY~h}@Dl9zgo;3Ml)Es|FB#=Uv%YW?s zqP^#Q5r8*|2h3yta$A{);7!us+0WNE_JpOfv@BHATA-8pO8?l{cOYnuEVhDx!ziIN zpwvo)=3|C}+P|M#3FPW|wz++JZ^mD&B{V)Wm=Le9BqqTv;TT4LcMFzi^b*5$z~ioa zIzPc3XE*z5?dqL4q&c-7GDe0lWT#$2=_JZU%Y&B_Pl?@0|?M;~*> z-|Jgs8*0zZdI+~gKvKwcG+`^qsbx<4@vi%gQ96nN**1Q)Kjpsm1K|^WTp1>LfP)s9 zQzUfv^lU?4?`65!mYq$DZ}EJ>`%9t9nX)nUVwA}tyX?g05+3c=_$5Di+XlYbandu2 zpMSu@3$bn)^xv!4ql#D-VE-;JrvpG!<+v>krF${4Uq>eN0W6k^1?ZXW zEV>8yBwNEYcwAcoMy}Vc{aHMYT5e!`d>8sULp}!bfQGW8F2{T?vhVEco4uVr9=s>) zx6U0?y%|S2jzSNNa?Jhj(You57T)WY^(-$W2nmfI9t&8a;P~^`gZ|K@Ax`O6!DqF$ z3Ic#E7KVGi3A*i#u2%VanFsP2kL|Kul)gd2R%R#H0f3_JV{08Y>ujl(|09UBsxf^f77HRBEj`TmWVkXyHU?H zf6TTc4>f(4+tplbrrxMY5QRdB=8cn%ME(M5kP=$kb&G^ejfrV*=1BxNvbJN{vGqf5 z0uTc+Ua(rV__bq3V&GjBH)En4(r&VA8)ADx)#n;sEx=j$WgMq$tQ+c`?p9ZPV=Z(- zNgxpzt?NMnre5)_Qz*Yt6{8=ZsC)YQW}is$@FZ?emZ8Nj8p!zRYj8n*7`iEk| zE)=Y+F@OJl^YZqFmV$F|{NCH!TjZO4y1zt1LFt~Dz*19F6OSaO4Uj6qzGW!$2_E2HGK=V~eAUzo>6iwdsnt z28z7=)77RU23=qPy4AGn|M#EDd@2TpaKQZRjHM5E{D71Cud^)=`}tTRNkL@jnFe4> zCawfrFGpXNI@JRYVKW47iMRLl`Z9)GddG6Rtqwn~edG>(E!&$o+H`XFa zYS?s~d~N)0Ls`3J7_rUuW=>u1Olp?qizQg(84+kcFS^I626m-8UIx5KyU1uhK8l_I zkhWcp@e>Pr{D%4;KpfcLof$x1v$?yu@62a9mL_$&(h5-jr3)Fm?~4~;XRUk$1WNJoZ(6OP2XbhPT*^bf5??xujtlA0T7>(74|b+MU_ z;Dct%bcVYzgtZHd@$dH$R72 z3k>dRE7pG%0bFi9lN2%VyQL_04&lIor5)>U%9x%f7Yle9(`<0UC~|V?RV{HWc&YMs z6aLmUc9&*SqK*L&2FTdI?Y_Ry;;@d5-Y_IIwANB1K;1#hqnlg~=WDI-R#sN7u`K;_ zr}l$|lsn4LHK3dP$@htNcj_oOd!Y2P2b2pSF97q&o8(iu!zxsUj=HziclH^q(W07E z@_H;ovm*-=$(6u>KP47(Qs3TGo{B*Bnjt=t469TXwm;uxkF6N6Zczik1MeePyRZtr z9Uwz+tjwfnUA@lL^UQRSvPPR1FAYi#OAI9`E+u7iRu{w^r`o4dYu<7y= zV(|$-tE9(pw5bK_1M)7#L^In&yBb9%9i=6QX$>aI;_jOH&;RP!;1f&_WewvYI?zly z$+*3am#roS-A_zv+>c|ZNobv%JDEUhBAhwm)PlfDe3Xh&H|YZu#TCS6{Pb^~m3;6b zf4nfxrzZ3Rj3v9jD8FytRNW`fXA$h z`dL)8XVq4($(a`5$0q+^qJ)-MLZ|==Zr)$Xr;A6{ zqUU5xo2jv{qpFdMOL;^drdxD~#re2Piyt}OYetDV}>WkV}_vfWD|*lJ%>`{#D< zlc)|b1Q~iH3P4BzHhq8m@LYgokYUa5`lFWBESbH5f-AvMapv2zSB;1D*1!j_e+jQU zJ>1ihn~C>F)@1mt_wkztdD@9J4MKp|#UK?JI9u;V!ob)a8Va4BR)Mb0KyU8x zx!=b7`Sa(2v$aO<-i|cgAN8t7O12M0A%wI4}8^O^qb zqE><0wD5n*E}++vmT7FB0zAJ1T}zm+Fn}J9X{cPE-(k18?QFh zyci4P<`7(NSnXpcYLjiKVowj4tHG4LY9$cYH{kn~iY6u+VeS?9%j&A zWQ9(O31&E_y}#rc9tTBMw)M5OCi{~RHxG|l=j~Ab#0;&j0IDE2+wagC12@jv-LN1L zb>UjPXRl?{$tHuMI(^Y^^sWMniQxV>fHv#i4>N%tAf#!J+W2{7`yt%#W?Z9vJ6_YH z&T0s5Ss)&Lzp3wums}rgX*&v^Zq5Vrp8oadrImT4!SU?Ckp)41OVO=SY2Ca>7+Ot` zni`gvIOlSzT%RwtjM zC5$*UNdV%cC5il7T{xA<6QleSm={=IPNah_xsOx$*ESD6$E)^_Z4pDdv0nAB;IqVY zJ7;hD{3X7A{VF9d&!sPlHh$kVEO2&B{?fefHv0MIcs%ejGuTy}^<=EGFBUK8?)>m- z!vT#4Ei#x|J6UXngsLTLxfa`USAFjK*dv(`?ZEa;G=|Qu-k}&gs8H63* z`We75I5$2=v;+vFHl9y-y8K^_Iwj6#YP32XrDo3OSOF@0tj1XM2FH z0H3ZvuN;RS38JB)frW)l;;?v^>2-kw*uR2?2Kk%wmf^>srDcSwQ`ZKHG$#Y@*7 zN{685%{O&*Noi?hAbduDv%VOcnnG{tnkrBr<8wVKbPlH8HI#zKegy|GgA3qT8WXvT zGmk>R)#At!s-7^VM3;x*LZ}$d1=7Wv?WvLMHe%u02B97<48_qHU=`-Or-L(jaqh!g z)3Qj6MNF_fQvxBVR25p;4tEByK=_S_C`5@h&*}QeNqIG~DVz1dM8E9f{JgWjKM0WN z(~Bb*qW@D1u#^xJ^TpVh5?Z$;ZmJoX$-Gf{`jJ=7K3RO$a)pG*yh`RP&%FA@V8Dl_ zI*Tr?cIryn7Xcb4>AQ!RvKm`MFP-_Wns_x!Xt5R`&j9IZRMNF*vt0HgT~pt#l$!mO z4S?MMH?GF(xPXa&=_$iSVJ1BPG%w0OZ>Bt*EBH6pYz+(ULot7wo^WA%P!iu8Wo^{W z)05qHiDzdtwZQokwSmMWJk+c={v3v$P*|SUrPg}VfhZNZ~wdsf4GWR1psG+>LUk+@@_V!0< zI<#`J0DC}$zlxsFH6Q8@3?|{T3Q5w-JVFv_+M@D~LIVL+S>9e&;`P)a&tWqGG`9=brmubwY0V#I>$ecu#dAC^@V!`4_uko z@U7f!;}4^&M&Y<;HPb*>oatSHPM82DqESd7F_&6wlQGXd9CRx6Gl zg(9{sP)~>MCneb0(41S`g*Br?mbt^Tr~$3t&b|t}vRswjj}|5IYOQa&E^p5y%8-*Y zY8XBPf!mgJmeI5|6-};66LvZZUtw2$9&OSyz`4J2N96#gOh9iqpWzG>kE~00)NkZ{ zpsJFJ>}e&_=Jx6LeQMO@bsOr+a!OaDOrpD!sM$0w+O132L-xc0&`{KSh}szd4=DcL z076HqYiY9N$)Knb4{jwkFwlbJzR5fATICkcIiAiU8(R%xyq>)H9xc0Tt6V=IoQ*S#RG@W#{P005 zH1xwzPM{}KQS!D=>C#1kMWd;!MYGBB<<7zFzq(|H+x!byThmVy@+Oo?VCeLNf&1_-1e1g$0*78z zzdw>>U|>L>Up6jAnYuf?L3Hc)_ih>MFn?zB-nGw{jWW~qA=$cjC$E7Wg9%Y#?(s>_ zGrYeg_acI7nCn~KTHO@Dfp!EB>@J3_o!>+aPmH}3gedKpNEk91Ux;6wzc=uzpjljZ zR{zY6AbQ99wfEaeGAz(0J&%1V>(w?A3k!?rL1TUVx6oRTuI}!bw6yT;;Y5m$AAcUR z$cbqBP?_5D1{q2TdNlE`{SEudXqj(khKBCk>zgRn{gqNwwFs5Hp7LXTOW<)ua7gSd zxBlRX@xTEQ*8b^FUx9_Fy5=$Nso{bV*Z$;r4dHk@-H%cjx zC~Wq5P1XL45?ej8h0^he^y>hEXF8q{F+Z=KChSYfmK=`z^Q}GKBU8qr56t&kH2k&0 zaJb>XuElUyT9%r&<#hW!vFjTLFx?@GmCVPo_miZPd5z)_-kAd1gZ(yZL|i4!Jdi7VzJ#gN$ZIUeY7_ zH^~_o!jD7jdC zyGs-_7C=l-Uz~Ahs!v1)b`2MwKDqHMRFNcnvF9O{MN+PZRZ>gaZ>Ep4!t>F+bt9|R zo%n2RmdmJrOM_EE9O1(L2>bWnabt88j;HcOmN83XJ6;`gaL#r7zll-B;8tdV(9R!a z(St;lYpIkydJtSp*dD5-p88!o|8AI<=lUOrhwCG*l5ek7rdB$TG{B1Ai}iU4Co_C| zS3j4Egut|oC1D!+7`>8Sjpz%54pK3sNaRM0!UpmDpc$ILs1==-hT5o&4ys{5oi4>m zRoGdJph)awIe^b-ccUg2X}f1uHZYg3?ksybD&|B!qlTDtHOc7J#0 zh6`F7rTNJ^7~VYZJX7C{{osoCC0k|PH&JFDJ*bAfTJoni-X#pkP7b{;c)b0SFW%KL zlZcLg?24y`@ddoTsw%Fbp&^UgnSQY70TLR36-;eU7cziC |R*hY*{fEb>wjH^}d zt6d|^4zaeLb*gCYHBeaFxXrV1!?J($_*`UocF-wvN4HMBXn=+Ll%yLxhpzV)#6UEI z9VU@9zR{!u(MZ?~su4+`ATL!c!f&}WqQ-yg#x{cDKVrVzegg~@12G{8H1EO`RYzqU z0)(ePKK-ZY$1gzRTu)abTn=ZmtE%4FwLfzOVvwO%s)t8MW7c(Z5Q^m@12fa8e?b7$ zv8Zs!#6{X1iQ&_TjMGY>5kWSEa(?mRY(wuc(j#OY zWUQiqcJ($0j&l2ZR#w-8ZNA*{O4z+*n+}pE3`V&y)`%;Q$Mw}`GM>f7tgIaLb`vEO z4x$m$3iWT_Fah_)Vm3y6Y}DA;C@C*rm!i>8_3*akPKR!6>$m^-_F9|<982v!dWJK4 z;&5kb++%iXE+$cwPJi80IhZY5nq ze$?arz-r!G`%l;Cv*CxMu9A6D2H#191YG|QXMX`ySI};YqH{rj1a}s03Be^ngF7J* zT!ORE;1VRbOMn0g!QBb&E87_I0POtBe4TzN(B*8Rr<5ijYh(AwVd zKgLF)!na@Q8f=i-A+F~9Btry7e312V#tK_9>-<9-YdFmY8?)s#NDr}#GWby!(h zSvh_g(qu3_DPwVYu@->No?!l@v_<2(Fc6G<`S}T#0cGn z>oYK?g&AWf){|_vDi0B`28xx%0twugmF1Ieg0pi+HM)!Yt#U74%${3llHQ-)U;!Q1 z;NM4w{5Qfdt+#*5(%}Dd%v1hE{gSmBuTLF}jr;x_PgoB^A(bC5Jv}{9=;Moc)^`El zShNra3yoo6VSb^ZXs!nf;fg7|#v8xB*@TwxCEI;}w#<;DnblokV8?&DrhM5nS52AR zgK&yPEYQ0~5ShhpskFC~@^xW6-`be0UgbU0z2-?I`ST6KK5Jjq9i>=AaHVL`XDmERt(E%_&w$zVJY9`a)> zlFsXBfG4hh0^_HsC@a84aML*zLj1;lD0(|syOo2IijMYepHTjI_P@C1YsXfpPah4R zua+r%Zd;U|Y3H8XvAY~GO)xo_Ch1rOmLHaJG505jGN-G%g?1Rf+QC|uXR9sn;PA(y zqN0~qS3kzb2|(c!BcEpznoq(^6u4htZ#=gt^(TJ1L)Rwa?oCe7r3oomTaq1HkB2%L z!}e`r9AyKZxA8}H6}cIiU%b`yH-@JUG4<8VNR=!le14Hc;lD^x;nttyZXy4e$ag=H za3}cSZ5>#R9BHnw=U26(M9UHo<_iM0#}{{Nw3=GCu#TflJ9qO64mhfIEeF;!e6mkQ zmgZ^zyM7ZPEiG+>`!OB(V-T3CtKH!QVUdwiA3qArue$UPHBTm1p9w8JY>*Mk6M@jR zmY_k8ySmlsYMI-g*p_~|yn`3^&rQ6x1<(2+#h;{2F^i9hV60Gm_3|5HaKQHm`guhWtVXUzVkMB}J@Ru^Qv?JqS^j7FBig zZp#?aokJ==T`fyK-t^->;)`na@!67=vwF|OF8Lyb-AGVSP_OC36LE3zxjwK{_hze& zO-uqoS^QzArLwYersh4tr-%qVeEb04hnr* zmDQpCyQ4uP!{Ur2%kO00m~CD+gpVkh=7{%qsQm2PsT9}d=_vx?Vjmy5TcSpWm$n(PKPW9{R z%&2{3aVt6N)3M7ngA)OkzAXZc)NCR+G!F)B`F+o_TXa5GP$6tGs8Mj9-#t*Kcv5W# z&Jz0J;o)80o{Z=cT1=Pecdh&gr2X3e%4_)U=E7vDR8NI**U_Ywq4CZ6euKvu z^X)Cw(%IbXqwgBr2zyo7+WnyT$66$S^LxGEQU;lPwbxL0ErJOGs%;bvz>kXERhx(R zu^QU-hfBp&5h<7F#r9(V_T^KD*!!s8nKii7%iO%wLW8AK?uA~B-J@Q0R1kZz|5*p= z%_rHz}hB1wv{%N`O#8g_Q<_WNt=>b1SwtCK_lXNH4??aRb( zw2~n-baY5y<(lgY2n^(M*-^JyXgJOqIg~@U&(rL`3U*5p7Wy$a7U+?gjj zzWPjV^9)7S9UwmbMrU^9EueVaNMxz}{p&QCK5KaX+Cpo1{OCjGvYI)+*vfyb>618C zZ!_3`pmqbEV%`o0~hi z^>f4N=Z3F$M?DYnm8mA_l0fq8nxif`JJs&Ge^RV4XH#@&kKLh(?H-C3E9E4?2o6Sf zdD1ifL?^UxzJXY`a^jB{$@b6cy?qFK&tCdmB#`rP_o*qNy75)IdI{w+d2uhQ@NK_GzV(3t)0dUJ zL$xTP__toE`=iUUk)R}kFZNMJ|D9nGo1h5tY7x$(CdNGLW4F+N2WIhsnKd zLy@TxkVWdZ+lAPJz%O3C$5%o#814?Z|5i;ckSo)$7APJ$kUUx4HpO_ZD}ohaq>|gt z;-mi(|AvEjMR|u(0mH&#Uqjzn%jWrC^@Fy{XHe-^)zV?OMp|thhX_0zj^s+lnr(+ ztqBwx88*Lb2HeNEq@>}OmzU-!MqXY*=|0!nni=$j=(L3QF3+=y(%8(ntU9q3@8SOp zBApW8NYHzGf9k7Q)S}P;CHl?8C!0f(>gt5qIXRUsyPAKtwoGQL%w7o_R-Z)gQ&mO(06XpPW4?2bnY_v!}(Q3_{P}Squ2bfDC&>!LOPWDNgl`oi7QGjM=sSeO(HAsQ6I z8L$r|QH)3kiHLqoPbchih6OV+GKxi02}e4y{nF1?ifiuHBYo+nh(+Y5PlF#06cYUG z?+@Rw3~T%%pLcY2Ho5tgOZfobJ9FXL=BKCy9v*eT)j$8{2l%3;Fu4TMYQYGnguet( ze@_WTcDxd7gPs;6U!6<)+V+Kc_cICA`3_h9?nf&BX`yx^rG@jcwKru8c{IP()1&C@?5sg9%*n}d+?%O1 z?*o|uULYz~fZWpnP=4jVP5V?*!rIy+#u`!SL;|2=U_eK_HZ(T=03j3^4O>cHUcR^w zDo+&HXx39DW7e46PH+}CrLC~5Sy8u zU5%Bc`r&!liuO%@a<3s*&jn}p(kC50J5LEX+)oK}V!9xIZBGP6bYf!I*XK%f0s^Ga z)Ca@c)Uf-zn;flj1GtboCo!8&5T&4NW=##DlJFTJK0f{ocB2|RMBKyURZB~Y!{5;? zbI~`7iqcY2K}%(mIkGs=O-OQY-^w^Sy@0}3kl{m?l$11Q0=~7v>&gk51Ux=IwgjJb zJzDg0axz(Kk2q@}mvenp6@FI;R$x>V)@Mq=kYcU!wQ(iU;Oc7U2ony2matOYdOB8C z1^eme@jZ!+x@u}5-lMJzaJ1W;!1b5hM6dtS+ssy>82Vg2E7NZdi59(wr|{Z#`re;p zn73=GrZ)i;_5KR;jQrrIAF?S*~Iv5c0aZ9+r@Mz#0NIj^sQre-f-1hxs2EVTBw7FDIAqRFP6 zf&%7CugjxQs^?$(6Eibu7(k^<%40=X?Qv!Wc5`yh!ahGR7)vip<)LSN*bIi<91QTp zUc>fy!As|@;YptieP**;6Cd;ImNQYArA$+Ji1DVK=*VT??Vy+QW_-F2#z~*dK{6RA zi0-`BgZWFz7b6ww3P32z~*9_symoP697pzSM0cSgbHcrl6#p zzkB6<{h@hUgZbM`y*;&+l@%@}<>ICze)uC`WoKtDc=?Mqoyq0_h?96kM8WT>&3|-t zA+N5kf}=qz-s zajMxsyz%+o?0))mUthswABSOEPbBa0qt50WE83Y1!v1 zfEWYeq~2~(0vxGRW%?0(OYX?8U%%F_wIT|VpxqqFzyY8rsj8{HN5DTanzSFfO6eXp zLvmT}>EHQ!4tF6ckY0>{@AY48EK*0V^01fQoXf!7p=F{E+tU>AsUqw^O1Q%zbBq z2@rF2`9h`0^#Ybv_kj*ZE2r*o}Ih z!520(HPO+~WcE1g>FWzF(oGjAQRfyEPg7*JFE_%mx8B91Qkl}m(F=Ab_OfNh~HhS$( z|5wfCcf9X2xSXd428w$Wp8xN-n{9SGr1xI^mf}sy7IiKC5%%!{aP5m>eXc`>Z z{gdu&G<% z9ScPz!7VB%unY`CtsC>cvva~M1qNBy_dK0ESTxfuxvyg>By&G$Yiop_6v;&PeX{eo1`4Xn0i&@L#*;I%hf9UB)H1lDQ9?o=rNKol(Xdte~) z!oq^(E%W={nMPfG{ixJbiv0^{Hd{$4w78h%_VyO|d)_8!_KVQ#=m*--$%&+#+~eQB ze?vz_V?+?MApxd*Y$w0R7x9GH3*KRhr+?=Es`ix>|nNZ1l-rxSB3_c^WflMaHi5E2~7IZSuQ~5Zm!ZK)WC$+ ziY`OX^1rcmCgyq@J$f75E7D-q*=;8Pxt2&W}95BoBPIJaE3@X7N>xfyY4!@ z$>;~tJ{Q<4@_DM}XFHRT85z_j*46{iojZdm_HF&3)Cw_%Gn|`ZuT!7Ataeheih#F? z`lcL5UnM17oT0Sd+iiY|_cw-Pca8pDNqhD`W)ZtVf}~IxD;w7=Ma}4@I578+zwTf~ zYe2e8&eDgR{2*Ya zNFbJMez&A_qFtFw_4^t%r&n6xt2CkCveyX~$XvoTiZv@Kp-rTS513t+7##)|L%enbSYKaP(JyVA508rK!Qf~aMT9?ws4W@zG7R9zwQkZ5c0nVm6`F5S zgc6hJ2y3uge8vCj)!OaV3FK&$iKZ$|Lsx=0B7vgH*5TawRN=kdxnkcG*)s;j++ZCr z0g?^s3a1Tup_{$x!qU842Cx<0+|p_JKdqzcfRfA+c|@-X(AwGx9qR!t+Cpoj0~W zJ3D(}VPVJf46jaCZIICt`PLFTx3~H2Raq?*s~A}zs0@T-U^?>O5(94Yvxm#m*jN1W z*DBPRWxbJk614@Z*?|51{VdfYTypXVErX9J;o;$MQ6HgsV>}5~Ww4AsT&|)6Mmu8m z-6A9~aw@dGvN$7qRoK879@3xfQItI!4!^IZO+BwNn*2-CXNmpZB(V2fEc-@_2cpt4 zGoypgC>+nO|LZ>wcbXwur9xL4no%q)l8P7d{MvJM@`*P9JbY@0eWb6=$32&`BBv=* z<2YC_fEC_1HU-B_Aw4x0bl+HLbZ+pvsye?BaNeY0WsU3!drTW&s5urUO!5x~Jh)d; z+t|7Tep{g^Ybh^3GR97=7epckBlmuoCw6~YEK??NKA@xBL)rH&xoMtVcghOV0m859 zfRNx|T47=R8*dpI)DTS4PJ~nt?weC*87|Yas3YgvzL_?vxjtKWy+P_79N_X=fjtW4 zY020yU}pmRd-I++JRd10Q6FO^Q#`O^)qy5q01s53nKWl1)3dX)a84rtjFJTFkQ<2| zGLTNWhQ^*SzGpFlecna#yvngu`_mvDlZ1D-O^{qkYJTF0r`!dH7z)i*#4M2?w3}1# z5h%SHKGu^Z9LO1Uts7~w=(h}?BMD9QAwx5x;Fto%mnNYldIxq5K&%Y2JR2;-m|0#n zym>~5?zsK?88pcNLg~@P#q^aBI84w&yWXEqwtBiD%e;Ne#n)FJ2LK`Gd#HS}#>IU? zW6NOvw-5A8U@2(FF8d>mYN-L5Q{s8qUbW`tT7VT;OVL1{(g#iGfT)M)4#NX0vtj5k zG-159hS;CJ`zrsKR!R+ci)nka>7ZbU^Y}^+8hwLwMU@itQpNNrnF%w*LhNXtWQcy5 zI!T7gC?w9hz^uJa$7G=ibv->j0QmOp8(i3fdvh=;=S1MrOQiN1PeMmSStVj17~esl z1Jnz)R9iHE)+|Uw+F@`D;0l#!Ukj#_4 zl0sw;)}bipss0!nQ@Lr)$)Q(g)(9Pj)^qIb?LqWJfT-KxvK!yl^)x84h+tYC2B^Tw z{!YnUl=mj)&;BT|Ti|im3tA{I{Ep3xiu|74WziZ#?J?{EsooQh2f!Y?tpA+bOWOXk{fm z_ueh!Bz0&GgH_E0fv7MWAO>OZ$M7(=qobocNGM!ESXph(29l-ys2D z8PiG`nQd^S??zOf4@^(j5459=VRdm{(w5>n*QHB#i>`BRpNwDrnP#2%Nj{4`#0ox^ zf_EVPmfa#U5V^B{Kf2x=VLMje{1_8C^;f>8n0NMqt)Y1E!D2g^ui~RDBq$)T$3Sds ztahsh4@B#J{`|?FHt^#I?C$15*45QDrLX1wdKcuZbL9a@bI22L(=)8Z3m5#UJ*B=Y zT_R{ARHu^AFn#bxkZ(rk8NL;sJN(z;=hweV#k>R2)89&*9usb#rm85G`aPMzD$$B- z-XE`7yK>D1KD^h#h-vaGO@2oRoeU&vYWnmM5^Qg850{GSGmgk5fFCYpKMfY#2G@Nu zP%q92$uJUj_V*Xfl5nxU_eS3INeg-@uu)OHYP^5B#NUK~Lw<)N$GWbxU#|ui(}Q%~ zif6GWb5#ypBlE+^C+GFI?06LUj%HJxbg*Tj8Qta+mTR!|WYNWje11%z*eoiv0Yacm z&sRu7vaqmdxmfaaf4IBgaz8S<-4cBe2YG+>{>m%p+AlddIWapsc6Pl6zoeHI63pKa zKLSuZ?y=-7xG)1<9)`qZ#EDY(Z5@o>acf3lNoeC(C}3Uuc+iXz6Xkdwr!`RN?)GAb zP`*x*fp|^9Rlu+P(<^t?Zzt!#DXzM*EG5cH@RbME;Ll)n`iVIl3s$Ur7Y&PoG9 z#8eU}OQ+E%t1e@F9bT$;yo)S1bLNGtyF}s8c z)FSD;1RruFLo_@_n9rJ3}#+l{L#eFW$gott0Gpf<%;YAr8pS&@|(|9H>S>Y@C?)fKW zl&M34B?(PNLh~Q6B`}Lxbo|MJ@%4rc-!%XI)m}8fU!ICCW^MKlWo5D14G$N0bZ3O0 z0$a0GCm7n1-;4DLPT0lies^TXNTvbDo4^{uI<~&JRB2u0s(zm_<$Le-t=BM?UEphu zUF$fZGbP@lT5Hfmqwa-e3LCEou8Se@M2HKZBle!gL58QsxoS=W*2+(dLjC%r ztDo$}y~l(2$kKAG zv$OeA?Q*^(Gf!Sk<>lku{WM(c2!gBHjX#QEuLXYV_-PLOsSglz*&zhZ+FF^InB)`{ zrRC+(KT`>R`u?3cZTc@=zLL(eUqes|;HO1(IKGNwRGEgAPE8!f@z!*GK-7tosYY-E zZ5wy!Pn2=1&N8+tNwT9Kv{a@3y<;os)|;10EqSffnayu!i%?@ye5r8n2rhxR2gm!w zV{~*0Wo29l#k^Vh_(=Nu`{mxeG2wcV%`vaEmHoXA)+$87#kNqOwx!L@3e3MoZXiPF64JGLT9W3jffqnaW z18LQ*Ky}W<%Ygc{w!Hewu#ExMFJd~&k7Hw^?{P{5W*R4Y%lT>xQX}fLv8)bH9&98Oe z>pc5%(#x4mq4ESER3}1$}~#QSc6?D5uL^>e?zKl_nCJKcAojlzx3p z;wR{Kus3lsr^BC|m|s5fCtD^Ts(l!I_V9}a{)qvqOQ^CV!P(QQC12; zHqVJU^ryy1qZWNnVGtQ8p_+eV|4^Po?I=mDgjXeR#vh~S*F&kN8ls(W!grXZ($fLP z1(pU^exLyZZrs|mDyKs`yu@076!SMNA47XQYMwM1=?E-rh3X0ko`zu`xvseUK7alU z=e2=Dl}NwYb@kUb7Ud#!n^5V~Kf5m*!5+%Ccf+GPoUX0(y;yhP*XHB~fTR7Z`5I2{ z_tQl-E*FPkGp{<{8k84+%K0(EVgYr7)?_bLj?Mfs;*F7*gs zT&=5h)pVe5-?LWE^4jnDA92jC6;n-8z{vEl%BntVEl#R_}%@MAoVFm<)t_^ z;pR~!%dT}R-IgNq8K@Lq{(i#&E#HGfDo}v|8LYKe8wq$Iy#M)Q%4*m~frp1@6Z-m< z@y&L-&O_6ylX{+WBoTAjD`iT$qR+M?4bX;g)ry zm80Z##=2Ilxb<|Zs|3Ac>d0LvC^)#15_LPXLDU{fMPKSlah(V5+eHe_x^gtY=%R5R zwHces&{t25#KI)->?QJ@PtHvtMu*z$1~>M%zlyz-PeA(f!wG9hpo~zy$%`FfFxUn> z-7ElkYx78PM5PPCHQX!yqBF}M@I&rLiJ9VqpXzkCfYc#2x!mC1Okq4KKtvj33{U^$ zfKSkhO{?$HSOj41j-A<}-dre?_z9Rz7HV+2UU_wUfgRCc+K&&-aD(M7W?(YV;Qc%J z$$7v1J7wRd zeYR)g4z{*UrnN3+q$_(B2H9hIGX{T*BYSt`XdIT#jh)hQV&ZNbBH1H?*t&XYcz-=m_x!&m7CS?*$Y-#Fg`}jT zkf?8+uz(7L0*8Az)`kCk{Kx2PNgIp66gO)~KFYLXQD_YjF}@A@+Q1^?Ia(W^JRRUA%!0J0VT{5eM7(7Z+pg=Vkfi zgx1C&AklrKJt z8GOWA5^_QTD%L6uWYR9j$Uv^5`S}SFaLaO!`hBD3;9G5@3UgPv*5sS2K~L@u%jvY2 z+O2G>;k$VaubY3zyyah1PmOx~dpUiQtScq{&#SM_(dNavPyU~|9p6wrEw%4hTkMD6 z{?{@}aW85mB|dPbE#}MT8Q1^k^0c0{u8UUYe1lrR2yTRsbKB?D?j0SF408OD4*tTr zlMNgK+NiH>k-wZweC0lEw5XNIZN~ufm$5wF+JcN=z?`_+&(_wK%lrC+LWYRG;P6Oa zU+6}moF_;@-M5?7!O-V4e`VtpC4q^Ia3AnJ@?yWg`oOs9`!y&n{G+>-A7AI)DPMX> zV%dP((>XIg&CmO0r@QU^P0zua{Hg1`Tg7dq!I5({p9f~Q5asJ;`~=Xci;s(ndx!pww8|+Co!OQtIyR7GutHiBM1k zb@4#EQcqH<(+r0D46i79x`$4Mx}nUuYhR56H790;l_A+vmm2BRJhkue7RNs?)G?1v z#q5LJ$$bUl!K93Qyc2@@EYQvO7x(p9_(udB)${9wiEIYtnneZ5=>xBvu)?aP2Og;F z0`&qDv8$PY1a|#TF)_G$E$(>WkDd97DGqDhPo!jIvRhKk*^)kaPC5-P`oNx-C5Zpr zTH8elJZ&;65-Ig%e$-ax%2-&~HE=WjUb7iTR$d>EXj1LmoJsocNnDF8ZcCjKzsBCH`YGmG6j`eswhcw{ zVq630nD7_Dou7s8>FBYefhr3P@FL5Q{`!YCb(YZY$B!S3Oz+#0UWIl+i_W1LKG#L# zTTxL_P*2*}+Cr0c;{{4?JnF^HCw4rVd#`&I$~iLCnPrvS&?r0{8L7OxKSYQ|RZtDj z+}b>H*TcpxAXW(1X#|$yJ6_?a|16UD~?XON=T>4C6V-1a+DGGe6JhC zrD@)GeXee7&|~!f86OQQ43nc>HVQHT8Y%J{mw&X9pIY<}<^y^S4zvdDtHHz(l5d{k zhxMfGA{N?xMW;&M1@bL>5xjr@9?pe$#Gsfwv5Hby8k<`N6U`?aoOdTXikr8NZ08ATXjh9;)*nm(K$8GES6jw&RJ10VpOTL_D%rYS zc|9_Q24$$;sU9$EV49v`fu$xVGaEZe8b5WIT~^fq1&+GnD4>0>%~3*1{-`W`|3U9v zPRDP1wU{_97~nwigsZWg1Qn$rJ0X-_NJ@f#b~bUZZ8EO;x{DKYhb)eLSFN6Aeu2>ay0?e83VXt+M~1~wM=>zb*h^D>h%&b{X$pPxqNh(5Ia z&Bq}BttoFSfOL#J7OmdFPLpM>8%HLY(w7A$|TaBjcMnmCltp2!`{X z62@ioXcd?x!!CdXZ6sh&Nc4L`z+g#TTwdO{F^~|Op8h#95))eDCg-&g)6kfd9MLi` zz|K>BqNFrp9_{JjQ3tm!gbqV#xjAUY5q9&=LEo7C6iQ`*EhQn6N@=rlz!(%AA&j)R zhqt#8^w^i=tL~gMb&wZTeAG_3bQh8%NJVawG(uQ{@&Ck6TZyI;MSXwkhEqM+Q(%4Y zYlz=Xy3Xs8f1dWIT9BNmMI_dFP-3*Pcj3}OBGDwWu-?qT>c=ryBJLdJ%~$lJt*GV( z6}mu$r7#drz&9x178R@tmA*(J7V~CdW)6ss#=$2bz@wxr^}Y?n5$UL3_KDb?)~#FF z_IR%IE=TMgd8w>T)5F1Dc{9%$B}McA4E)L%2=*HZju{CCllD*3;LJp!I<%`zC-@pm zxky#Khr7Ml7fwwg$S+cHQNmQ4Ew@9aOjpDwhpC^{udle|B(UREyXt6c-4&P6(HS;r#7I|RDO_f+q{ zJ$2c{CcND1>ged`C9YmDvGPC}&vH-;gVOZq{CtYt&D1ik8ArA}kZhyuDr;JNXu@fF zLfac=$;jQI9`%3o5DujppgDh3S|F#NZB>ulJ52B0xe|r$~Dx1af zRI@;mN*DW>0y}iyAuI_bOD~czmzI`RoT8kSWE$U``a2~gvhac}Rff%=NJ~o}ot>Ep zVj^h4D2?XG;^gT4FCSp<;D31z!=&hdokvbiP6z=qL|LvXlb5%5?UgC!f30$dHzMf& z%XXN#?c?C&l+8!?+pKqQID6sQwN#_^5mL=}IE4>_sS3f@X< zYHC)SOqCnrV26(H6?-_v2&hJ2@s;YWQ%)*7!`%7Q{-@l0> zwZs=@=q;_SKUN)%x^P5m_@miqky!s^vg3Z8SX4(xK)B7>HO2AHq2?Mmb_fmYCR3kvALkZcX7 zbqOA}1ormIL{f`V5j=k@3dVd1iw;hdS7fex#Mwu32oD1nIgQA3wPun?M-}z## zCt{|_g{8Bz^TVg|sycD7wM4a)xBFqU2<;cnn+lu1^WNYHm$kK}&$fJg3Ek67FEz=9 zA07njkbC-tJ?%LQO9ZULW~n8qvcBHJ%F3!nlpO4alU{QB4?GjR-@bk00uqCQ9_^bP zwwz$UH0tnk%^)El=)$BHSuk>ZN8r7IM?TGMez5Z;;R3k>x{4#KulAfz?1NL z#>$VtB(s5d9?uf{<+efV^46)%Mx6ouB!%8bC@5TD+y)C(auO$twM4n|6D|t4?zu!? zLfW&f?Y~kD-~lMVIzy|TLM ze{|$vGhbI)%>m<5>m6PU-7vGLIpY&!B~LkQ#!+P?lzaX92iVz5d&hm$z7*l%;kF{> zwF_Wx3ot09M>?+eZhPG^Tfu7NOW2XC-H#DcQc^T8Uf_cr>{nXKmapJ5GHfa#=pKj_u&*pFEyX~R7O};uTZaiMV!MkL0E9q$zY!6U{rlAa>ESCpKcITte=B}K`ANZzzPUeVG$8sbG6pTU{BI8F+C~PEDca*?7qD^ z5jine9_Dp*ExXJyTFvEq!KjL zC?35R%;*Jlurx6`z+naoFk}!2R!*)UPaqu~U5)fA_^Z{Gl@72h#Qy)W^ffOaD*v~2 z@BR5@?(45L4`8~BX6NLLRa+?S&DCO9PnV%&E=9j&W=01I6#xW1R^lAPb|0G%OF~c+%{32=<;Y6X;I2#? z`bt|_F?V)#>3YA_YNq=iwmh6gw6XtMbGGNj>w~;`4=F1vE9S`Y;TtCOZag{z1E2r? zTm~wIY&H)9n_PBLsiLQ|Ys*VYN~DWvMm-Mlb8|CGN-$$%W5G~@UvhGC9VCWcFp&(L zJk{CFGnPYZSV@VkUrk%qkV*IkXLy=eKtKR4Az_e!^QI;Bz|fGilhcdezkh2-4j$Pp zHU)qG&b*(-m@(29$E4=)ca&DI(J61pks@vXlMx)1MXRhE#Q3AF3|~4<&Z8`@UmH$( z&8}GVGLfSqyf!lx!~-}B4UWiQk;f6HuEYfMAFVb%q3Q1+*aTLat%JZvNH98jdOT|C z=>O=0=;-LaaV62xySCN^NmnkXTjCwQI6u5Lv8!`)%2NO`qB9&~0Ng^K~AK=3cRD)-r z8eH~&+O#pZRh!UE7hBlFuZf0+*3j5kE7+!?x%_x&lX@}DMuo1_zK)%y(uj3fjMwuz zQRaTmx||CS4{wMH)C(3;ba=>qji40J>asxf8#-5aGksc*lFI{x<$VPj(hx3{-9 zzdqeoFgG`ka0V%%l7girBRwH{Qk(b93*CCVN$ZlSPQ8TRoUTy)eXf*l!_!W}s+}n; zlh6*!CjX_s`1eZ!W*27)pS}6*-5sgFSO-{%YP6FJa~g{WUT`+L6_Y+BW+eGM@b@ZQ zvid{zJOP)S5|;L|cbR5J@Q^OfNNKyPie#l;1(aw1GjT&QrK{tru^X`x}THb3vbFpw1?gFNxgKzVn~j1(wSvYKaz znd(_8&Fr2sgw!qE^?y;ON*XojxXvH=v<;yin5Y%@%Oa@)&Zv`^s6}I89f4-<>S=rqdOnu*473;Oa+GUJs@A>G<`H<#?}pB)hK8 zG%8{c-)_xsDf_7@Iqoj1QvQjWU~bgB$AmtqZDm(55W zK~6YQnVd`(_wAeUQp-nrd{I2<^e&?M$_t8SbwaYag`V^tQ@;d<{atUAz(W3Aqxh(c`W_eZ2E2XS8|`ySKbR^ozGdqd7IwS+ z*J;i0K`&s~jwC;?&&0zcV2&2=YVv^UAS&ubM`1$y7DdO;%zo?@aKJ-en3bQg^k5B> zkd#E`wOdnH@Di*CzjsnuYGf+VCiNJt^mM}qh z*lkhCdvS|;48Y+C5%p^f7fYdWKcy%4kOwUW#pvm%DcUHavSY&uTKSB5`UI?cT~%fS zA*X_<&VdBgvyry*b;b~R#`+faGiT4ddg~^T#QHd=N6nj^nf`Zx>xF~8LL3Y=5BaAw z4W9T{$_|j)t?$<_%NQF&^5_UeW0SL;Nvlw4sXx;_1}H0&*SzXeGa8X0mPW$Pz*%}! zrmnyv3Ow#v7_=G&N{qi(AA3F#{Tv$7;G(0;_SbJ=vR9=bHxnz)A%EUh*jMwROy}AD zjxgG2Rzd+sjWeE{6xaw(K0cDXyu9=b$jF06?WNoUa`xNZzfRdZbqmLy+|P*shvKXB zWLi9|pD)okf2}l{c`nNNj$X`P~i)b9d8r-a^_L0V+9dGFBOUWl9H0r z53Q}O9nbgP$$bgx?3BoG+oGZS9WZ}@uvcgyeSho(b5EfZ`(43%aa8`gJg`4A(BS#m z$4C?-_whdz9ik6cBhbCw$78b85GMoC5C(1Gs_mkq`&Tr4006l?DO|mau##%Q6(_8g zXGF1Nak_2qvj0dZm4C&xRkuUy3#M6pt41l-!{Kv`kt*y-{QC84zz>v1tU5K1Su{(O z1zyn69mXL!e9$A5>5~zE4qIgRzve|#T_0qYm7Oq5sY1&vDj*CC|GjRI@D)Rx&W}OR zk39hcREX&rCMG7UPL1l3VPQW-%^vf!Enx;4gpN6?03E?Jx$jyp0QK&ML$rn7Qlrjv z@NLWY=k<`08FD)VeiuugI3UAl`K1E&NZ%q%!G*-fWNoNnaZSwD8O^D$n3Ocu&9?Ob zDf?&v`*VtTsIVfjUwO+vbSmFl$J=%42Y_aIdfKY?I|UJ5tbMDXpYVh2ImPqeWM)R5omV__~oj|CvenF<2u+73i`!LOd=hD(${^bs30aWF$|iZ7ron7;<6Z~ zfDT&$+0kETL4I^Y>KVH=IWx-7T9fPjwtbG1ty%0UgE6)7sVMSm|E>{xA04JE)4TdlT(LRC3N4 zB&kF}$w)>7B_l}AIY$Wx6FGw@L69Ih35aBroFz*Z5Xm{`+3owj-}g;T&2Oga&i&)o zol{zeKHa^0hqa#dJZtanRs?v*$jhsmeU0OLO)A&<;Se{};V5;RcKUd&n%c|UI34Uu&OpZ$j;CIWLj2J*m{5}G2&oJVPGO5F9sTVKZQ+BTE?wN} zG})cfJ8-bg;dCZE8REw*^+d3HLV7yTYq1e+#MMqzKnt8li;TP^&R$qpsHVCXE>Pt; zw`(r~r5j~DJr+xf!-DRep3siWGg0LWbB9T~318>nbs>K|&XS~{RJVcpV+iq!!0C&+ zJ7GsNzC?K1`7gH8$qgO~oQ}@U`~LpoyL)@amAddNA9Lbv#(dLsc1L{)eUfYSa!3|n zQ);-{dsU)ytiJsCHB$bF3PsDmxLo8Oe)OHrNpU0%N+eA_84?g8R7gWp^TU@f6gO_% zVCUj$gBAAhFq&fQFisT8de4X)Bstf@(qD$z)+Wu?DQ)9^n^qy>NK6)~aI+22g5{ z-T}yS#eM2bQ&r8e&`)~3Ug78r)=%XVhlXyih@s)^yP<8BqAF1RswQ_T!7~ySK@v(6 zqo?0}^5jWJSJyY!ez6=6b`ppyQL?p&3~K7xx$Z}oEX0M-bT9gHtcgw6t*Fi{nsHzK z_tm=y;TG8gKP7X_=!(Lkp;vTHrHL>G&W4P#7iH#Av)Re*h{Sl^=(3e zK;a{@wXqWT-BG9ngy`zkwP$w)ZsW=M&Gs`sK|JO3hW8W79U_e-`5R#>+etA5Qp`1trXwtviN zmyShFyull;5*BT?lF`*zY8cxZTe?CtS>Yk#^we;XFjZaWnm6EiH`1u&RflXk94L z)&^LTk~4T15T~u~b28!qC$HDY`a&J;nz7!anxD(ReK`m5vQgiXn64k`JSEw}q`9s^ zJH^e(DW#}*J(6A`GF8OW#{oAnIT^cP1gl^q^!(hnCrLo5m&f|$beYIwOHavIWD0Ks zcGDGCfvkCiSjPY%ZbT01GC=H%)@zvtQi0-1WVDfazV-oz-;+azc zV49t1uMPRTwM63X8)_R+2%Q?Z3hP9k3C{nj)S=rX~f~KK-M6|pKteZOnSuv>e(z9b2#q!>CwjMkl#6{u~`zlrev z&>6IP;PZ%AVJ>t1<1)yWGEP6VN;(KEw^RL%pMyIU4U^4y^ z;b3Xs-^zlTypBq@lszCt>3BvX!vZA2yF?Pp5WoU=!!Z;5a4grQzOQr9L8@PFKlHM- zTiZWp+^`F0Xbp`&&^@lsQ2jLZ$LB$xoj?ZBM!O z8v_Mr6C}l)HMt_SQ;7jJr4+^^`V(2Mm$!ANO5vf#k?A+Cy+?gZv#>`~uDVste2x5h zG}*nHRI9D{qp^cg#6#x*1=YhZS=(8<2pxTWv8+AEnL5gJ1$G@y8{O}tD>8NC$UM_| zi~|~TIwKAcJSh^De!hNY9Ms5Y1W_ciZxME&L{~^sUXCb=HP_d3ocLB1#qB}OZ*9PDe-~m;Jlxj}TD=b!4=@RNG(cC}55QJOCm)CUc zWTZ7G2Lpj}!4bZ0N?hlNQuLVY`IHkrC2?LS|G6#i{iVhFdwq8MZioc$ViJFiWK^9e zPeLN7@0zR(X7f-*z|UR)x~a#4pNg7VRaG?*I1dR43A{k)k?77#-4BD;g#2VFFO})` zU9rkz3G9b6`!*$ImBVK1QA3_9o;2TM97lOEb7i@*K70?Ve6P7JV`F2dt_7XVgfKF< zKCiV;&2_M%AI=cN63cgo;^N}Uy!R}oYTU6!Jl31yIn*B;8b;;kb3#p34F|VL$jKL4 zLWpl&enZx0Vz(CnhD>G_MYX?|f#cJ2_FB|%dn<7ZjYTkHj@@&1=#)Csum4KDr|+i* z#$Y7(R)?qduBYQ_2#1S-p=-Zu_IBneCy09C@cr52pUmHT`vW6T7viL{mgjVA9@zMn zHNg8;aP$h}yJ3~ps+F%x9-ch+?AJzdVIfIEBHZ3i?kyDPw+fZl{kTuhFKqm%sV0T1 ziQZ~t`+DtoiuBYSKKP><3|)Iw18UGaw-3gnqk$@;=3W?wP3kj zURd^+cPXN6lBAV9LG5JcCPBiFG5G8-4degAG;vxuk=XZMEUd0h6oehY^pbkN zGY`<*uXcjXHkI?oI|~TEdWDJ;$HajRA}^kslyT~qRTtf3epfL-tXzru(Qc$prNj?s z<6Mc~;b>jseBfwd*^!)u z!Jzfx+fAssMMWE~4`W&2H&Y&OH{+uo*1*s4gcHWvhJK`i$+*~Lx0;9_iSqsu!|JT+ zFP>_BCg|<3RB(0PR2tBKl_H4oj)UyZ?3&cTELW6(%WvzM*Uv!A?E=FKEMnG&O_8)BD2GWgQr7~d%`1;L-vilXezFiP z;(uJ_GPK5WJ{u5=-7(EpY&^-Px$IIS$D)<-l}O5HEvDZ&bB=D1xJT!Wg#Y*c-iIMO zDOX|AiEkT8-1Jq*XFLNK=&fwY6=O@iPAMQs&>~m*QJySd?vPHSzZhiH70;pb*^Y8| zclW28@J^1z#dTR(S%i0QU!Mx_G1KPPxPFn`26Yr{kKWgn%=*n0u=@IN)K%RgT=wQC zkh*OCu}Jh~Y^8;^7M0vVa8B-p@7N3U@yzY=9Lt#_TdsHF7;E(XZ6%Q!a}sy~k?;Vc z6URkUS-x(~3c6qN0xkV0NcXz5o)eMPRcvBpp(?5)b;cW|%G<|!tA^Z;>FK=Te*p=^XY)%x3bE8sR}|L{YRTNw=GGP zEV39ad%kOJV-Ag+3evFGXM0d2W&U!`8G04F7IDwD55BEiVT=GMNj;*j@hZ>CV@Rsn*r7#P;*=;$|#^11VqJ@(tTTi=U4191%<^s4m` zD}1M7M{lpJg9FzaE`4F9k|&0S_<&qgR8)w!4-F2U;>zEo!Na7%3tF-?EQs$97zrB@Lm#pTdE6AyE&*8mzjk_81TG*SP z-7_+~LmYCpeJ4ip-6W_>Fb0iW@ABVZ&2aJLVY$w;>=qm{$jm(cTt4R`lfcYX(d@am z$#0#WWKk^o#IgB_X~7Dilx2_VJE>L36lxm^yUmYma{QhbJge>$i4+khaq`Bxr$4bz zuJ&{Uf2%zB8p6GqQC82-;&QO|F{>?~_BpE}lN9Emh=_=BnNF$amI3^HjQ0y`qearz z)*O*^ViQ|JGvXmCZ+Q?P+6G6BZGOLIt*4}fO+-Y5k|f`wkA?P|>s7wdui{bb4AfkV zxK1=chMR-%k2oNFdhu{a|Gi{`G-6WHSGl?D{rwL%dT4JNixj^JG~G9b z_$xkqk2r{`KR*P~ecSRs64B7sHtkIkc;VA9&E@z}$oSh!$=8a6-yEti-A2yF46%(i z2Rpu<-fw~+On!JRfYGK}E|$fnjdBYf;^t?^*+|37-Nf9yyszPu{H1W3akqo{g->z! z;19@*5P3VjNQM6sGW}7-*;Pf`#SzEq*{c>0F%=7?*6uL zN5z|3UE&>dqjV?#!0@sAWDmv|<4tf`3^1Rt>jym!6^QntHLJtBl(iUFl97=i`KkjE z5fT#0Qe429?3#)DFK^ZxL;-U7)hKwmQm_us;e#8k6vwFKynFPkLHBFfI;bII_uJNf zQZM~|A|8VH&o>zolqW;@8l3@bs=hQvD8S_#>p(L>8J99szNl(wX@vsM!FutPK?Q`zUmfzJk4u_2r)y<> zee1xmtn87l;?aq+8~7b)mIYk-BP8MJF#rBWm&H&(z2R~*GBTvrCn}XFBIL>(XB&|1 zT~mMRGyT1JD&dvdrVSx>lW`KiWu#8Ig>ssykdTlG0BwVQ!v^1Kof6CF(u{#}Fr?cEIetK)ViHnyDpvXKJqFrfZ1`0g;GxoUr%+K8&A zrg~d@VJ+ra87$q@{C#{p1Tgtq26bQhcUr^97k+id8Kk+c4(ItVaXW4(5zb#ZYiwd|*(x#t80oOBx+2`0nR zRs(=Avsv$U0{-;$G$;Z8#~L5pHFKt$`C-)}zBY@+T01lBNB3j1#3is9Pibq@_M}@M zU+M$F)H+q|DkjmNCP9C4asod@3%Kem+8vdm9gpccTKrL4e$RKcFIuNG>0Ld&hI{mj zj4_=kkVeNhP3qc^&20Nhr*BeSvh~qld`t$E}V+@>eVa4BgJ8DaS9xatgI}w zb@z9*e7`rhc|Av3L6a+8v^LG zHceX!dOr6wsDxMwex8+SdnB$*aMI#UdMqp~qdQc9&-Cb1x_TbJtz9+Q#7onyE_lLu zOY}N+7vDMjno|u8jo^c|F?cf9d1FEVUNEIyyk5?<2Kcg6$yk_jvS3J8*Mlg$DS-C4)Q-g>*IE+E@uW zV9MXj_(BFU<->sCg~hQw$|pGJINcw5gpTw#>}tGtMQ%WVP0qExmjcf}n;fo>!^ejZ zvxU-%c%a=j_+1=c-93e?3B0CTzbNLVQ$Lc0iHRvAxJl@LhAc$?HSqkin}i@rboUzI zbR%2GhZ~b1i~t^`860Y51n^z6N1qCd40&h%<(b@qRzc;gX=`gMJbS+HQ;!uxc+x(V z{kU~1`-2HWfw?$8AIqb5dWkxwYC%S~=|A@X;p|gSS6yB07b!x67oV887ZH3XtXkfD&xVV7XEG83Jll^lDIzvP* zu&4CAyz9VkF6eG@aLnSpW$s$ykh?%-KFJzbCpumxCTDH*UW;!_H`$PC$12L zZw&uOo9tv)>!0n zlUX8f^j(I7Xw8yC|IH0bJxCl#N|)-Vs`Rq5vN0E+sxk6(s?N@382EIBPs%M;7NR5som;>;D=}s0 z=`8v#oi}W(tb%=qhlgbji=LFr^^5L>zYnPS&q*>OA_6ETm0z~5E-rUme?Q##QO3Nv14zj5 z%naQDINj*((mOGv#ydg#Bm@5w^Mu62wV_86BgBtHKTz^pEQ2O#nVM#@ zY;5*(BrX)^oPMk`B06{|MNqvyM@Hg}jE=5$-7L`ny%QQ5TH(ECq5O_dHgg5^PY zO&@G*Z7V%CbSo<=aw{q{yuH1#UPvOjYGtt zP5_yA`p%CKFR0T0!UwMBd>D6Rp?qI|zqYw~HcB8;(C$b2*;b>WLzSu4mW-LX`8zQ* zhRiSFI{m>43ZOxvbWuQ(>fm2v=>+^)&1I^lrsfN}5vcM5e+$Dqf`Y593#tYN z22mFu9xK?}+w0bOIS=bcq?n8rKC!?@P`(^<#1sgGkO{OB{!dd=Q@eV5BdR6t-qi+w z3;UmYCCOtj5Sy>xFoF)<78A5T{`vxnceaH5#0}w##J_p7n(8?-5KPEy(x*^SS*aLn z(U+3vywH~I?FRATqBGFk00aN?>kDCz^|!C_X!C(Yz2_p9kdSRtq^+f;HHGpWYpuJh ztLxW+&*IV9;*M7apCWGacc4ZW_f9mlwA!woK&WzvYwZjMs9L{0YAcPtqM*OHkB)*! zh>w@4LE_U2XFY$8qN%B=dX7-Q!jim)vO`shS~6C!y1Lpyop5^Uv-taW*V0~Ietzgx zCKUS`<$GWufp?5bi|xMIE2a0ry=w>g85z&Bu^_cjuf~qpIM97O z6zgTP8y>c6qeWB_{&lAF0>Q(jf>o2snV87<`P(=6=ys?eNX8W7d!QO=xv^%8Ed&un zfTB`9Uxs<%HjM`dp!j4!GHuKTs7_q-X~-Y!(A+g}RG;)YF#$d3qZeMd_acUv()wi1 z)lE`9vs?=REtY3AKvkKq=*ZzGyo#q_`jFCH?f49WD8r7~eonnbpz-0*MlJvTYd=~F zH9AA|{bjF~cPVuZe~w^>0&(uQGjChANFbKv5W(RKEq(IBbf&JmI252leXmdmDBD9) zqufN>ubb$%n15>RY_h~m8wDX7(&3Hy8g5~_7>;0Gl))D$ z1Rao@C=)S(DOwgjxIW{u*UMBwutEM;Wh%JY+Q{9R+t_(kp3XcU25|+h^Wm?S;|zxu ziW{$Wm|5anXCEM-TZ}OjnSFgKCgDgH6AS|AR#V81>@<9hC={^8pg!h#gJ?}EoG{uS z;5{;4;m+tQ3TB22b_f9}j*aP%gA-v^fFxlU&E6{GT3lSzEi;krQEGd4c-AwlbXkfI zLYo4@L?450ZE#|OYVV965yQnpjm-%wtwL-gza9uhp;|`xK0j_5S4$0~1qOg2n`i^A z_l}Ce!k;6+mDN6$q^5V^fc>NLh-+E*Kf3dp~`o&yn=!-J?OR6hxn^=exYpC;XhXQ&3R1%A(9TEMH7X%Jr}%t zcj$W3+?2!J2s$LXDuiPFvnBFzf&25{i;JzPi2V~^j;j;M=mrT=lIosu@7Q?Z==G$8 z1kx(VZT_T;icJn3VUYos<@TIORV4xorAG#|!h9J`)-f1t+=rXAj3#!MO%BbyTvsES zv`iF{Ck1|8Ttqti)8D5O z{A2;$VaKR$DwR^#)@`GD9~t);LMatBk>Vo;8{S}Kbf|b4-$mpyG|`YCk;?aYxme!u zG6L*#-a<1Oo3;P=RGgLbnt}wDWD{!&1$k`G+vvaWi=E`b#M>c`I3_7j>A|onZRl3m zTLMS{*8)N$JDEIH;4_ul-QAt2jy$O5gR*zU!_*c(brlOS}uZPg#; zr>9E-v`V_`_#Tf|*leCX&=nw!%apw5Lj^y-`lnopQx2ZLnipMGw&$8_j_1P!Jl20G zuvzm^!~?G5x4v==Auqd*DVtJ7JXc?l8+`xO8JGE)`^hI6IDb4^q?bd;EN3e>#x&`Hs0;#pyF^Yr6vLj4XZjcO zCeq>cn6&CnZR8XJ< zAbJQ;j1oB22!|}dZH8|dsjJ6X^t>fJ`hqXIjeNtUPjymPU!TfEfJ%fHgz0;P%GTDb z;1?Clh=BoJ-eSgvpXpWrO0}7UC-VWJSarNQbkx+<&0k-)v!AJZa(C9ZoFxxXsAq@k z14#n5^K^IF*yM$Tgg9tLQZhe(7Ie893Un(GcCXj=lnI~RN^D;UNvgk#BOQcK|E7OQ z9gvzShpJDXKFyOMRW4{zoQlJ5K|vrIh7hyK*>`t$4}k#6G`aO<0q?fn=ZEmkG{7V) zeNSAJlkO;vJ#~$slz!bj47kt6M5T?jwW_tXHOoZH*3nH0ocx-aq-${oEJTFrnwqT_ zW(bDYpm1V%rTl8$Uf)pf-PuB7HWjPG_3_j3e*d6vbO6gRA_4(FG!(1-DbXE~-JF7gn-1hoob7V!{nbLwl-G>V z!SNv}!}T?OXCCSS{s4uhPrQfvHqwk zu;mI`mRPEvMnm3tz;c*VULN;@9!i0ra)bN9zHSizMbLr$js#`Jq}E%4GeLj&3;1x@3(XwQ4gNU6$c(c0891P8COeN0tO=rmN84Z0P@P6{b- z!O+Mnq9VH*z~&HegZV*gAq53g$^@|0qQ_v%T#n6wFGHgzG!OyZ{M2*f=TOe0wE=m0 z<%uK;sB>&OWbNoo^YP^Y; zlP{7C3~BwIeJWFdpzIOblKPK+XNR)A;(cinjbr2O)3u&C>qI>K!N~gUQ~NTdCfM$Q zsb#>^mKONv_t=D;&&EQMariBIzQ4Rq>~?;m`$2L52Dw@zBwo(nmaq5<@ zeC&P?fB*iE#Ntvd212dn7Ue24xA~+HE+goe9~AmWvq0CK=aWj{-#LbAkVCYss0`&g z`H-mk5sD);`h7Zt8V7^)8k&Zh8c{X<)(7bzPz*DFh8JW;rX;5B>F&O=(@sBzPX09g z{M>hAw$U(l|K|(*RiKITIx|s!9{6bL>leY{0N{OoLBUWAv-~f>g!T26cVaLsWOtyB zTW^vOzPp`lu+FsZO%gagtZJkk7q=c{5j$HiuUUu`c}PPGt%Na^0)kP|9d)6PJvA~y zdPG#JKT00dY+Pj6XaE$Ns;cP8$w_`;p#}rN_t&`8YE-SJ<{owN_-|X)(BW5v6p|Zw z-7mN79|a7ba$MYd*KA|DcB~}^aW8IiL$GKdvyo6VFFa6@yg+70htyeAa{krpF<$? z*2!9>pW+(nb1{$VnXtS}Rzy84dv@e7X^65(R!|DmBtXDV4~<7?wTA`=GX@6-6+Tem zU;rnU{0oD3#f5il*oY1`(JGM3g8rw@p1nnzwW}0_(cUPLUhQ3YIDk89! zeyY5+{%IY0*A3bO48m=#sA{$TXl5v}VgH7rnyqaC2v#4o4ZLMRxMPhXfQW`>za(8+ z+1p1@Vw;l9+xWh>CPm20n36BBbloqGz6+hnMP zjOgDHp_=A+k&3j8G`;h>_}ql^)3smB@{j=fp%7rI0+=8Om0H0niZ7?!*O9#Jg?r`Qs7BK3fE2ONT8{Go-4XBjbJ(|b~*l^?E ze1{T?L3VZZJ1FkrN^lJuW8648BgMhEJ;I3lwig(|`gr-l?A66A;Lrp?V{4l3?(Sx9 z3jP&*;p7ECtj&%^K999%!7zP`+87^D_|-|;)srx?+k*ga?>NuPm|I%PZ=?ZwtH_KB zt!)FF&COjd>$3;(>|!(avq-I{LJoH7;Ic&?iW4vhpQCBd1Ny5Sda=U+G!!TRsA<~i zWoJR*tiXmc4S02E5%9FtubA{2#>OAUn-P%Sier!)Yq~VmGK{J6ku(E-N7M2$Iioz7fVj+-~bYWo*mtgkLFtZ9_DCMG8r zcf0`iA#?4$9*v`7=Z_WCtXuC>dF~&ekrlMxx>)xou#Dv>M1c@8d8xt>Th6`58fuQZ zj9W5GmW&W}`iUHuhHkZol+on{@8|xb#1K(J`oTNPw4f1NUP@?9E<_>w%j@ebgQ(DH zry4&Jc$@grCApz5uQD^DusLJ^KDz_3wP2+tKtaHOMA3;+9h?Am zbEwRJH!f|NP2wcvaz95xF_x?P7m?Whcj=nVnl%hML-$8(jNo+RE&ol0$n`y#o08<9zY5Y@xTQFTIzGam=< zL&)Jz8stu*XNMa@g-^ z!kag%fKCoiRl5=k)9K47RC=khe>HjxbO_g!>p9^9iDJv{)h zI!r?PDmE>xq^s)zU>^2AN`{Q>ziQM_~sq7W%-2M1Oxbm+S! zsP(M(Qfd~RBdALZYcz<1KmJKJSX$ijD+(1=WF5d!jGW3tL|%sf%w;_xw%R3yRRt8P#5 ze9YM4^FnenX8`}|UT0*GezJOd1irrrUP~Kuc7y`)X)~$H14$}Fi{LNf4`opBdjmaG8AUuw{ceNAs zJ>CIWC;{@U!t=p@P`~NU|Bm|6OsC!wdUhyG3=|2J@pj;ysA(IH@>`z!aD}58rSI4b zvsYGByyYXSi6fQ{`c$t9j(fP|QF~kc;f=x6si~=lo_uydwLt`}ovEF%Vl{DrfpCoV z+A+Dfn7yihjYmsWDMo^|GV^R}B#gVkYOn(``ngWTF~47_ueGIu5&KoG>UnopmzuQn z3%Lj?(#rbmY+3LT_B*XmWDbka>x2%C4~u>VlawX$*CI{b*@n>V9P9k`%rHvp>-Aq7 zcktlQig?_znU>YoPMq=G>!p)rgnppMFdZ)Zq13b-6KiX03CYPf)B1Dr^Evy-?fQw( zf5H^3`Y9Lqmynov^kd7V8lEyq;^6~GyvZJ=BOx(^V+4d`R!8-d6PI!jt7TQKq5<1&ipi!19xofJqk8i2;Fj%rHQ|YSzD2>O(Z#-y=StHS7@W zH!fAmW$N-VAM?LtsgU45C%_K7GvBIP?UMZ(htlNG$p|HbFWP#_p^EI5PP!3Kt|S8i zzgkv)e&{^Dr0x?=Odf9TM*;!@HQV2?d+8oXbr3!%_EM$%Y7{+n8|N*lD^zd59m9mY z4*1*;9Qb7;2f)LBe1skfl$-JohUA97?`bd#{l%n)hD?B2!{zB3V98hj)!{V9`1rW$ z*-~$^-RCT5C@i@)&lSQK)9+`E`caYIMrc>&r0uVv$|+kS4* z|Gr3MtB(JZ-|nwYc={`ZS-OTGVNP^q)uZGeMpF?!_t32&jaq+O(gvVNVM)mt5XwGw z_(9FW?yK?FnV(YzcOi+tpWrn;1fho;5EJ{^27S}^$cTwcI5ew&e8_990jY85Pj7mA zdw-m{{%hBtZ~dRO0CsK-s626?RzdXsI@sh~8h(ED-{wl;i7FuFR~8cuN2wj_j1RYw z{*3brQ-hVGH^IjNVb$it4+iCJCkQVDyvuMS`x?TZ<$_Aj}%@U(_mTvu0DLsyr|+Zi9@1w1P5?gFw#R2wt( zlWK{`+E$-9Ag5P+G1Y>{8~~?4Sii%1q4z6${SsTwmIOKPvmXAem^>!jA*3Fl6B_{C zf%7+af3{gIulLz*pfiq*c@nqzr$LioVsS2h0mD4AV1}4U<{qtN4d~gri^Itqig5A^ z3mZF=OcvEcss$5j>VowGi1QPbwuzjaoIP|{vA6V9>9y|vYv)O~p%xEC7StoDqIV|S zBLACd6jPEF&Oh{Do%h%5^@(VLU=~8m#wj8qVsrJYGY&ZW_q#H_W~?TDkQwX+y}iAs zFsX9%ThoiN=SmD?CHN8hmx#zR$}K6;c5o>66{jbp7bM{xZT}MpiPlh?w7})L4C-{i z5D(lesgokU#wYbr*7wJWvW~X4_7qLi3>|ry_|or+8+3nFsFN*i8Rj7Z#^2D&gHTwM zS;5gleL2O&Tn-2dtXF*}^)06<)H~wH3!T-X)e&vOqN~fTE5TB0{?0!q&@s-Up&VM0QJ&sP1>9g7}{#oxc>=KX>X5JvJ~`1%`;xFMad4Sr|C zQ&WW!=X3bra;pCR4WXf-z)JFo*;EKXKX}|avvdDlk3=-xn%?0)I5>c3zF8l>Wk~GK zhNrOq%;Uz!5-}M;KO#Zh0c)*rS$uFYBykZ3KY0yK#ec)*9)5u@o(n&#D^g^W=UFh4 z+5FpX*kwyoOK1R~Y2g4f0FnFv&_Jd+2p{%BzIo4KZ4Uz?=pG2$@Y~STVA4KX_$1RY z;A}ExBaT&hu)JSf332UhOWzLn^eHjwkluOs_YcyQerF!~ZUeF{jZi76l0J(6$ztLG zsKWB$q+^tirO>|zuM8!hEg?S@np!@%$6~{>1Vi%qEK( z5idPm-CsLi8ag_;aGrX$!OzCdZn&swCLf!-j%V|zcG}^#PN^i&832-R+_=%z-L1|} zG6)bMH~?X!2M{6Oin#t@R1c109vc(cPWNsBLg;!~4uitfH!_dtoQHsOa|Z zMaCx1lpi0cKyCiyCjIq&cUr;q!q@q6q|c7F;Hh_*y#peCp3?pwdCC>wie0_EnqWU* zRF0F!Zo$o^*S)++bqJa>irnpSxbs#*NQ<@t?*auKg68X=+)^mUi& zif2a-K;|vCgb>%$IktsUCh-_QkEE9Ngw=y%}o2Zp~c;A|oY@<4%iuzPIH zZ$yEsmzS4k9tK>VEXCj%v0eH@_mcm%zV1Zsv4tp!VxV3*xw*M*8r4gNP*F*3>$QRK z`uYMBUi5+sc(Z9+cxSG~Yf#j*gXb@zQlPIK9JD|;zyWl$8R0|&FbRGV#0|y0-@R*u zRZhi|0@b~QP+ao>kOwLumn{Fwy);hb92fdsxyG$%vD#kYkYv+vcCco1##jlJ{@p%K`Fpbh8J`I)Xw`o6^5r67%4A6 z{L)96`D{NabY$p0qIn_iadbQRH9kF;A{=T|QiVwqxOJXBEdFVNhBCkz)Ld#vJG{300?}{E^`YD zk}uOhP^?`)?x765>>H(c{UKW^aTOp!KC4O!`$`!3H_#0S!6_WKZiQooiw*s~(&*3@ z2T8h577?4AT=4t11+9d?XfHD=lm#<}O!+7|KYyV2fzOE@|9s)I+sSt3HwDLAdcXac z8NZ`xgT~Lbb1ZK#>AWPB61gILXa177{!!S!U3Jgf;ci-c&ygnN`mrs7I`Q{Fww8gx zdy^tBV_XJ^4XBOf9WNid+QSLkgt)lOv(4H^wH_O*d%Pf67+jn+Ue$u`BO964)6mEQ zp@~+^=Z+0oXS|C+vgk~RuWFsiZsVKd7W zSD27d9tzL<7-o4JfLDH}bj>{QjPswRNuSJ>qMTM>w-OQ2OTOzk+Ix5QmLi;#crW&> z!lE}RD}uH(wmehy!Gi}(`b|N2v}*V72j=JJpK8?CS>ddXzm26iC|@BX#2NiBdZh7Vb}^Jjtx1ged)qRPsVPv{Ik z&1AYDq#xjBPwRaufq{@$^S8rup8;o80Yl3JnXo?{debsQbE$;I#jKNaNaKWy{rHec z(}4!3R>!VR4NqIS{TUS}>b%PZ9cK!alLc1{uMP}hiBb>v9%~dqaBqlhgBmGE9A`_D z4n6In0RsC91 zgn(yS#-(|bO-(a_3$RADXBzSRFm0UN7#%Ir6YxLxQgVw=ezobs0{bbfRp8src_@Yf zz{aPgaj%3aHN#2A&P-H-sL74bGMSj`A(1W&7oZvzmX@|ByNhD*s)&}C1CZqDy|TEv z)$xL&X|#wBKS>G)xCf$1M>DJLS7)&zg;gQlbM2jXLV!=m=YY10mnL!xNN_A zEG_+lWySIGw_}CfIA5GR=Dsp(yDot`ED& zFcK0HVhz96G)@%qWo3PxWgY{%)dMs++*)|ytk~u8LTrPVB^`*a`$Lb`Yz@znfK~gr z=+;G;%?2me)zy9BM~bKQza&H;y3FVD8d+UWxd;mJ1u_wRvP z^>La{;;xmO`fJX=S2(I=8R7{8Z3^RV(19ZxJQ=`#^QIE0=iOja{WN>--|2^uu%D)6r!}P9e}XV6^q`pIbD0(_4*ePxg5|+c=0rf zFa=J2dV13#?|tM5jE0nwio)KzS+Io(2ncY{;K^?!!!tvfC{BWZ)<%T-8-9D}iK8NH zVnk#-Mv>~~JpT#nkM6x|KatgVxz?zqq@)z(x8fg*@m-{9a(YWHszGiUN8?{7Fh0Ok zz*nlTuUmTmY(FvJBoV~4$MUG2V8y~BB5>Lj9z`X?)7UKT_xLLq7}1=d{qQZEt3YoIf+$*P z-u)&Q`6RJh7IDoi-K}V=@&S1%3%uf5WcDQT&27DpqoZ1jOG|6}D??ObJ{5n#jpSXP z@-)c@aBll6zJGc<-Lr?YXER4MCoeJYsn*m477e0L6@mB``}eHfO~4F5FcWs3fBF0? z!*r7yDk_ILfJ@LfRZ}k5t@Oxn&+}vFRAKiAva+(gmbEUft|L=Z5AX6IiX>p{(^&y#HIFO%cbO_s&pg8#8tiBemZ@9907%_CZOzYd8o9Ozcvsx5D9abBQ$ZhaQ2>IEQ^Gfo-Yi(s@gdqjAzc*NRm}t-a8^Kt{;6WK)f2H_l?f?;f6O<{Q7Dd63 zMT5?u_RX3QUq-m)_kOppq+<#dD%zeZ#B?Kt?7y}59#C0j*}CWk1QierAW=XC6;zNU zIh3M+2#7?rWNs34gO3olimLMpIBndw`3X*fqyt%NtPn~zV`t<4B_l`I2 zYsRQj{P(~2T5HcW=l9LD=EP$!-AZ6O`X^mm+`?=ku_9t8sR%2~? z#C{hOH$Q{DIxMHLaqo@P-olr7*z>p89{)1pS7g@539!Rg@I1ZsRQ5h_a40}prWlZL zXP$W!*)hS`?ru#j+m3soVy-2KK3}|8_)^cR^Ewa4`~sx`Pm&GMaOwV=ypzB?&R zv_j0)q_{YixWQ-=0k{~C2_>^ucbnODdZkJfz6=Ws8@oh=NkwB4L=V*WX5CGL;khlr zYzAW;dx9wSR8pIskz$!1-p(xTxu#)eALp)8Qc(e<9Ls6b_UbgJc4u%HIt2e@GTZ-Y zSH*6x>b<^7QUoj=rzkHHhv++U(=s@Y9v4|^Q3M{$Yopnt^%Z1;cn>~GG#hJ9z_qr< zi!+t^{KUf)xR!BD6vSgqG8{P23XE1ea9fnmGrBes)m9%Nqzu3x$h->hBbf!M+f+TA zXa7bHrHSLAc4xh{|B`^?>4t|%Skk!P4i@oLvMSK*%(sjI<;?8NxuFWgi$x}I6eSI^ zp!hD^xpOV!xRfQaospgEqxn-)Q}g!1Fk-KQX|~rPlsR_-E%;M$DXKLyE0SMQrF1r3 zVMg<74+A}KNw={W7hOr6>@C)KA33eb9`sN&YX3=m-J_>VUF@ueXv=DY489@8UfV z*?oL`{A0_@%M+)bJ$v@b`F*N-&NT+V)AI`I>Y+`R<${Nw-$K??pAgd}@eRkDv>m=D z3?R_Kz+9-AfL!=3aKWph=H}*h_YaMJNTj@@f?*+_UlHoAuoS{(1LG!`Rh*QRRF&Ev z^5DUPfY+~I&(C!{>S=X7K*4!dd?uCp)gGpcl1JTuQGohv8m|6{A7u6pqD}V8#r%w?&k@Oe!B)LnJUQ(YZd(4%KZT3GQU(CsH zK0oiU(ph}P>*$z)*ha|TDFCUX<+@+?tKgtUbM%u@qg2nYM9hr%`1nN6*v-v{$|@>s z2lp9PxE~1gl;ZMg4y-L>pnpXW%r>+SZ~4_Yec5@9T+&r)+m`nBL@J$!3ZU7*mWV8S zTsjU5g918%FQl%WKJhx$qs`o-vJAugpLjBtMRF}4z5rajIr>%Mp;vp_j}S7s1z6rZ z=c(A{@Ymm&j((=`l)|5Uj(;9d`|m}+F(GSYO4q5&QOysx6jxX<=9N9_%MozroCtSYX~fm^R6Sk zXV+AGBSGf=-cUR#wmZLJW$U1EZ=Si;$YeVnpJRMZN?Tj|+LO0X{wPoKwP2?&{V6`or+U*zM{C=6?zven4tC&W&pJ20oj zw6dpm%93E2V#MO@?#-<(w{pT-&)L2x$HP)$us!Eqn01O}YfzA3YRBLmv~1`5EuKPDrW!J~vaA!_S&KI*i8915WHF z(yNpD26UeZ{|n~jU>zjHiol-SKXGfKc#WDMg!01TVS)=Tj?@m}oa z6a0e~`-&=)4n8BmswMIERlPSEsuE}Bmit2b*UF2w=aCKeoEJtSs+yUwk%16$-)!d% z+iZ`Qw$?;meV41D%_v^F198lHL^&&^bACQZ~W_M{vkW^qa%8{q$(M0!r{OZty5 zWtiI$1+3gMoO`kDwYMB@E)ZZ`7kval(Z|S zK7;F7(g1LSVgLZiQRxu8)vLbrtgLg?VX!vrK{X1cir7n3E78)SRjRW!uDaCgG?SE` zE_eR?`A|+hcOr6z`zdHf#7HP6d*5|3zAtoHSXfwklkpHCV;tkos$}s>uT$Y-lZ;Kr zy&$9Q%{4X-4zZCLHpN1>khDF!xzb-)vM1Wu>v2m$pac&v2&+l@G2lKJHCtD+1SNuA zYwJ^tx5jIpBVnS+*Tye<09Q^tDB!vCdEd{Wp)(vDel;c7ck)^c|LDkvbRB7w10po@ zG3L<8tD~C>4Kwe>UZwYCHxiO!sBlTGTfFVzvxDF!v_8fCGyd}j1IoWwntSr)Ai1hAT zt1pRCH_gn_foj;me=aI&eqmJV9RW9m7B z-4>GbvXR!0tr5SYl&b$rY_a=ZgyL>r4vpdOitX+t$D4kV7}$w>RNM|#!`?>G>tNff zRbs8^*bR4gB2)tFI9)J8ud3Vb1G?+bSW_OEAtpYTCTs^1~l_`&o}oyDk5DtUU8r&_Ms)7O;%fBMW1v#nzmSxP1{L!V|mIelrT)p)upoj@w1BCD5l_2Dwa&3 zzJ6z!`{2H>BMV<9-9)>9FY6KgR;0?)=nNF(FfUDriMb5W?NvipSC^vT)&A?h@4!>ngWnET`I&ruc^rB7)z#Ixa=8UV zj(z{DLHod&a3tb8wqEm|P#1V=VPT{suhb&3NF$>zNqi#L*F7te(C#Olu{|K{Gjf{-0RqeI5x?^?t*mvwD zt?zk^gqS$s%^RFvZpJRq(#`wjzK+$i7blm~oVl`>2a9#ri{jjSfHpQsF~N9YxK5#q z&%8GkV16(U8)opp5nb{^&j4dF8f%I}K|1!*%?GkkqFN~1fQ&&n5jJ@9%ru7pgZ=7NS zxP8hGK44DXu<5GiyBvejy8%L$_P-`0Bg?QK!$wo!S3oJjBBY8I9Q}CvJSqequ4cNE z9b*fsc{3l?wi|c23EHQsx9IygzhrBbK$d9zE0Oj!VH)>$@a3fZd^NK?Zq*7-0)EUa zPlJVp#b~EsB+9j9YP&t$-RiD$_{oVu`i>`B4X+gqwhINaCkbJs7k-Vj+HvTWBm<|@ zK`BiYl|cQv(71*A!SP~cOW#z#-LE?J3wKb-GNHou3KrczpZ4oxLblgYt|u%nul=lU zT`=cmA-iesNj}oce!}yfFTA{V|2dRbhS7D?vV{|Y(Pt!7-v;Nt4$c`4ls_{zHI-et zTy0n!k$|c!b_?fKH7&2SaQ@92x`1g9|3kT^UGa409~HiQ`2yEFPUW;kFQ1;60sMNE z+n~M$6a%u1qwlL2tQLzeZ=9eRL7`%i?evnB;+ZSXbttwt=)(s_N*Sh zv>lToUANbQH(t-#HG{w)W(h0OSzwj0J;1pA8_GXe0`?iZId|{z-SCsrXiG$Ir@PEP zVU$<~NElsrd**e7Rj*m zJy+{y&6?Kn0u^|^JY-BIYFc`Hx7DZZd z!CRb3SBlGJqpR;~$?(=nA0IOf;0DnP2_NQ0k~cWbN4Ng<&wwLCMe zyBe^j1GYszvSAnF^!v$J2U2-Rju{+i_8t!OwDjsQPOgkjHOl*x4_Z4_G8K zs$3mw5a1%MK&OeI0<=nzX;@IB{2sHIk`mFZsW+n>ElDjo)SLWkRKS`!6qUDunVnih1qlunE!b^)j71xSR51PWiqupcWy z-{H56hMjh@Yq&ll#n+b%MUt1uKMx#s&>p0}MMe1bIjVby8BL(VpMeSwG&7yeA5`lZ zJL??0JUnPt{>Wx5Iqwp3uo^ICB|kksT3)t!Uh3f9`q#`T^>!UO6h_5%7Q~if7`LQQ zqA<$3foDqVgeUy=gvh8p^R_~j=z``60sket=Js~#1!{wB(na|bBU1h)8$C0|J%u)o zBN2Ai4I(R}1tXEoa9c{7@N7SH#NYv19#5+PqUXXOY^h#OdBKw-Vmk##IkZzr*z?72 z-MWQJ6Q**zK7jmuZOhk8Pn_W-J!#z4v7zU@yWG~G{Uu6le|&z90yDg}lWJSAv07#M zdMH-%!xLJhEBjE;A^RRBXSD@A=EkI1o(iWk;sndMU#C~97vl%yaf;&og!lx}KC zQhXB;amZ%L)0j%!>9W7X}X!x zof33PAl>UzvRo(}mCbPTs$iP~${cuqzk>@*;xTjtgtuLl=ePF^?i+zy?Z`80>6sRi=@?-VuFmn`=l=P+Fec*)7PUp+J zsTz6pFp#U4FJB(|rE*if@W0a=vnI(sDk>|;#qLv3 zQql+r2vjPoYTZM9WR4921gf0M>oX%gCfI^H;8?=q(}CdZ8ru;;`6@p~zTw&sJ#Tt} zrmyw&7c0GK^v-^Cjm)3u7cK_${K_*H$atK$W8=xmWXTv{4>NwdS?m` z!@{5P{tvtIl+7W01NL;&b*l5d4A*U}YOZg9?$K$F1AEV3ydq=rn15e^Qd8$m3g7by zUbBf`n%R5igZC%~kqwjV-y0};p$MCUXg`mNrVZmb4Y-6tbh%G7B zPEE{&og7+lxD~i`kJO_gGX}4$a%;a3hTZ=*HkPEG)>rxJJ_<;7T<(rEZ~ z#hH?r@aKYqG8-k!>uSn)Lbs&>GA2xqo{@24ZjRG_yUwVmzTP*qE?!PfE5*x2Zs&ORrzu=en_nP7xVN88K_<#UO}<>mP5yCFE4cC;L5WpM-fK9t>BJ{qUH5;`9A&Su%+TjOwaoeOeD4OTiX%%)6MP> z1{UbtWkh~6N%0P1SQ*zIarmU>;J-ikU-{s-l1MkPl$4Za0L-`&XD7lvdsZC#YXU!!80|RIvo8EOn<{b_{=wJYM{^gV0v*w78^*0bN zA2a{ZF#GfjN9`Ug)R*$lzSO-3AERmvpcBT*7~k(&8yb8jkhpvMVDDjp$B`LQbI1Pv z2QK^fxV>3@Lyle4Bm5V-{R>Y|&7QyT^pMmK>Pub zk<(3)k&)@9UAergA?4-ej$7PZT$=0Tv^-KKt*MPL$GhH*l|FP!9R^XXCr=JWTO}nY z`&SHDIKV2ehB4{LzvVnFAE{(&nid_cxg+TN;>ArM)K8*V&z?2;umiL#7NtD{J6F7M zSEi-DN!(9}hlj^LCnw^rD z4jw5PnaYC)S@R1ETAd3E3vU7gTl)=eDk@Ga^=00ElItk$e+>-ppWDTB}oRVTS zI0@fXk33q6e@41FURDvrf{2v*ZA**Nd96aFxVSjvR$(?Of4YCXGbd%ix1Uup-tgiSEs0%_pI@ATKH1>fd!| z=LhM8;}f}&q_^Tku-Q}p@=bhXXaZ&!8n4?ocXfkQ-gCYj0Kd@xCSCC#Mgzk%Z)DEo zXCz>litfKc=l!!Y?@svR%s<>s{f89x|KNj4bNv43D_#^-R0-d|-(Wd?+N37%e9&a+ zoiD^J%*+Wq;9jvPQeR;5;ygX~<(T(r=XF2ux+Wtdq71hY_v)v@L%m7A zRaZB9({L-@ycrGT-el*{cm9ob=8paI2YzW-3gPM$nTNlmS0o->2`_NuJ0 zd;0p8+}-*5Rf>?1koEZSPjT`OA8xYD;75u&7sAqnXH%@GkK4t5sK5 ze`{#4GiqpPP*qn~e-b4vE1TBlTH$`6e*4p>PpBrqroRj?NECJJxud7YAsF=XXu<~@$trEO>ro|4kKcbK62{%jka{H2)jMWvz@tJ-Bx6EUGeJ&OU zU^ijiH5_ty1hcNG(|F=FpNRkK20Ph?wYIc?wZ7ixyb_UKysU0B+1_}h2<|M-oo`Aw zTN}cS9Jv{{E8JO6oVe7c<5~=u#Cj^HjqTz^71ZX>cDj2ky$8PVHl+J7wlgTUeHmSa z`mCVv5Au4W9*wBsGDs`(-`i|&I`~j?UJ3~b5wxFuD`X2$!%g3!hP5B;ywPB#&p6o4 zCtZ8V=o8WDQtDZEpQE_BF_^3Ij_d>>+XHZY%}^k=+NxpBo1_Cc31utVO&&(zcqlha z9F_aBv9d~ojEuKsL|wXkxrdhrt;bhWyjc+OX5>dkeN6}nlTzX=|seq^Xa&)0v^3Slu}?a6QkH zt|)@`YyrJ@L2D_LhYmOIE*8Sd?iqiZ;UfRtO1zL{*t4Z)X=(Yw%d1&r zrF+l^tziG*RaX8#QNP*;+|tx@noRS0;-jNam`@*t>%{tT*lN_yS`@9Ib>b=c$jQ1M zc&q(;f1$QnRWgsuhYlS=X_of(U@r(h7|J><)2ow35_qQ2^!1r(<7t+<6#Kt^O-WCm znwfc_$GTZIRLCwJCfZ#96amx3%Ofl+FTd2|FcN{PnRpE9=|vW5xoP<=-XcFcD<`Ky zk?qujwsg4Ig0S0`*uFNHlNdmWwsxen?FFf>-#?aSCv-;}T94SL|eWUi)LaRuNJe+y&>cZIHsYr8b?u{#-XV z4D~57X-l_hn}!E}zivk2mXni1eBeOw=XN)*nCa zzy&D!$;QE<3d*p^mJsCfY=16h|64<(2o8k7hmNnl!*=xhap%QGNf1cO6*OLpkZuNy z?l4%j{Cxf%s|JzYb~IM@PicIQ;Lo=|t!jvMH8nM})8|mUZK(Qn4<{ie`Z?nv=H_ZC zgFS@Z7rIp7CBjb&mq!i;V*J*lH|$7ZsW6V&RbID(f&y7%0>s_zg|U*QR*j`jfLFrB z8t9citzsZp_;xnK5Y(NZMZX&n8U5&eJAQhY_k4>P$ldpSeuQQ3^{R|Ndmch6X}CUO z;=Va>OhMbp(Ig~cZx1mD5z-$z^ll=fRC}ezE=tg5JQk2yF(@n4@^o=5A51E~( zRT+5bXgt8NLa4@hXGv!#4wT6lYPyn zM9|>mNXNkg};AQ~OnWcdq_{YxAKgcou znG@y*x0pT2Ab+QHdnx11n>TTFb;RR99(A+ZDc=+%mkg}efv%mIh1b%v2nzEKw z#KIOTawt~BG?tKyuD->`$6qWb;n18DM%0fhZ*LiDyHf15(pn!Ov@}=6 zIDsBVj~D^HgOxN|-b4`VnMMX)KrGl5=gqk)^p>;hsAEuX9h-^Pbr87j zR7=I1Hv<9!V!wD%33dJVu7%nT+o7+A;$mY*Q5VOlyuS2b=~p~Pa+;U1hkhp5)Ms9| zS=EQGPP&8_L`Ajl-Mhzb3fmDuukHDuQXZ4qzYC-D7Xis4O3dIi$+-k;-sqsDtXw@{ zk^c)bcDe2!LbYfaE=y-o^5=MOaZz?Lykju4b+ikhsX;d>D+NsGgefQ}ER_!v!U%kp zqkuazS2?v=}VP# z=BR?AH2OUOKE7rzo!)p3Z;mT$D0K8QC@h@^-i)Y)M!am;FghjOy2qxC*Jgd!A=E{n zd+SyRygAhRV5&1m2KA>xxn*EFQLj?e6A+hYuCp}Nr3xramX`Z(yCnLu?s>sK3}juq zrJ<1`$>1P?lGI8z=W87G+@C%L1{N!AMFeU-soUj<)MXH-1QDlYLjcb8y7_i9+HZn_ z_#C_g0w@r7YmB~9pZoOb9^eV%(R`1TR8&w`<@#LM!sDSP>5U-6$V+fc$hP(;Hw_I9 zYNLjt_~Wn=_pRkLC5Y)Ph^eV6WSM?{!bW;}QjkCys_F;d)3kYWuLgxrXvnR5G;F9zVuMt%v;qqMx=IZ>he= zn5L1ZVr^~xx||WrQ|bIQ%-SP0^&F}Svl^~#1PMkSt{jtH2<)%)3&qdG4}bW;^YiCV zqv6^RR9NGnrNk<;+_YzyX-N!~z zAi5tf9Kpms|5t|cFZ3q-rw<-K6v4wH6@K@l7`{UnMj3x&Jo{@bWenl?ME{&mSBLN) z^XZP{hKACj%2;4Kex13dDZ#d#zud3)xrx{?W0q%m`_B z9A#w8XpE6O%g2|ey#wMdKu-qypuQszj3n7HLDb*iz_i{-eFOskCF|p7Z2;3Z3JCns z&ebIn&xXfumaz46te)4+KLSKF3m^4218gjo(Tl#TCPa9c}`wfAQkM7Rd!LexOU?i>axTS?f_kc5JXJ^G0Q{||oJ(B{(k_^YVA9)Qu&^Xk@(a}TVnc%7Cf zYr=W^UBY__G0pv@{Dky0G)ZM;WsWI!r{%&E^S;^rhSb1%TT6@KWczzTz$fzy3mgvT z%NEC)l@$~Q9v%~Ym>)ukoI$=id&>IzbwSYq5bcW-0W@l4WOTsg6fbX%_Rayii}{PC z2e*+927Lo+X3T$K9~f#NW;Mvu-nUxE+W+4xFbB!m09HkwyI&91MTzOeT=i)h$EBba z;_#}j`PWB@WkhW+CIIqk)!A8B#kIBys*(!XPA0<0`Y4Ej8{T#~^53{))pOQM3G4n$ zmtFB+1QLe4-Ttvan2RP+V^5SwW=9M!ZkDUmflOECv(?l7v)y~}yJqRdcZJ}=FKL_|^h@w(4E4)n(MP0V>x zVQ!}T-13tTO(PHR;c``U-w(P^CF~#T^z6y{M_}q$Ru6Ry>b1hj$w^E?(wup_VzK&k zlzrA7FA*=v!?F}eH#9TqF z>qXJ=V5qYx>f|S6H?3rC{o$%N^*dlP=5K&p_S@i9QV&i@^IZ8A-1=ZYcgFy+sjGer zY2Dr3j$5!#hAj9|!&L(fef@aY{%`Nq1XfeWfTo7wi~@2FVg^3S3c6`wq@nEzxoG=n zI6axii4!N36cpag&RXy(J+Jw+WQl$34_(f;npMWaBbR>oy^g_7+r~nhVBSNhOBEVoTdywJK*GSzE9ZpVuP?C|D z_;PWpRBE-iQeJ-5kr>mBz%gEQ86J;vv@!eOSF|nmk#%D}j-;>4C5veIkjo)VK2kWT zvs3-bm4fYOM<^-dz?Kk+DXOU0R&P^nXTKBGab8k>qBlG|Y`i?#vDBFX)+3%0q)6Qx zRuEMFsw1uF02SWsW5AR_E0w{Hra@gh4=Tg|0!ejLRTIoh)~*63o!6%_1@lee+SxoFyH-N=ceOGGa3|JP$=i(5&3KvL?$ z_~Zc@D>X!l@S67u=NL6#9jx+OxfOvNI7e4o6y-&JB2ffgjC?5$3&TdA9uo>WEvIU` zZrLnNw8{G~{_x^pJ99=VM$#{Eam$AWg+B5v-VN7F%lpo@RYRJtbgwj=RBUcl<(KH6sHJE&*F*I#2YJw3+)H##M#g3&!YPHLNV^q=$M2+f1lL2v}W3y-R@#IxDwcDTvXQM&|)qPquT8Ostgd6mxtl zh?T(^dpDvUOJAsT441}npdA(N-%mw7sLV^ZS_Rxo={4+H*m7zI*MTi3>y)}UuKf(6 z>)!6RYA_loe};a`fp>UCQqttJ=b_}(RO@9LF_(|1n;)zM>H&!gcJ=pw`kVpHkLWr^ z=7H2F=st8pBgzR!aq07#`QHbZwqEiX7r%Y`w(O@`pKX^ZLrTi=>BOAELUok$>#{YW z!^eNs)HKbkbjudGpHT1wP4mtdVv+f;CmE8)?tE03AVdkjHI~IQB^%x9UQ49u#0bak zCne=^g*#|qwvV` zNlD*el{EDb^fvP!*9`hP`f%!xet)O`Er2yLYL?P4wUuBm8MU-NK-hZIk%lNfV4XY} zALn0iR=WD82^sTNiutB6XCs)tQCKK=U6Yp z(^bU%%0D*O7LcF8&@fTP2{z(Fn>^70+E{@5{HixEnsq0lHaS3`{bK=k<62tIfZ14E zgE6q2I)%>M={+02A9f%iXy^D7Cr{#v?94^hM`PGslXCa?;MA3sS)|#!2nYx$>FFmo z4F))>T94&ue{b@e@Ix^TD%YMc&>BNO%RvPa>jnZwnq z?kI7Iqja-L=nGG|R*tqso0eS?EK4DBN-uYrJ0*CYGQO53tDvTq`1!NMm(QR1ySTWx z0L;h1*z@WO3c8gi2w$I)y?OU;tXavLK58UjV`DQNp1!otP4)TmRM%NvUS-reA3X`| zS@B#YUFUl}42=Dv-@%2VcnJNfq^iB+)T>-gP9@@1G0D&_rl6xsK}iX~0+Q3y zTT#q3D~oGHjKozScREpd-fwVSkIMHl+*M|S+)v@wtq?)GY0a7J2q8P$vF7j8V%;7Y z1u3nt%XT**>51 zw}Ie})6Vu5i!@j=OQWhcw+Uz2nqnl$QHlwQR1Z{mpu+@`f!JAES_=Lpc?K2~b@N5u zE!5l^upz2#W(NDNj>F;5Gi&k@xm%&qlCg+ey2#ARDh=c*-A*+rLn~QD8Zq{Cts)fU z13rza!%!zzt0J*uU1!sBj)NK6z5z%wc0I*fO!#VJVNHcFfO^2dtl(;6+_$Bddx?kyGg74F<@_OCT zxkRY#ZVd09=ZM8SYlNA9(Tnc_UE2nTR{I1L^8?H*el- ziuScqG&X$!xCr&Z{dv;^&=k@P0QJoxtgJx9-2(yw&hYRoM*I2v#=Sb+d&ItceQBbu z&}O2RO6TG2i}%}g;dcWfGJ3=0loTJ%zCM}`P)FI8gI5>8rUk3&IOCqWyN6!!Ix0{V z45(`iKnClXGZ$d|!?|5%-0*(5L4U6(-=eaKTiTv7cjX2ve-ZFLj|bT?F0du(W4M%{ zg0c88b;3!pU{yGGaYk{Iy2iX^T52i}3kyqpZ0u(# zE)pts^viHcMhRCX^}=L!L#U_cK9I1PU;uQ%QCpJ+@vw_32Q2ujzSh?AfJxv5TM(a_ z`5n|lRkvl;vHB@MV5ivFzHLq#h;chDPjb$DT=3mbPR>Vi^j!Y9vGY_;TP>>B18ru2 zS{*NJE;Q^Kvu7k!>so1m#NV*8vT~w{O?Bp2`KsQ!W#qlKZMFg`bK!N==4T2js<6k8 z@te~$__&RlVpGyv!59PsP!2_7EZv$iA4VmB0|Nv4^F!6zX-meE*vA*=bV|B6E5tUR z>?b3;2$DS9w>{N&-NB*2Y6D1A?bb@4IY6!JU}g$Hsn(&?wjjyN;UZ40*F$-#0f{s| zefBIwXga?lATV$}I{d1ZQ89pHeL%wTuy;;@(pnsCthgR2+z~El^G%gS3s7Fk>JP7Y z&>`_Q6Rq)1E7SUb!s3&YzuhAR<*BM}A|skk$7jalKVCdtG$X24yW-bq53(%_dkzpU z=Vfw%_~2D=gEHcL{`~pyGb-&1uwT*2Qt5gJF_M!b>iUEB zpLcDj8No6V@7rhLKuARNZNOsH(HQHu+@zq#CaCo&4jp=jdI_UBXBPty9`7#5CO06$ zdJiGbty>-n!2w}Ji7C-BG0q~#uvtMlJc3L)1%+3SA3tsZ?WiM9h&_Ca?JE<6y@cRI zf?PyPx|UFmNt;|!Qc{qJ97sYSyVm$48hkec%L7c=6M{&0D2AD?7|ck1GJP zUSAxG5p`ZiQJvS!sF)fMl_KbVUa*=`I07ulD@+_y`{~6|as$Q$Tz%ZME6+R%d946` z5boK7x;UfePbiZ*N%1r~ZKIM}Sw-cVKdKD{sF#VquWSbWkL}8pE66gIx@?V`I_+$& z0Fy}^8oIk%Nxr?c4m%H*AkT^Qx#RNIl^XFyvlF`pDwczz&e*Pbl!G?GMd;tKE8p