Some initial "optimized" guesses looks very different from posteriors

[avivajpeyi]

Hmm i wish I had added the TOI number in this...

I am a bit concerned that the initial fit is so different from the posterior for these TOIs.

Check out this pdf for ~1.5K phase plots

[avivajpeyi]

Its bizarre that the initial fit is so off. Some things to check:

• is this the optimised initial guess, vs that from Spoc?
• is this just some plotting error?
• does change the order of optimization help?
• maybe store the guess after each stage of optimization -- plot it on top here.

[avivajpeyi]

Focusing on TOI 2164

NOTE: initial fit == initial optimized parameters' fit

The initial fit is not in the 1,2,3-sigma of the posteriors.

initial fit plotted on the data

The signal has an S/R~ 12 --> maybe this is the issue?

Plot of the initial fit on the prior:

Initial fit in orange

What is going on with the noise params??

[avivajpeyi]

Plotting histograms of posterior + prior samples with initial fit

Ok, something really strange with noise parameters.... how are we getting a posterior outside our prior bounds?

Ok, I guess the tmax and Rp/Rstar results could make sense -- the posterior is inside the prior, the initial guess is kind of close to the posterior? But maybe our initial fit can be better?

[avivajpeyi]

Optimization code + logs

theta = model.test_point
theta1 = pmx.optimize(theta, [jitter], **kwargs) # 1 noise param
theta2 = pmx.optimize(theta1, [b, r, dur], **kwargs) # impact param, Rp/Rs, transit duration
theta3 = pmx.optimize(theta2, [rho, sigma, jitter], **kwargs) # GP rho, GP sigma, jitter
theta4 = pmx.optimize(theta3, [u, f0], **kwargs) # limb darkening + baseline flux
theta5 = pmx.optimize(theta4,  [tmax_1], **kwargs) # transit timing

2022-06-08 23:42:34,204 - TESS-ATLAS - INFO - Optimizing sampling starting point
optimizing logp for variables: [jitter]
logp: -40938.29 -> -39062.89

optimizing logp for variables: [b, r, dur]
logp: -39062.89 -> -38961.56

optimizing logp for variables: [rho, sigma, jitter]
logp: -38961.56 -> -38839.13

optimizing logp for variables: [u, f0]
logp: -38839.13 -> -38838.41

optimizing logp for variables: [tmax_1]
logp: -38838.41 -> -38818.97

2022-06-08 23:45:24,923 - TESS-ATLAS - INFO - Optimization complete! (logp: -40928.32 -> -38742.42)

Looking at the theta5 'optimized' fit, it looks like

1. t0 is correct,
2. f0 is correct,
3. transit duration is incorrect
4. transit depth is incorrect

(code to generate plot init_lc_plotter.py.zip)

[avivajpeyi]

• To help with transit depth, maybe I should run optimization with Rp/Rs again
• To help with transit duration, maybe I should run optimization with duration, tmax again?

[avivajpeyi]

theta = model.test_point
theta1 = pmx.optimize(theta, [jitter], **kwargs) # 1 noise param
theta2 = pmx.optimize(theta1, [b, r, dur], **kwargs) # impact param, Rp/Rs, transit duration
theta3 = pmx.optimize(theta2, [rho, sigma, jitter], **kwargs) # GP rho, GP sigma, jitter
theta4 = pmx.optimize(theta3, [u, f0], **kwargs) # limb darkening + baseline flux
theta5 = pmx.optimize(theta4,  [tmax_1], **kwargs) # transit timing
theta6 = pmx.optimize(theta5, planet_params[0], **kwargs) # Rp/Rs
theta7 = pmx.optimize(theta6, [planet_params[1], period_params[0]], **kwargs) # duration, tmax
theta8 = pmx.optimize(theta7, [*planet_params, *noise_params, *period_params], **kwargs) # all params

A bit better! logp: -40928.32 -> -38714.38

@dfm -- do you have any thoughts on this?

[avivajpeyi]

Fixed by looping over optimization step