⚡ electrical engineering

Statistical Efficiency of Single- and Multi-step Models for Forecasting and Control

This paper provides a quantitative analysis of the trade-off between single-step and multi-step predictors for linear dynamical systems, demonstrating that while single-step models are optimal under well-specified conditions, multi-step predictors significantly improve accuracy and control performance when the system is partially observable and the model is misspecified.

Original authors: Anne Somalwar, Bruce D. Lee, George J. Pappas, Nikolai Matni

Published 2026-03-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Anne Somalwar, Bruce D. Lee, George J. Pappas, Nikolai Matni

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to navigate a maze. To do this, the robot needs a "predictor"—a brain that guesses what will happen next based on what it sees now.

The paper you provided tackles a classic dilemma in robotics and AI: Should the robot learn to take one small step at a time, or should it try to learn the whole path at once?

Here is the breakdown of the three "brains" (predictors) the authors tested, explained with simple analogies:

The Three Approaches

The "One-Step" Robot (Single-Step Model)
- How it works: This robot learns a simple rule: "If I am here and I turn left, I will end up there." It learns this rule perfectly for one second. Then, to predict 10 seconds into the future, it takes that rule, applies it, gets a new position, applies the rule again, and so on.
- The Flaw: It's like playing the game of "Telephone." If the robot makes a tiny mistake in the first second, that mistake gets carried over to the second second, then the third. By the time it reaches the 10th second, the tiny error has exploded into a huge disaster. This is called compounding error.
- The Benefit: It only needs to learn a simple rule, so it learns very quickly with little data.
The "Long-Haul" Robot (Direct Multi-Step Model)
- How it works: This robot doesn't care about the next second. It looks at the data and learns a massive, complex rule: "If I am here, I will be exactly at this spot in 10 seconds." It skips the middle steps entirely.
- The Benefit: It avoids the "Telephone" game. Since it predicts the destination directly, small mistakes in the middle don't pile up.
- The Flaw: To learn this complex rule, it needs a lot of data. It's like trying to memorize the entire map of a city in one go rather than learning one street at a time. It's "data-hungry."
The "Hybrid" Robot (Single-Step trained with Multi-Step Loss)
- How it works: This robot tries to learn the simple "one-step" rule, but during training, the teacher grades it on how well it does after 10 steps, not just 1. It tries to be a simple robot that thinks ahead.
- The Result: It's a middle ground, but as we'll see, it often gets the worst of both worlds.

The Big Discovery: It Depends on "Truth"

The authors ran experiments to see which robot wins. The answer depends entirely on whether the robot has a perfect view of the world or a blurry view.

Scenario A: The Perfect View (Well-Specified)

Imagine the robot can see the entire maze clearly. It knows exactly where it is.

The Winner: The One-Step Robot.
Why? Because the world is simple and clear, the "Telephone" game isn't that bad. The One-Step robot learns the rules so fast and so accurately that it quickly outperforms the Long-Haul robot, which is still struggling to memorize the complex map.
Analogy: If you are walking on a straight, flat sidewalk, you don't need to memorize the whole path. You just need to know how to take one step. Taking one step at a time is the most efficient way to get there.

Scenario B: The Blurry View (Misspecified)

Imagine the robot is wearing foggy glasses. It can't see its exact position; it only sees a blurry shape. It thinks it knows where it is, but it's actually guessing.

The Winner: The Long-Haul Robot.
Why? When the view is blurry, the "One-Step" robot makes a small guess. Because it doesn't know the truth, that guess is slightly wrong. When it tries to take the next step based on that wrong guess, it gets even more lost. The error compounds rapidly, and the robot ends up in a completely different city.
The Fix: The Long-Haul robot ignores the confusing middle steps. It looks at the start and the finish and learns the direct relationship between them. Even though it needs more data to learn, it avoids the trap of getting lost in the foggy middle.
Analogy: If you are driving in thick fog, trying to guess your position every 10 feet is dangerous. You might drift off the road. It's better to have a GPS that tells you, "In 5 miles, you will be at the exit," even if that GPS is harder to program.

What About Control? (Driving the Robot)

The paper also tested these robots when they had to actually drive the car (control the system).

In Clear Weather: The One-Step robot drives the smoothest and most efficiently.
In Foggy Weather: The One-Step robot often crashes because its predictions are too shaky. The Long-Haul robot, despite being "dumber" in terms of data efficiency, is more robust. It keeps the car on the road because it doesn't rely on a chain of shaky guesses.

The Takeaway for Real Life

This research gives us a rule of thumb for building AI systems:

If your data is clean and you understand the system perfectly: Use the simple, one-step approach. It's faster to train and works better in the long run.
If your data is noisy, incomplete, or the system is complex: Don't be afraid to use the complex, multi-step approach. It requires more data, but it prevents the "snowball effect" of errors that causes simple models to fail in the real world.

In short: If the world is simple, take it one step at a time. If the world is messy and confusing, try to see the big picture.

1. Problem Statement

In learning-based control and time-series forecasting, a fundamental challenge is compounding error: small prediction mistakes made at one time step accumulate when a model is rolled out recursively over a long horizon, leading to poor long-term predictions.

Common Approach: Train a single-step predictor (autoregressive model) and roll it out recursively. This is statistically efficient but suffers from compounding error.
Alternative Approach: Train a direct multi-step predictor to minimize error over a specific horizon $H$ . This mitigates compounding error but requires learning a more complex model with more parameters, potentially requiring more data (lower statistical efficiency).
The Gap: While the trade-off between prediction horizon accuracy and data requirements is empirically known, there is a lack of quantitative, theoretical guidance on when to prefer direct multi-step prediction over autoregressive single-step prediction, particularly regarding statistical efficiency and closed-loop control performance.

2. Methodology

The authors provide a rigorous theoretical analysis for Linear Time-Invariant (LTI) systems under two distinct settings:

Well-Specified Setting: The system is fully observable ( $C=I, D_v=0$ ), and the hypothesis class (linear predictors) correctly captures the system dynamics.
Misspecified Setting: The system is partially observable ( $C \neq I$ or $D_v \neq 0$ ), meaning the Markovian assumption (that current observation predicts the future) is violated.

The authors compare three classes of predictors, all using a static linear hypothesis class:

Single-Step (SS): Minimizes one-step prediction error; rolled out autoregressively for $H$ steps.
Multi-Step (MS): Directly minimizes the $H$ -step prediction error (unconstrained parameters for each step).
Intermediate (I): A single-step model structure trained with a multi-step loss function. This attempts to capture multi-step dynamics without increasing the parameter count of the underlying model.

Key Analytical Tools:

Asymptotic Analysis: The authors derive the asymptotic behavior of the prediction error as the dataset size $N \to \infty$ .
Error Decomposition:
- In the well-specified case, they analyze the reducible error (variance), showing how it scales with $N$ and the horizon $H$ .
- In the misspecified case, they analyze the irreducible error (bias), which persists even with infinite data due to model mismatch.
Closed-Loop Control: They evaluate the predictors within a Model Predictive Control (MPC) framework, analyzing the resulting infinite-horizon Linear Quadratic Regulator (LQR) cost and system stability (spectral radius).

3. Key Contributions

Quantitative Trade-off Analysis: The paper provides the first rigorous theoretical comparison of the sample efficiency of these three predictor classes for linear systems.
Well-Specified Results:
- Proves that for fully observed systems, Single-Step (SS) predictors achieve the lowest asymptotic prediction error (fastest decay rate of $O(1/N)$ ).
- Shows that the Intermediate (I) predictor is less efficient than SS but more efficient than Multi-Step (MS).
- Demonstrates that the efficiency gap between SS and MS grows with the prediction horizon $H$ and the system's spectral radius.
Misspecified Results:
- Shows that under partial observability (model misspecification), Multi-Step (MS) predictors significantly outperform SS predictors by reducing asymptotic bias.
- Proves that the bias of SS predictors scales poorly with the horizon $H$ in misspecified settings, often leading to prohibitive compounding errors, whereas MS predictors maintain lower bias.
Control Performance:
- In well-specified settings, SS predictors yield lower control costs.
- In misspecified settings, MS predictors are more likely to induce stabilizing controllers because their lower bias keeps the estimated feedback gain closer to the true stabilizing gain.
Nonlinear Extension: Empirical evidence suggests these theoretical trends hold for certain nonlinear systems (via Koopman lifting).

4. Key Results

A. Well-Specified Setting (Fully Observed)

Ordering of Efficiency: $SS < I < MS$ (where $<$ denotes lower prediction error/variance).
Reasoning: The SS predictor leverages the true Markovian structure and avoids accumulating process noise in the training loss. The MS predictor suffers from a "curse of dimensionality" where the number of parameters scales with $H$ , increasing variance.
Control: SS predictors lead to faster convergence of the LQR cost to the optimal value compared to I and MS.

B. Misspecified Setting (Partially Observed)

Ordering of Bias: $MS < I < SS$.
Reasoning: When the Markov assumption is false, the SS predictor accumulates significant bias as it rolls out predictions. The MS predictor, by directly fitting the long-horizon trajectory, effectively "learns" the non-Markovian dependencies, resulting in significantly lower asymptotic bias.
Stability: In numerical experiments, SS predictors often fail to stabilize the system (spectral radius $>1$ ) under misspecification, while MS predictors successfully stabilize the system.

C. Empirical Validation

Linear Systems: Experiments confirm the theoretical decay rates for well-specified systems and the bias dominance for misspecified systems.
Nonlinear Systems: Experiments on a nonlinear system (with Koopman lifting) show that the qualitative ordering of errors (SS best in well-specified, MS best in misspecified) persists.

5. Significance and Implications

This work provides principled guidance for practitioners designing data-driven controllers and forecasters:

When to use Single-Step: If the system is well-modeled (fully observable or the model class is rich enough) and data is limited, autoregressive single-step models are statistically superior.
When to use Multi-Step: If the system suffers from partial observability or model misspecification (e.g., unmodeled dynamics, hidden states), the benefits of reduced bias in direct multi-step prediction outweigh the cost of increased model complexity.
Control Design: For safety-critical control where stability is paramount, direct multi-step prediction is preferred in scenarios where the model is known to be imperfect or the state is not fully observable, as it is more robust to inducing stabilizing policies.

The paper bridges the gap between empirical heuristics and theoretical understanding, offering a clear framework for selecting prediction architectures based on system observability and data availability.