Keep the gradient flowing

Policy Gradients Part 1: The REINFORCE Estimator

I have a dirty secret. Well, I actually have many. But one of them is that I never understood the basic algorithms behind reinforcement learning. So I plan to remedy this with a series of blog posts, where I will cover foundational RL algorithms, from REINFORCE to the frontier.

Here's the first one, where I cover one of the oldest practical gradient estimators for reinforcement learning, the REINFORCE estimator, and show how it can get away without differentiating through the environment.

$$ \require{bbox} \def\EE{\mathbb E} \def\RR{\mathbb{R}} \DeclareMathOperator*{\argmin}{{arg\,min}} \DeclareMathOperator*{\argmax}{{arg\,max}} \def\defas{\stackrel{\text{def}}{=}} \def\dif{\mathop{}\!\mathrm{d}} \definecolor{colorpolicy}{RGB}{102,217,178} \definecolor{colorreturn}{RGB}{217,95,2} \definecolor{colordynamics}{RGB}{176,170,224} $$

Setup

This is the notation I'll follow throughout the series. An agent interacts with an environment by observing states $s \in \mathcal{S}$, taking actions $a \in \mathcal{A}$, receiving rewards $r(s, a) \in \RR$, and transitioning to a new state $s'$ according to the environment's transition probability distribution ${\color{colordynamics}p(s' | s, a)}$. We write the initial state distribution over $s_0 \in \mathcal{S}$ simply as ${\color{colordynamics}p(s_0)}$.

The agent's behavior is determined by a policy ${\color{colorpolicy}\pi_\theta(a | s)}$, which maps states to distributions over actions. This policy has parameters $\theta \in \RR^d$, and we'd like to find parameters that maximize the expected return: \begin{equation}\label{eq:objective} J(\theta) \defas \EE_{\tau \sim {\color{colorpolicy}\pi_\theta}}\left[ {\color{colorreturn}R(\tau)} \right] = \int p(\tau | \theta) \, {\color{colorreturn}R(\tau)} \, \dif\tau\,, \end{equation} where ${\color{colorreturn}R(\tau)} = \sum_{t=0}^{T-1} {\color{colorreturn}r_t}$ is the return of a trajectory $\tau = (s_0, a_0, s_1, a_1, \ldots, s_{T-1}, a_{T-1}, s_T)$, with per-step reward ${\color{colorreturn}r_t} \defas r(s_t, a_t)$, $T$ is the trajectory length or horizon (which may be fixed or determined by the environment when reaching a terminal state), and $p(\tau | \theta)$ is the probability of observing that trajectory $\tau$ under policy ${\color{colorpolicy}\pi_\theta}$. For simplicity, we focus here on undiscounted finite-horizon returns, though all derivations extend directly to discounted returns ${\color{colorreturn}R(\tau)} = \sum_{t=0}^{T-1} \gamma^t {\color{colorreturn}r_t}$ with discount factor $\gamma \in [0, 1)$. Throughout, we color-code the key quantities: ${\color{colorpolicy}\text{green}}$ for the policy, ${\color{colorreturn}\text{orange}}$ for the return, and ${\color{colordynamics}\text{purple}}$ for the environment dynamics.

When Things go South

Suppose we want to maximize the expected return $J(\theta)$ using gradient ascent. Writing out the gradient of the expected return gives: \begin{equation}\label{eq:naive_grad} \nabla_\theta J(\theta) = \nabla_\theta \int p(\tau | \theta) {\color{colorreturn}R(\tau)} \dif\tau = \int \nabla_\theta p(\tau | \theta) {\color{colorreturn}R(\tau)} \dif\tau\,. \end{equation} To evaluate this integral, we need to compute the gradient of the trajectory probability $\nabla_\theta p(\tau | \theta)$. Using the chain rule of probability, the probability of a trajectory $\tau$ under policy ${\color{colorpolicy}\pi_\theta}$ is a product of initial state probabilities, policy choices, and environment transitions: \begin{equation}\label{eq:trajectory_prob} p(\tau | \theta) = {\color{colordynamics}p(s_0)} \prod_{t=0}^{T-1} {\color{colorpolicy}\pi_\theta(a_t | s_t)} \, {\color{colordynamics}p(s_{t+1} | s_t, a_t)}\,. \end{equation}

Now we hit a wall. To compute $\nabla_\theta p(\tau | \theta)$, we would need to differentiate this product with respect to $\theta$. But the product contains the environment's transition dynamics ${\color{colordynamics}p(s_{t+1} | s_t, a_t)}$, which are typically a black box: a robot arm colliding with an object involves discontinuous contact forces, and a board game advances by discrete rules. In either case, we do not have analytical or differentiable access to the transition function, so we cannot compute $\nabla_\theta p(\tau | \theta)$ directly.

The Log-Derivative Trick

To move forward, we turn to a classic identity in stochastic optimization known as the log-derivative trick (also known as the likelihood-ratio method or score function estimator). For a broader perspective on score functions and their connection to integration by parts and randomized smoothing, see Francis Bach's excellent blog post. For any differentiable probability density $p_\theta(x) > 0$, we have the identity $\nabla_\theta p_\theta(x) = p_\theta(x) \, \nabla_\theta \log p_\theta(x)$, which follows from the chain rule: \begin{equation}\label{eq:log_derivative_trick} \nabla_\theta \log p_\theta(x) = \frac{\nabla_\theta p_\theta(x)}{p_\theta(x)}\,. \end{equation}

This identity lets us rewrite the problematic gradient $\nabla_\theta p(\tau | \theta)$ in terms of $\nabla_\theta \log p(\tau | \theta)$, converting the integral into an expectation: \begin{align} \nabla_\theta J(\theta) &\stackrel{\eqref{eq:naive_grad}}{=} \int \left( \nabla_\theta p(\tau | \theta) \right) {\color{colorreturn}R(\tau)} \dif\tau \nonumber \\ &\stackrel{\eqref{eq:log_derivative_trick}}{=} \int p(\tau | \theta) \nabla_\theta \log p(\tau | \theta) {\color{colorreturn}R(\tau)} \dif\tau \nonumber \\ &\stackrel{\eqref{eq:objective}}{=} \EE_{\tau \sim {\color{colorpolicy}\pi_\theta}}\left[ {\color{colorreturn}R(\tau)} \nabla_\theta \log p(\tau | \theta) \right]\,. \label{eq:rl_score_sub} \end{align} We have traded the intractable $\nabla_\theta p(\tau | \theta)$ for $\nabla_\theta \log p(\tau | \theta)$ inside an expectation. But is this any better?

The World Drops Out

At first glance, it doesn't seem like we have gained much. While we have rewritten the gradient of the expected return as an expectation, we are still left with computing the gradient of the log-trajectory probability $\nabla_\theta \log p(\tau | \theta)$, which does not appear to be any easier to compute. The key to resolving this comes from expanding that logarithm. Taking the logarithm of the trajectory probability $\eqref{eq:trajectory_prob}$ turns the product into a sum: \begin{equation}\label{eq:log_trajectory_prob} \log p(\tau | \theta) = \log {\color{colordynamics}p(s_0)} + \sum_{t=0}^{T-1} \log {\color{colorpolicy}\pi_\theta(a_t | s_t)} + \sum_{t=0}^{T-1} \log {\color{colordynamics}p(s_{t+1} | s_t, a_t)}\,. \end{equation} This is crucial: turning a multiplicative structure into an additive one makes differentiation much easier. Taking the gradient with respect to $\theta$ yields: \begin{equation}\label{eq:grad_log_trajectory} \nabla_\theta \log p(\tau | \theta) = \underbrace{\nabla_\theta \log {\color{colordynamics}p(s_0)}}_{= \, 0} + \sum_{t=0}^{T-1} \nabla_\theta \log {\color{colorpolicy}\pi_\theta(a_t | s_t)} + \underbrace{\sum_{t=0}^{T-1} \nabla_\theta \log {\color{colordynamics}p(s_{t+1} | s_t, a_t)}}_{= \, 0}\,. \end{equation} The environment dynamics ${\color{colordynamics}p(s_{t+1} | s_t, a_t)}$ describe the fixed laws of the environment, not the agent's policy, and so their derivatives with respect to $\theta$ are identically zero. Likewise, the initial state distribution ${\color{colordynamics}p(s_0)}$ does not depend on $\theta$. The unknown environment dynamics drop out entirely, leaving only our differentiable policy: \begin{equation}\label{eq:grad_log_simplified} \nabla_\theta \log p(\tau | \theta) = \sum_{t=0}^{T-1} \nabla_\theta \log {\color{colorpolicy}\pi_\theta(a_t | s_t)} \,. \end{equation}

The REINFORCE Estimator

Substituting $\eqref{eq:grad_log_simplified}$ back into $\eqref{eq:rl_score_sub}$ gives our fundamental result, the REINFORCE estimator:Pronounced as it's written, but shouting.

Let $\hat{g}(\tau)$ be the REINFORCE estimator for a trajectory $\tau$, defined as: \begin{equation}\label{eq:reinforce} \hat{g}(\tau) = \left(\sum_{t=0}^{T-1} \nabla_\theta \log {\color{colorpolicy}\pi_\theta(a_t | s_t)}\right) {\color{colorreturn}R(\tau)}\,. \end{equation} This is an unbiased estimator of the expected return gradient, that is, \begin{equation}\label{eq:reinforce_expectation} \EE_{\tau \sim {\color{colorpolicy}\pi_\theta}} [\hat{g}(\tau)] = \nabla_\theta J(\theta)\,. \end{equation} This allows us to optimize expected return by sampling trajectories from the environment, without needing to differentiate through its transition dynamics.

The REINFORCE gradient estimator is commonly attributed to Williams (1992), Fun historical detail: REINFORCE is actually a backronym coined by Williams for REward Increment = Nonnegative Factor × Offset Reinforcement × Characteristic Eligibility, referring to the structure of the update rule. although the log-derivative trick was known well before that. In optimal control and simulation optimization, this general paradigm of directly parameterizing the control law is known as approximation in policy space, and the resulting estimator is referred to as the likelihood ratio method (see Chapter 3 of Bertsekas (2019) for a comprehensive exposition).

Notice what has happened. The gradient $\nabla_\theta J(\theta)$ is expressed as the expected value of the estimator $\hat{g}(\tau)$, which involves only the policy's log-probabilities and the return. The transition dynamics ${\color{colordynamics}p(s_{t+1} | s_t, a_t)}$ appear nowhere in the expression. They are implicit: they determine which trajectories are likely when we sample $\tau \sim {\color{colorpolicy}\pi_\theta}$, but we never need to know their functional form or differentiate through them. This is the key insight: the log-derivative trick converts "differentiate the distribution" into "evaluate a score function under the distribution".

Variance

An unbiased estimator is a good start, but that's still far from practical. Ideally, this estimator should have low variance. Under standard assumptions, we can show that the variance of the REINFORCE gradient estimator is bounded as $O(T^3)$, and scales as $\Theta(T^3)$ whenever the expected per-step reward has a non-zero mean.

We'll make the following assumptions. First, per-step rewards are bounded by a constant ${\color{colorreturn}r_{\max}}$, that is, $|{\color{colorreturn}r_t}| \le {\color{colorreturn}r_{\max}}$ for all $t$. Second, the second moment of the per-step score is bounded, i.e., $\EE[\|\nabla_\theta \log {\color{colorpolicy}\pi_\theta(a_t | s_t)}\|^2] \le C$ for some constant $C > 0$ and all $t$.This assumption holds for standard policies: for softmax policies, whenever the logit gradients (network Jacobian) are bounded; for Gaussian policies $\mathcal{N}(\mu_\theta(s), \sigma^2 I)$, whenever the mean network has bounded gradients and the variance $\sigma^2$ is bounded away from zero.

Under the assumptions above, the total variance $\text{Var}(\hat{g}) \defas \EE[\|\hat{g}(\tau) - \EE[\hat{g}(\tau)]\|^2]$ of the REINFORCE estimator is bounded as: \begin{equation}\label{eq:variance_bound} \text{Var}(\hat{g}) \le T^3 {\color{colorreturn}r_{\max}}^2 C = O(T^3)\,. \end{equation}

To derive this bound, let's write the estimator as a product: $\hat{g}(\tau) = {\color{colorpolicy}S(\tau)} {{\color{colorreturn}R(\tau)}}$, where ${\color{colorpolicy}S(\tau)} = \sum_{t=0}^{T-1} \nabla_\theta \log {\color{colorpolicy}\pi_\theta(a_t | s_t)}$ is the trajectory score. Using the definition of total variance $\text{Var}(\hat{g}) \defas \EE[\|\hat{g} - \EE[\hat{g}]\|^2] = \EE[\|\hat{g}\|^2] - \|\EE[\hat{g}]\|^2$, we have \begin{equation}\label{eq:var_first_step} \text{Var}(\hat{g}) = \EE[\|{\color{colorpolicy}S(\tau)}\|^2 \, {{\color{colorreturn}R(\tau)}}^2] - \|\nabla_\theta J(\theta)\|^2 \leq \EE[\|{\color{colorpolicy}S(\tau)}\|^2 \, {{\color{colorreturn}R(\tau)}}^2]\,, \end{equation} where in the last inequality we have dropped the non-negative term $\|\nabla_\theta J(\theta)\|^2$.Since the gradient vanishes at a stationary point, this bound is tight. Let's now bound each of the two terms in the remaining expectation.

1️⃣ The score $\|{\color{colorpolicy}S(\tau)}\|^2$ scales as $O(T)$. The score ${\color{colorpolicy}S(\tau)} = \sum_{t=0}^{T-1} \nabla_\theta \log {\color{colorpolicy}\pi_\theta(a_t | s_t)}$ is a sum of $T$ per-step score vectors $X_t \defas \nabla_\theta \log {\color{colorpolicy}\pi_\theta(a_t | s_t)} \in \RR^d$. Crucially, conditional on the state $s_t$, the expectation of the action score is zero: \begin{align*} \EE_{a_t \sim {\color{colorpolicy}\pi_\theta(\cdot | s_t)}} [X_t \mid s_t] &= \int \left( \nabla_\theta \log {\color{colorpolicy}\pi_\theta(a_t | s_t)} \right) {\color{colorpolicy}\pi_\theta(a_t | s_t)} \dif a_t \\ &\stackrel{\eqref{eq:log_derivative_trick}}{=} \int \nabla_\theta {\color{colorpolicy}\pi_\theta(a_t | s_t)} \dif a_t \\ &= \nabla_\theta \underbrace{\int {\color{colorpolicy}\pi_\theta(a_t | s_t)} \dif a_t}_{=\, 1} = 0\,, \end{align*} where we've used the fact that ${\color{colorpolicy}\pi_\theta}$ is a probability distribution that integrates to 1. Now consider any two time steps $t < t'$. By the law of total expectation: \begin{equation} \EE[X_t^\top X_{t'}]=\EE\Big[\EE[X_t^\top X_{t'} \mid s_0, a_0, \ldots, s_{t'}]\Big]\,. \end{equation}

Inside the inner expectation, the only remaining randomness is the action $a_{t'}$ (by conditioning, everything up to $s_{t'}$ is fixed). Since $t < t'$, $X_t$ is fully determined by the conditioning set and can be pulled out of the inner expectation. Furthermore, since $X_{t'}$ depends only on $s_{t'}$ and $a_{t'}$, and $a_{t'}$ is drawn from ${\color{colorpolicy}\pi_\theta(\cdot \mid s_{t'})}$, we have $\EE[X_{t'} \mid s_0, a_0, \ldots, s_{t'}]=\EE[X_{t'} \mid s_{t'}]=0$, giving us: \begin{equation} \EE[X_t^\top X_{t'}]=\EE\Big[X_t^\top \underbrace{\EE[X_{t'} \mid s_{t'}]}_{=\, 0}\Big]=0\,. \end{equation} Therefore, score vectors across different steps are uncorrelated, and all cross-terms vanish when expanding the squared norm: \begin{equation} \EE[\|{\color{colorpolicy}S(\tau)}\|^2]=\EE\left[\left\|\sum_{t=0}^{T-1} X_t\right\|^2\right]=\sum_{t=0}^{T-1} \EE[\|X_t\|^2]\,. \end{equation} Under the bounded score variance assumption ($\EE[\|X_t\|^2] \leq C$ for all $t$), summing over the $T$ steps yields: \begin{equation}\label{eq:score_bound} \EE[\|{\color{colorpolicy}S(\tau)}\|^2] \le T C = O(T)\,. \end{equation}

2️⃣ The return ${{\color{colorreturn}R(\tau)}}^2$ scales as $O(T^2)$. Using the bounded rewards assumption ($|{\color{colorreturn}r_t}| \leq {\color{colorreturn}r_{\max}}$), we have $|{\color{colorreturn}R(\tau)}| = |\sum_{t=0}^{T-1} {\color{colorreturn}r_t}| \leq T {\color{colorreturn}r_{\max}}$. \begin{equation}\label{eq:return_bound} {{\color{colorreturn}R(\tau)}}^2 \le T^2 {\color{colorreturn}r_{\max}}^2 = O(T^2)\,. \end{equation}

3️⃣ Putting it together. Combining the bounds from \eqref{eq:score_bound} and \eqref{eq:return_bound}, we can factor the upper bound into score variance and return scale factors: \begin{align}\label{eq:variance_split} \text{Var}(\hat{g}) &\stackrel{\eqref{eq:var_first_step}}{\le} \EE[\|{\color{colorpolicy}S(\tau)}\|^2 \, {{\color{colorreturn}R(\tau)}}^2] \nonumber \\ &\stackrel{\eqref{eq:score_bound}, \eqref{eq:return_bound}}{\le} \underbrace{\vphantom{\sup_{\tau} |{\color{colorreturn}R(\tau)}|^2}\EE[\|{\color{colorpolicy}S(\tau)}\|^2]}_{\text{Score Variance } \le T C} \; \times \; \underbrace{\sup_{\tau} |{\color{colorreturn}R(\tau)}|^2}_{\text{Return Scale } \le T^2 {\color{colorreturn}r_{\max}}^2} = T^3 {\color{colorreturn}r_{\max}}^2 C\,, \end{align} which completes the proof.

To put this result in perspective: a cubic dependency on the horizon $T$ is terrible news. Cubic growth means that if you double the length of your episode (say, from $50$ to $100$ steps), the variance of your gradient estimates does not just double: it multiplies by eight ($2^3 = 8$). This makes the estimates incredibly noisy for long horizons.

Empirical Variance Scaling

What we derived in the previous section is an upper bound on the variance. For all we know, this bound could be very loose and the real variance scaling factor could be much smaller. To check how the empirical variance scales in practice, we simulate a simple synthetic MDP. Similar toy environments (such as linear-quadratic control systems and contextual bandits) have been used in the literature for isolating and studying gradient variance.

The setup is as follows: at each time step, the agent observes a state $s_t$ drawn uniformly from $[-1, 1]^4$ and selects one of $|\mathcal{A}| = 4$ actions according to a softmax linear policy $\pi_\theta(a | s) \propto \exp(\theta_a^\top s)$, with parameters $\theta$ drawn from a normal distribution $\mathcal{N}(0, 0.01^2)$. Each step yields a reward ${\color{colorreturn}r_t} = 1 + 0.1 \, \theta_{a_t}^\top s_t$. For each trajectory length $T \in \{5, 10, 20, 40, 80, 160, 320, 640\}$, we estimate $\text{Var}(\hat{g})$ from 30,000 independent trajectories:

Empirical scaling of REINFORCE gradient estimator variance with trajectory length T

The complete simulation script to reproduce this figure is available at gen_reinforce_variance_scaling.py.

The message here is that the $O(T^3)$ upper bound is not wildly pessimistic, but painfully accurate. Fitting a power law on a log-log scale matches the theoretical $O(T^3)$ bound with remarkable precision.

Why does the empirical variance scale tightly as $\Theta(T^3)$? Intuition comes from considering an environment where per-step rewards have a non-zero mean ${\color{colorreturn}\mu_r} = \EE[{\color{colorreturn}r_t}] \neq 0$. In that regime, the return is dominated by its mean, ${\color{colorreturn}R(\tau)} \approx T {\color{colorreturn}\mu_r}$, so the REINFORCE estimator behaves as $\hat{g}(\tau) \approx (T {\color{colorreturn}\mu_r}) {\color{colorpolicy}S(\tau)}$. Squaring $T \mu_r$ multiplies the linear $\Theta(T)$ variance of the uncorrelated score sum ${\color{colorpolicy}S(\tau)}$ by $T^2 \mu_r^2$, yielding a variance that scales strictly as $\Theta(T^3)$.

Summary

The central achievement of the REINFORCE algorithm is that it allows us to optimize policies without knowing or differentiating through the environment's transition dynamics. By applying the log-derivative trick, the product of transition probabilities turns into a sum, and all environment-dependent terms drop out.

However, the price we pay for this model-free elegance is high variance. Specifically, as we derived above, the variance of the REINFORCE estimator scales cubically with the trajectory length ($O(T^3)$). This stems from the combination of a return $R(\tau)$ that grows linearly with $T$ (contributing an $O(T^2)$ term when squared) and a sum of score vectors ${\color{colorpolicy}S(\tau)}$ whose variance also grows linearly as $O(T)$ due to uncorrelated steps. As a result, even a moderate increase in the episode horizon $T$ leads to a massive explosion in variance, requiring a prohibitively large number of samples to obtain stable gradient estimates.

In Part 2: Taming the Variance, we explore core variance reduction techniques (Reward-to-Go, state baselines, and actor-critic bootstrapping) that systematically shrink this variance.

Citing

If you find this blog post useful, please consider citing as:

Policy Gradients Part 1: The REINFORCE Estimator, Fabian Pedregosa, 2026

with bibtex entry:

@misc{pedregosa2026policygradient,
  author = {Pedregosa, Fabian},
  title = {Policy Gradients Part 1: The REINFORCE Estimator},
  year = {2026},
  howpublished = {\url{http://fa.bianp.net/blog/2026/policy-gradient/}},
}
  


References