Are particles equally important?

Usually not. Their weights determine how strongly they contribute to the approximation.

Does SIS automatically resample?

Plain SIS does not. Resampling is added in related sequential Monte Carlo and particle-filter procedures.

Can the weights reveal every missing possibility?

No. Weighting adjusts the contribution of sampled candidates; it cannot guarantee coverage of regions the proposal failed to explore.

AI Glossary / Generative AI and language / Advanced

Sequential importance sampling

Definition

Sequential importance sampling, or SIS, is a probabilistic inference method that builds candidate sequences incrementally and updates a weight for each candidate as it grows. The weights correct for differences between the distribution used to propose candidates and the target distribution being approximated. SIS lets earlier computation be reused as new steps or observations arrive. Plain SIS does not include a resampling step; adding resampling leads to related sequential Monte Carlo procedures.

What it means

A hidden process may evolve over time while observations reveal only part of its state. Instead of evaluating every possible history, SIS maintains a collection of sampled histories, often called particles, with numerical weights.

A particle is a candidate explanation or state trajectory, not a physical object. Its weight expresses its relative contribution to the approximation under the specified target and proposal.

The concept extends beyond language AI, despite the supplied category. It can apply to time-series models, probabilistic programs, and sequence-based inference. The word “sequential” describes incremental construction, which need not always correspond to real-world clock time.

How it works

Start with particles drawn from an initial proposal and compute their importance weights. At each step, extend every particle by sampling its next component from a proposal conditioned on the available history and, where designed, observations.

Multiply the previous weight by an incremental correction involving the new target contribution and proposal probability. Normalize weights when estimating expectations or comparing particle contributions. In a state-space model, this correction commonly combines transition and observation likelihood terms divided by the proposal density.

Monitor concentration of the normalized weights. Over many steps, a few particles can dominate while most contribute very little. Good proposals and related methods with resampling can help, but they have their own assumptions and design choices. A large nominal particle count alone does not establish an accurate approximation.

Three practical examples

1. A tracking system maintains weighted candidate movement histories as new sensor observations arrive, under a specified motion and observation model.

2. A probabilistic sequence model builds candidate hidden-state paths and estimates properties of those paths using their normalized weights.

3. A research workflow extends candidate structures step by step, using importance corrections to account for a convenient proposal that differs from the desired target.

Think of it this way

Think of SIS as keeping several developing explanations on a worksheet. Each new piece of evidence changes how much weight each explanation receives, while the explanations are extended to cover the next stage.

The analogy has limits. The weights are calculated from mathematical distributions, not subjective confidence alone. If the proposal rarely explores an important possibility, reweighting cannot reliably recover a possibility that is absent from the sampled collection.

Worked example

Suppose three particles currently have normalized weights 0.5, 0.3, and 0.2. For a simplified bootstrap-proposal setting, assume the proposal equals the model's transition distribution, so those transition terms cancel in the incremental ratio.

A new observation has likelihoods 0.2, 0.6, and 0.5 for the three newly proposed states. Multiply to obtain unnormalized updated weights: 0.10, 0.18, and 0.10. Their total is 0.38, giving normalized weights approximately 0.2632, 0.4737, and 0.2632.

If the current state values are 1, 2, and 4, the weighted estimate is (0.10 × 1 + 0.18 × 2 + 0.10 × 4)/0.38, approximately 2.2632. This is an estimate based on three particles, not an exact posterior calculation. With a different proposal, the omitted proposal correction would generally be required.

Illustrative case study

Fictional case: Cedar Tracking prototypes a model for estimating the state of a moving asset from noisy observations. Its SIS implementation carries many candidate paths, but after a long sequence nearly all normalized weight rests on one path.

The team recognizes that counting stored particles overstates the diversity of the effective approximation. It reviews proposal quality and investigates a suitable particle-filter design with resampling, then tests accuracy against cases where a trusted reference is available.

It does not claim that the largest-weight path must be correct. The dominant weight describes the current sampled approximation, which can still be poor if important alternatives were missed.

Common misconceptions and mistakes

SIS is not simply choosing the most likely next state at every step. It samples candidates and retains weighted alternatives.

Importance weights are not arbitrary quality scores. They arise from target and proposal distributions and must be updated consistently. Forgetting the proposal denominator can change the distribution being approximated.

Another mistake is using SIS and sequential importance resampling as exact synonyms. Resampling changes the particle population and is a separate operation. Finally, a numerically stable weight calculation does not by itself solve statistical degeneracy; both numerical and sampling quality need attention.

Business use

SIS is relevant when a system needs approximate inference over evolving or incrementally constructed hidden states. It provides a foundation for understanding more practical particle-based methods.

Business evaluation should focus on the complete inference system: assumptions about dynamics, observation reliability, proposal design, computation budget, and error behavior. Compare estimates against suitable reference cases and inspect performance over long sequences. Ask how the implementation detects weight concentration and how it responds, rather than accepting particle count as the sole measure of confidence.

Check your understanding

An implementation changes its proposal distribution but continues using an old weight update that ignores the change. Is the target approximation necessarily preserved?

Answer: No. Importance corrections depend on the proposal. The update must be derived consistently for the new proposal and target sequence.

Related terms

Importance sampling
Sequential Monte Carlo
Proposal distribution
Likelihood
Resampling
Effective sample size
particle filter
posterior distribution
state-space model
weight degeneracy

Learn next

Work through the three-particle update and verify that normalized weights sum to one. Then study how the update changes when the proposal differs from the transition model. Compare a plain SIS trace with a resampling-based trace, recording both approximation error and the concentration of particle weights.

Sources

Sources: Alexandre Bouchard-Cote, UBC Lecture 7: Sequential Monte Carlo, https://www.stat.ubc.ca/~bouchard/courses/stat520-sp2014-15/lecture/2015/03/15/notes-lecture7.html — primary SIS derivation and degeneracy discussion read. Naesseth, Lindsten, and Schon, Elements of Sequential Monte Carlo, https://arxiv.org/abs/1903.04797 — original author abstract for broader scope. Carnegie Mellon Monte Carlo notes, https://www.cs.cmu.edu/~epxing/Class/10708-19/notes/lecture-13/ — primary weighting and particle-filter context. The numerical example is original.

Previous: Sequential Feature Selection Next: Sequential minimal optimization