# Reparameterization methods for MC gradient estimation

April 4, 2018 — May 2, 2023

approximation
Bayes
density
likelihood free
Monte Carlo
nonparametric
optimization
probabilistic algorithms
probability
sciml
statistics

Reparameterization trick. A trick where we cleverly transform RVs to sample from tricky target distributions, and their jacobians, via a “nice” nice source distribution. Useful in e.g. variational inference, especially autoencoders, for density estimation in probabilistic deep learning. Pairs well with normalizing flows to get powerful target distributions. Storchastic credits pathwise gradients to Glasserman and Ho (1991) as perturbation analysis. The name comes from Diederik P. Kingma and Welling (2014). According to Bloem-Reddy and Teh (2020) the reparameterisation trick is an application of noise outsourcing.

## 1 Tutorials

The classic turorial is by Shakir Mohamed, Machine Learning Trick of the Day (4): Reparameterisation Tricks:

Suppose we want the gradient of an expectation of a smooth function $$f$$: $\nabla_\theta \mathbb {E}_{p(z; \theta)}[f (z)]=\nabla_\theta \int p(z; \theta) f (z) d z$ […] This gradient is often difficult to compute because the integral is typically unknown and the parameters $$\theta,$$ with respect to which we are computing the gradient, are of the distribution $$p(z; \theta).$$

Now we suppose that we know some function $$g$$ such that for some easy distribution $$p(\epsilon),$$ $$z | \theta=g(\epsilon, \theta)$$. Now we can try to estimate the gradient of the expectation by Monte Carlo:

$\nabla_\theta \mathbb {E}_{p(z; \theta)}[f (z)]=\mathbb {E}_{p (c)}\left[\nabla_\theta f(g(\epsilon, \theta))\right]$ Let’s derive this expression and explore the implications of it for our optimisation problem. One-liners give us a transformation from a distribution $$p(\epsilon)$$ to another $$p (z)$$, thus the differential area (mass of the distribution) is invariant under the change of variables. This property implies that: $p (z)=\left|\frac{d \epsilon}{d z}\right|-p(\epsilon) \Longrightarrow-p (z) d z|=|p(\epsilon) d \epsilon|$ Re-expressing the troublesome stochastic optimisation problem using random variate reparameterisation, we find: \begin{aligned} \nabla_\theta \mathbb {E}_{p(z; \theta)}[f (z)] &=\nabla_\theta \int p(z; \theta) f (z) d z \\ &= \nabla_\theta \int p(\epsilon) f (z) d \epsilon\\ &=\nabla_\theta \int p(\epsilon) f(g(\epsilon, \theta)) d \epsilon \\ &=\nabla_\theta \mathbb {E}_{p (c)}[f(g(\epsilon, \theta))]\\ &=\mathbb {E}_{p (e)}\left[\nabla_\theta f(g(\epsilon, \theta))\right] \end{aligned}

That is a classic; but there are some more pedagogic tutorials these days:

Yuge Shi’s variational inference tutorial is a tour of cunning reparameterisation gradient tricks written for her paper Shi et al. (2019). She punts some details to Mohamed et al. (2020) which in turn tells me that this adventure continues at Monte Carlo gradient estimation, .

## 2 Variational Autoencoder

Diederik P. Kingma and Welling (2014) introduces both the reparameterization trick (in its current name) and the variational autoencoder as an aside. Nifty.

## 3 Normalizing flows

Reparameterizaton at maximum elaboration. Cunning reparameterization maps with desirable properties for nonparametric density inference. See normalizing flows.

## 4 General measure transport

See transport maps.

## 6 Incoming

Universal representation theorems? Probably many, here are some I saw: Perekrestenko, Müller, and Bölcskei (2020); Perekrestenko, Eberhard, and Bölcskei (2021).

We show that every $$d$$-dimensional probability distribution of bounded support can be generated through deep ReLU networks out of a 1-dimensional uniform input distribution. What is more, this is possible without incurring a cost-in terms of approximation error measured in Wasserstein-distance-relative to generating the $$d$$-dimensional target distribution from $$d$$ independent random variables. This is enabled by a vast generalization of the space-filling approach discovered in [2]. The construction we propose elicits the importance of network depth in driving the Wasserstein distance between the target distribution and its neural network approximation to zero. Finally, we find that, for histogram target distributions, the number of bits needed to encode the corresponding generative network equals the fundamental limit for encoding probability distributions as dictated by quantization theory.

## 7 References

Albergo, Goldstein, Boffi, et al. 2023.
Albergo, and Vanden-Eijnden. 2023. In.
Ambrosio, Gigli, and Savare. 2008. Gradient Flows: In Metric Spaces and in the Space of Probability Measures. Lectures in Mathematics. ETH Zürich.
Bamler, and Mandt. 2017. arXiv:1707.01069 [Cs, Stat].
Bloem-Reddy, and Teh. 2020.
Caterini, Doucet, and Sejdinovic. 2018. In Advances in Neural Information Processing Systems.
Charpentier, Borchert, Zügner, et al. 2022. arXiv:2105.04471 [Cs, Stat].
Chen, Changyou, Li, Chen, et al. 2017. arXiv:1709.01179 [Stat].
Chen, Tian Qi, Rubanova, Bettencourt, et al. 2018. In Advances in Neural Information Processing Systems 31.
Cunningham, Zabounidis, Agrawal, et al. 2020.
Devroye. 2006. In Simulation. Handbooks in Operations Research and Management Science.
Dinh, Sohl-Dickstein, and Bengio. 2016. In Advances In Neural Information Processing Systems.
Figurnov, Mohamed, and Mnih. 2018. In Advances in Neural Information Processing Systems 31.
Germain, Gregor, Murray, et al. 2015.
Glasserman, and Ho. 1991. Gradient Estimation Via Perturbation Analysis.
Grathwohl, Chen, Bettencourt, et al. 2018. arXiv:1810.01367 [Cs, Stat].
Huang, Krueger, Lacoste, et al. 2018. arXiv:1804.00779 [Cs, Stat].
Jankowiak, and Obermeyer. 2018. In International Conference on Machine Learning.
Kingma, Durk P, and Dhariwal. 2018. In Advances in Neural Information Processing Systems 31.
Kingma, Diederik P., Salimans, Jozefowicz, et al. 2016. In Advances in Neural Information Processing Systems 29.
Kingma, Diederik P., Salimans, and Welling. 2015. In Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 2. NIPS’15.
Kingma, Diederik P., and Welling. 2014. In ICLR 2014 Conference.
Koehler, Mehta, and Risteski. 2020. arXiv:2010.01155 [Cs, Stat].
Lin, Khan, and Schmidt. 2019. arXiv:1910.13398 [Cs, Stat].
Lipman, Chen, Ben-Hamu, et al. 2023.
Louizos, and Welling. 2017. In PMLR.
Lu, and Huang. 2020. In Advances in Neural Information Processing Systems.
Marzouk, Moselhy, Parno, et al. 2016. In Handbook of Uncertainty Quantification.
Massaroli, Poli, Bin, et al. 2020. arXiv:2003.08063 [Cs, Math, Stat].
Maurais, and Marzouk. 2024.
Mohamed, Rosca, Figurnov, et al. 2020. Journal of Machine Learning Research.
Ng, and Zammit-Mangion. 2020. arXiv:2007.00248 [Stat].
Papamakarios. 2019.
Papamakarios, Murray, and Pavlakou. 2017. In Advances in Neural Information Processing Systems 30.
Papamakarios, Nalisnick, Rezende, et al. 2021. Journal of Machine Learning Research.
Perekrestenko, Eberhard, and Bölcskei. 2021. Partial Differential Equations and Applications.
Perekrestenko, Müller, and Bölcskei. 2020.
Pfau, and Rezende. 2020. “Integrable Nonparametric Flows.” In.
Ran, and Hu. 2017. Neural Computation.
Rezende, and Mohamed. 2015. In International Conference on Machine Learning. ICML’15.
Rezende, Mohamed, and Wierstra. 2015. In Proceedings of ICML.
Rippel, and Adams. 2013. arXiv:1302.5125 [Cs, Stat].
Ruiz, Titsias, and Blei. 2016. In Advances In Neural Information Processing Systems.
Shi, Siddharth, Paige, et al. 2019. arXiv:1911.03393 [Cs, Stat].
Spantini, Baptista, and Marzouk. 2022. SIAM Review.
Spantini, Bigoni, and Marzouk. 2017. Journal of Machine Learning Research.
Tabak, E. G., and Turner. 2013. Communications on Pure and Applied Mathematics.
Tabak, Esteban G., and Vanden-Eijnden. 2010. Communications in Mathematical Sciences.
van den Berg, Hasenclever, Tomczak, et al. 2018. In UAI18.
Wang, and Wang. 2019. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers).
Wehenkel, and Louppe. 2021. In Proceedings of The 24th International Conference on Artificial Intelligence and Statistics.
Xu, Zuheng, Chen, and Campbell. 2023.
Xu, Ming, Quiroz, Kohn, et al. 2019. In Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics.
Yang, Li, and Wang. 2021. arXiv:2101.12353 [Cs, Math, Stat].
Zahm, Constantine, Prieur, et al. 2018. arXiv:1801.07922 [Math].
Zhang, and Curtis. 2021. Journal of Geophysical Research: Solid Earth.