Why Gain Schedule Choice Matters
In SAEM, simulated latent random effects provide stochastic information about the sufficient statistics used in the M-step. The gain \(\gamma_k\) determines how strongly iteration \(k\)'s simulated statistic changes the running statistic. A large gain reacts quickly to new simulation output, but also passes more of its Monte Carlo noise into the update. A small gain smooths that noise, but can make the algorithm slow to respond when the current estimate is still far from a useful region.
Choosing a schedule is therefore a balance, not a search for one universally best formula. A schedule that is appropriate for a long run intended to approach a stationary solution may be less convenient for a short exploratory run. The key distinction is whether the gain continues to shrink toward zero, and, if it does, whether it shrinks at a rate that still allows continued learning.
This tutorial focuses on the consequences of choosing among decreasing, piecewise-decreasing, and constant gains. As in the earlier material on Robbins–Monro gain sequences, the usual sufficient conditions for diminishing-gain stochastic approximation include $$ \sum_{k=1}^{\infty}\gamma_k=\infty \qquad\text{and}\qquad \sum_{k=1}^{\infty}\gamma_k^2<\infty. $$
These conditions capture two competing requirements. The first prevents the total future learning from being finite; the second limits the accumulated effect of simulation noise. They are not, by themselves, a guarantee of SAEM convergence. The model, the conditional simulation, and the stochastic approximation must also satisfy relevant regularity and stability conditions. In particular, a schedule cannot compensate for a poorly mixing latent-effect chain or an unstable parameter update.
Three Common Schedule Choices
A decreasing schedule lowers the gain throughout the run. One common family is \(\gamma_k=k^{-a}\). For \(1/2<a\leq 1\), it satisfies both summability conditions: the series of gains diverges, while the series of squared gains converges. A larger \(a\) makes the gain decline more quickly; this usually reduces late-iteration noise but can suppress useful movement earlier in the run.
A piecewise-decreasing schedule uses different gain rules in different portions of the run. For example, it may use a relatively large gain during an initial segment and then decrease gains according to a power rule. If the initial segment is finite and the later gains satisfy the two summability conditions, the finite initial segment does not change whether either infinite series converges. The eventual decreasing portion still needs to allow enough continued learning.
A constant-gain schedule sets \(\gamma_k=\gamma\) for every iteration, with \(0<\gamma\leq1\). It continues to respond to new simulation output, which can be useful for short runs or for following a changing target. But \(\sum_k\gamma_k^2\) diverges, so a constant gain does not meet the standard diminishing-gain conditions for convergence to a fixed value under persistent Monte Carlo noise. It generally leaves residual variability rather than settling exactly.
| Schedule | Typical effect | Main caution |
|---|---|---|
| Decreasing | Progressively smoother updates and, under suitable conditions, diminishing noise influence | Can become too small before the statistic has moved near its target |
| Piecewise-decreasing | Allows a more responsive early phase followed by increasingly stable updates | The gain change can produce a noticeable shift in update size; the later phase still matters |
| Constant | Maintains responsiveness to new simulation output | Persistent gain leaves a nonzero level of Monte Carlo variability |
Worked Example: Checking a Power-Law Schedule
Suppose a hypothetical SAEM run uses \(\gamma_k=k^{-0.7}\). Check the gain values at iterations 1, 2, 5, and 10, and determine whether the exponent meets the usual summability conditions.
The gains are $$ \gamma_1=1^{-0.7}=1,\quad \gamma_2=2^{-0.7}\approx0.6156,\quad \gamma_5=5^{-0.7}\approx0.3241,\quad \gamma_{10}=10^{-0.7}\approx0.1995. $$ They are positive and decreasing, so later simulated statistics receive less weight than earlier ones.
For a power series, \(\sum k^{-a}\) diverges when \(a\leq1\), and \(\sum k^{-b}\) converges when \(b>1\). Here \(\sum\gamma_k=\sum k^{-0.7}\) diverges because \(0.7\leq1\), while \(\sum\gamma_k^2=\sum k^{-1.4}\) converges because \(1.4>1\). Thus this exponent meets both conditions. This check concerns the gain sequence; it does not establish that the SAEM simulation and parameter updates satisfy all other convergence requirements.
How Gains Change an Update
The update is a weighted movement from the previous statistic toward the newly simulated statistic. Equivalently, it is a weighted average: $$ S_k=(1-\gamma_k)S_{k-1}+\gamma_k\widetilde S_k. $$ A gain near 1 gives the new simulated statistic substantial influence. A smaller gain retains more of the accumulated history. This simple arithmetic explains why large gains often move quickly but yield visibly noisier traces, while small gains tend to smooth traces but move slowly.
Worked Example: Comparing Three Update Paths
Consider hypothetical simulated statistics \(\widetilde S_1=10\), \(\widetilde S_2=6\), and \(\widetilde S_3=8\), with initial value \(S_0=0\). Compare a decreasing schedule \((1,0.5,1/3)\), a piecewise-decreasing schedule \((1,0.6,0.3)\), and a constant schedule with \(\gamma_k=0.4\).
For the decreasing schedule, $$ S_1=0+1(10-0)=10, $$ $$ S_2=10+0.5(6-10)=8, $$ $$ S_3=8+\frac{1}{3}(8-8)=8. $$
For the piecewise-decreasing schedule, $$ S_1=0+1(10-0)=10,\quad S_2=10+0.6(6-10)=7.6, $$ $$ S_3=7.6+0.3(8-7.6)=7.72. $$
For the constant schedule, $$ S_1=0+0.4(10-0)=4,\quad S_2=4+0.4(6-4)=4.8, $$ $$ S_3=4.8+0.4(8-4.8)=6.08. $$
These short paths do not establish which schedule is best: the initial values and only three simulated statistics are illustrative, not a convergence comparison. They do show the weighting difference. The first schedule follows the initial statistic fully and then smooths more; the piecewise schedule also reacts strongly at first but gives the second observation greater weight than the first schedule does; the constant schedule moves more gradually throughout.
Why Constant Gains Leave Residual Variability
A simple scalar calculation isolates the effect of persistent simulation noise. Suppose the simulated quantity is \(Y_k=\theta+\varepsilon_k\), where the errors are independent, have mean zero, and have variance \(\tau^2\). Consider the update \(S_k=S_{k-1}+\gamma(Y_k-S_{k-1})\) with a fixed \(0<\gamma<1\). The error from the target, \(e_k=S_k-\theta\), then satisfies $$ e_k=(1-\gamma)e_{k-1}+\gamma\varepsilon_k. $$
Because the new error is independent of the previous one and has mean zero, the variance \(V_k=\operatorname{Var}(e_k)\) follows $$ V_k=(1-\gamma)^2V_{k-1}+\gamma^2\tau^2. $$ If this variance reaches a steady value \(V\), it must satisfy \(V=(1-\gamma)^2V+\gamma^2\tau^2\). Solving gives $$ V\bigl[1-(1-\gamma)^2\bigr]=\gamma^2\tau^2,\qquad V=\frac{\gamma}{2-\gamma}\tau^2. $$ For every fixed positive gain, this variance is positive. Reducing \(\gamma\) lowers the steady variability in this simplified setting, but it also makes adjustment slower. In SAEM the simulated statistics can be dependent, especially when they are generated by a Markov chain, so this independent-noise formula is an explanation of the trade-off rather than a complete variance formula for a fitted NLME model.
Worked Example: Quantifying Constant-Gain Variability
Suppose, for illustration, that the standard deviation of independent simulation noise in a scalar statistic is \(\tau=2\), so \(\tau^2=4\). Compare the steady variance under constant gains \(\gamma=0.2\) and \(\gamma=0.5\), using the formula just derived.
For \(\gamma=0.2\), $$ V=\frac{0.2}{2-0.2}(4)=\frac{0.8}{1.8}\approx0.4444, \qquad \sqrt V\approx0.6667. $$
For \(\gamma=0.5\), $$ V=\frac{0.5}{2-0.5}(4)=\frac{2}{1.5}\approx1.333, \qquad \sqrt V\approx1.155. $$
Under these assumptions, the smaller constant gain has lower residual standard deviation, but it also retains more of the previous statistic at each iteration: its weight on the new simulation is 0.2 rather than 0.5. The calculation illustrates a trade-off, not a recommendation that one of these gains is appropriate for a particular model.
Practical Selection and Diagnostics
A decreasing schedule is a natural choice when the aim is to reduce the influence of simulation noise over a long run. Within the power family \(k^{-a}\), exponents closer to 1 decrease faster and are more smoothing; exponents closer to \(1/2\) decrease more slowly and maintain greater responsiveness. A piecewise-decreasing choice can make early updates more responsive while still having a diminishing tail. The constant-gain option is useful to understand as a deliberate responsiveness-versus-noise trade-off, but should not be mistaken for a schedule satisfying the usual diminishing-gain convergence conditions.
A schedule should be assessed alongside the traces of the running statistics and parameters. Warning signs include persistent oscillation late in a run, statistics that remain far from a stable range, or changes so small that the statistics barely move while the parameter estimates remain sensitive to initialization. These patterns have multiple possible causes: an unsuitable gain, weak latent-effect chain mixing, poor parameter scaling, or an unstable numerical update. A trace alone does not identify which cause is responsible.
In particular, a noisy trace does not prove that the gain is too large: strong autocorrelation in the simulated latent effects can also make successive updates look erratic or slow to explore. Conversely, a smooth trace does not prove convergence. Gains that have become very small can conceal continuing mismatch between the current estimates and the stationary solution. Examine statistic and parameter behavior together, and interpret a gain change in the context of the simulation and model.
SAEM is stochastic, so an iteration need not increase the observed-data likelihood. Nor does a small update by itself establish that the likelihood has reached a maximum. Schedule choice controls the stochastic approximation's responsiveness and noise; it does not remove the need for suitable simulation, stable parameter updates, and convergence assessment.
Derivation: Why the Standard Power Range Works
For \(\gamma_k=k^{-a}\), the two standard conditions reduce to checking \(\sum k^{-a}\) and \(\sum k^{-2a}\). The p-series \(\sum k^{-r}\) diverges for \(r\leq1\) and converges for \(r>1\). Consequently, the first sum diverges when \(a\leq1\), and the second converges when \(2a>1\), or \(a>1/2\). Combining the inequalities yields $$ \frac12<a\leq1. $$ This is a justification of the usual power-law range, not a proof that every SAEM implementation using an exponent in this range converges. The stochastic approximation conditions must be accompanied by the model-specific and simulation-specific assumptions needed by the convergence result being applied.
For a piecewise-decreasing schedule with any finite initial segment and a tail proportional to \(j^{-a}\), the initial segment contributes only finite amounts to both sums. Adding finite amounts does not change whether an infinite series converges or diverges. Therefore the tail exponent determines these two summability checks, provided its multiplier is finite and positive.
Common Pitfall and Limitations
No schedule is universally optimal. The calculations here describe how gain affects averaging and noise in simplified settings; they do not predict the exact finite-run behavior of a nonlinear mixed-effects model. In practice, the choice must account for the target of the run, the stability of the M-step, and the behavior of conditional random-effect simulation. A schedule may satisfy the standard summability conditions and still perform poorly if those other ingredients are inadequate.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- For \(\gamma_k=k^{-0.6}\), do the usual Robbins–Monro summability conditions hold? Explain using the relevant exponents.
- Starting from \(S_0=2\), compute \(S_1\) when \(\widetilde S_1=8\) and \(\gamma_1=0.25\).
- With independent simulation noise of variance \(\tau^2=9\), what is the steady variance in the scalar constant-gain calculation when \(\gamma=0.25\)?
- Why can a smooth trace under a constant gain not by itself establish exact convergence?
Solutions. (1) Yes. The gain sum is \(\sum k^{-0.6}\), which diverges because \(0.6\leq1\); the squared-gain sum is \(\sum k^{-1.2}\), which converges because \(1.2>1\). (2) \(S_1=2+0.25(8-2)=2+1.5=3.5\). (3) \(V=\frac{0.25}{2-0.25}(9)=\frac{2.25}{1.75}\approx1.286\), in squared units of the statistic. (4) A constant positive gain continues to pass new simulation noise into the update, leaving residual variability under the scalar noise model. A visually smooth trace may therefore fluctuate within a narrow band rather than converge exactly; the trace and gain rule must be interpreted together.