TEST: add unit test of likelihood reweighting

test/core/result_test.py

@@ -751,5 +751,71 @@ class TestPPPlots(unittest.TestCase):
             )
 
 
+class SimpleGaussianLikelihood(bilby.core.likelihood.Likelihood):
+    def __init__(self, mean=0, sigma=1):
+        """
+        A very simple Gaussian likelihood for testing
+        """
+        from scipy.stats import norm
+        super().__init__(parameters=dict())
+        self.mean = mean
+        self.sigma = sigma
+        self.dist = norm(loc=mean, scale=sigma)
+
+    def log_likelihood(self):
+        return self.dist.logpdf(self.parameters["mu"])
+
+
+class TestReweight(unittest.TestCase):
+
+    def setUp(self):
+        self.priors = bilby.core.prior.PriorDict(dict(
+            mu=bilby.core.prior.TruncatedNormal(0, 1, minimum=-5, maximum=5),
+        ))
+        self.result = bilby.core.result.Result(
+            search_parameter_keys=list(self.priors.keys()),
+            priors=self.priors,
+            posterior=pd.DataFrame(self.priors.sample(1000)),
+            log_evidence=-np.log(10),
+        )
+
+    def _run_reweighting(self, sigma):
+        likelihood_1 = SimpleGaussianLikelihood()
+        likelihood_2 = SimpleGaussianLikelihood(sigma=sigma)
+        original_ln_likelihoods = list()
+        for ii in range(len(self.result.posterior)):
+            likelihood_1.parameters = self.result.posterior.iloc[ii]
+            original_ln_likelihoods.append(likelihood_1.log_likelihood())
+        self.result.posterior["log_prior"] = self.priors.ln_prob(self.result.posterior)
+        self.result.posterior["log_likelihood"] = original_ln_likelihoods
+        self.original_ln_likelihoods = original_ln_likelihoods
+        return bilby.core.result.reweight(
+            self.result, likelihood_1, likelihood_2, verbose_output=True
+        )
+
+    def test_reweight_same_likelihood_weights_1(self):
+        """
+        When the likelihoods are the same, the weights should be 1.
+        """
+        _, weights, _, _, _, _ = self._run_reweighting(sigma=1)
+        self.assertLess(min(abs(weights - 1)), 1e-10)
+
+    def test_reweight_different_likelihood_weights_correct(self):
+        """
+        Test the known case where the target likelihood is a Gaussian with
+        sigma=0.5. The weights can be calculated analytically and the evidence
+        should be close to the original evidence within statistical error.
+        """
+        from scipy.stats import norm
+        new, weights, _, _, _, _ = self._run_reweighting(sigma=0.5)
+        expected_weights = (
+            norm(0, 0.5).pdf(self.result.posterior["mu"])
+            / norm(0, 1).pdf(self.result.posterior["mu"])
+        )
+        self.assertLess(min(abs(weights - expected_weights)), 1e-10)
+        self.assertLess(abs(new.log_evidence - self.result.log_evidence), 0.05)
+        self.assertNotEqual(new.log_evidence, self.result.log_evidence)
+
+
 if __name__ == "__main__":
     unittest.main()