Yes. It processes a continuing stream while reporting a state a fixed distance behind the present.

Is it a moving average?

Not generally. Probabilistic smoothing uses a state and observation model, rather than simply averaging nearby measured values.

What happens with lag zero?

The target is the current state, so the inference target becomes the filtering distribution.

AI Glossary / Time series and signal processing / Intermediate

Fixed-lag smoothing

Definition

Fixed-lag smoothing estimates a hidden state from a fixed number of time steps in the past using observations available through the present. With lag d at time t, it estimates the state at t−d from evidence through t. It offers a deliberate delay in exchange for using later evidence about an earlier state.

What it means

A noisy observation may become easier to interpret after a few more readings arrive. Fixed-lag smoothing waits a specified interval before producing the corresponding historical state estimate. It can operate continuously: as time advances, the target state advances too, always remaining the same number of steps behind the newest evidence.
This differs from ordinary filtering, which estimates the current state using evidence so far. It also differs from waiting until an entire record is complete and smoothing every state across that fixed interval.

How it works

Maintain information from forward filtering and a bounded set of recent observations or messages. When a new observation arrives, use the model's transition and observation relationships to propagate its information backward toward the target state. Combine that backward information with the evidence already available at the target time and normalize.
The exact computation depends on the model: discrete hidden Markov models, linear Gaussian systems, and particle-based approximations use different representations. A fixed lag limits how far into the future of each target state the estimator looks. It does not remove modeling assumptions, and numerical shortcuts can have additional conditions that must be checked.

Three practical examples

A speech system waits briefly before finalizing an earlier sound label because the following sound can resolve ambiguity.
A location service refines a recent position after several new sensor readings arrive, while labeling the estimate with its original time.
An industrial dashboard displays a short-delayed estimate of an internal operating condition. A separate immediate alert handles events for which the smoothing delay would be unacceptable.

Think of it this way

Think of reading a sentence and reconsidering an ambiguous word after seeing the next few words. You are improving an interpretation of something earlier, not predicting the words that have already arrived. Fixed-lag smoothing formalizes a limited look-ahead relative to the target time, while remaining sequential relative to the live data stream.

Worked example

Consider an original two-state example with lag one. At time one, all evidence then available gives probability 0.6 that a fictional unit is active and 0.4 that it is inactive. An active unit stays active at the next step with probability 0.9; an inactive unit becomes active with probability 0.2.
At time two, a positive sensor indication arrives. Its probability is 0.8 if the time-two state is active and 0.1 if inactive. If the time-one state was active, the probability of this later positive indication is 0.9 × 0.8 + 0.1 × 0.1 = 0.73. If the time-one state was inactive, it is 0.2 × 0.8 + 0.8 × 0.1 = 0.24.
The time-one smoothing weights are therefore 0.6 × 0.73 = 0.438 and 0.4 × 0.24 = 0.096. Normalizing gives 0.438/(0.438 + 0.096) = 73/89, approximately 0.82022, for the probability that the unit was active at time one.
The original filtered probability was 0.6; the one-step-delayed smoothed probability is about 0.82022. They answer the same historical-state question with different evidence. This result is not the probability of activity at time two, and it was unavailable at time one. The example assumes the stated Markov and conditional-observation relationships.

Illustrative case study

Illustrative fictional case: Elm Tracking supplies positions for reviewing warehouse-cart movements. A two-second delay is acceptable for the review display, so the team compares immediate filtered positions with positions revised using two seconds of additional readings.
It keeps timestamps and publication times separate, ensuring that supervisors understand which moment each estimate describes. The revised track is not used to claim that a collision warning could have been issued earlier. The fictional case demonstrates how smoothing can improve a retrospective display without changing what information an earlier decision maker possessed.

Common misconceptions and mistakes

Fixed lag does not mean a fixed historical time. With lag three, the estimator targets time seven when observations reach time ten, then time eight when they reach eleven. A fixed-point smoother instead continues revising one particular target time.
More evidence can improve expected inference under a correct model and appropriate loss, but an individual smoothed estimate can still be wrong. Model mismatch, correlated sensor faults, and approximation error can undermine results. Do not report smoothed performance as real-time filtering performance without including the extra delay and evidence.

Business use

Organizations can use fixed-lag smoothing when a small reporting delay is acceptable and later observations help distinguish plausible past states. Examples include inspection logs, delayed tracking displays, and sequential annotation systems.
Choose the lag by measuring the tradeoff between delay, estimation quality, memory, and computation. Keep the chosen lag fixed during an evaluation unless the adaptation rule is explicitly part of the system. Test periods with missing or late observations, because a fixed number of samples does not necessarily equal a fixed elapsed duration when acquisition is irregular.

Check your understanding

At time twelve with lag four, which state is estimated?

Answer: The state at time eight, using evidence through time twelve.

In the example, why is 0.73 a likelihood rather than a posterior?

Answer: It gives the probability of the new evidence conditional on a particular earlier state.

Could the 0.82022 estimate have supported a time-one live decision?

Answer: No. It depends on a sensor indication that arrived at time two.

Related terms

Filtering
Smoothing
Hidden Markov model
Forward-backward algorithm
Latency
fixed-interval smoothing
fixed-point smoothing
state estimation
observation likelihood

Learn next

compare filtering and one-step smoothing in a two-state model, then extend the calculation to two later observations. Track both the target timestamp and the evidence cutoff so that accuracy comparisons reflect the information each method actually receives.

Sources

Sources: Russell and Norvig, official AIMA algorithms collection, Fixed-Lag-Smoothing: https://aima.cs.berkeley.edu/algorithms.pdf . Kevin Murphy, Dynamic Bayesian Networks thesis, section 3.7.2: https://www.cs.ubc.ca/~murphyk/Thesis/thesis.pdf . Primary indexed algorithm and thesis excerpts consulted; AIMA temporal-model draft also checked: https://personal.utdallas.edu/~vibhav.gogate/ml/2020f/lectures/dbn-chapter.pdf . The calculation and fictional case are original.

Previous: Fixed Effects Model Next: Fixed marginal totals