⚡ electrical engineering

Remarks on stochastic cloning and delayed-state filtering

This paper demonstrates that a properly derived delayed-state Kalman filter (DSKF) achieves state and covariance updates identical to stochastic cloning for handling delayed-state measurements, offering equivalent accuracy without the need for state augmentation and potentially reducing computational costs.

Original authors: Tara Mina, Lindsey Marinello, John Christian

Published 2026-05-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tara Mina, Lindsey Marinello, John Christian

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 navigate a car using a GPS that sometimes tells you where you were two minutes ago, rather than where you are right now. In the world of robotics and aerospace, this is a common problem called "delayed-state measurement." The challenge is that your current guess about where you are is mathematically linked to your guess from two minutes ago. If you ignore that link, your navigation system gets confused and makes mistakes.

For a long time, the standard advice from engineers has been: "To fix this, you must clone your past self."

This method is called Stochastic Cloning (SC). Imagine you are driving, and every time you get a delayed GPS signal, you magically create a "ghost" version of your car from two minutes ago. You then carry both the real car and the ghost car in your mental map, calculating how they relate to each other. It works perfectly, but it's heavy. You have to carry extra mental baggage (computing power and memory) to keep track of these ghosts.

The Paper's Big Discovery
The authors of this paper, Tara Mina, Lindsey Marinello, and John Christian, say: "You don't actually need the ghosts."

They revisited an older, somewhat forgotten method called the Delayed-State Kalman Filter (DSKF). Think of this as a "smart calculator" that knows how to look at the past and the present simultaneously without needing to physically carry a copy of the past.

Here is the simple breakdown of their findings:

1. The "Ghost" vs. The "Smart Calculator"

Stochastic Cloning (The Ghost Method): You take your current state and your past state, glue them together into one big list, and run the math. It's like carrying a backpack full of old maps to help you read the new one. It works, but the backpack gets heavy.
Delayed-State Kalman Filter (The Smart Calculator): This method looks at the same math but rearranges the equations. Instead of carrying the past state as a separate item, it uses a mathematical "bridge" (called a State Transition Matrix) to translate the past directly into the present. It's like having a translator who speaks both "Past" and "Present" fluently, so you don't need to carry the old maps; you just ask the translator for the connection.

2. The Result: Same Destination, Lighter Load

The paper proves mathematically that both methods arrive at the exact same answer. If you use the "Ghost" method or the "Smart Calculator" method, your car will end up in the exact same spot with the exact same confidence level.

However, the "Smart Calculator" (DSKF) has a distinct advantage in efficiency:

Less Mental Load: Because it doesn't have to store and update the extra "ghost" variables, it uses less computer memory.
Faster Math: In many scenarios, it requires fewer calculation steps (arithmetic operations) to get the result.

3. Two Ways to Build the "Smart Calculator"

The authors didn't just find one way to do this; they showed two different ways to build the "Smart Calculator" that are mathematically identical:

Forward-Time Version: This looks at the past and projects it forward to the present. It's like looking at an old photo and imagining how you would look today.
Backward-Time Version: This looks at the present and projects it backward to the past. It's like looking at your current face and imagining what you looked like in the photo.

Both versions work perfectly, and engineers can choose whichever one fits their specific software better.

4. Why This Matters (According to the Paper)

For the last 20 years, many people believed that the standard Kalman filter (the engine behind most navigation systems) was "broken" when it came to delayed measurements. They thought the only solution was to "clone" the state (add the ghosts).

This paper corrects that misconception. It says: "The standard engine isn't broken; you just need to tune it differently." You don't need to add extra weight (cloning) to handle delayed information. You just need to use the correct mathematical formulation (DSKF) that accounts for the connection between the past and present naturally.

Summary Analogy

Imagine you are trying to solve a puzzle where a piece from yesterday is needed to finish today's picture.

Stochastic Cloning says: "Bring yesterday's puzzle piece into today's room, tape it to the table, and solve the whole thing together." (Accurate, but messy and takes up table space).
Delayed-State Kalman Filter says: "Don't move the piece. Just use a special lens to see how yesterday's piece fits into today's picture without moving it." (Accurate, clean, and saves table space).

The paper proves that the "special lens" method gives you the exact same finished picture as the "moving the piece" method, but it does so more efficiently.

Technical Summary: Remarks on Stochastic Cloning and Delayed-State Filtering

Problem Statement
Many state estimation problems in aerospace navigation and robotics, particularly those involving odometry (e.g., visual-inertial, LiDAR-inertial, and wheel odometry) and time-differenced carrier phase (TDCP) GNSS measurements, rely on observations that depend on prior states. These delayed-state measurements introduce correlations between the measurement error and the prior state estimate error. This dependency violates the standard assumption in conventional Kalman filter formulations that measurements are conditionally independent of prior states given the current state.

A prevailing view in the literature over the past two decades has been that the standard Kalman filter cannot properly handle these correlations without state augmentation. Consequently, Stochastic Cloning (SC) has become a widely accepted approach. SC addresses the issue by augmenting the state vector with copies of prior states and expanding the covariance matrix to capture the resulting cross-correlations. While effective, this method increases the dimensionality of the filter, potentially impacting computational and memory efficiency.

Methodology
This work revisits a long-established but often overlooked alternative: the Delayed-State Kalman Filter (DSKF), originally derived by Brown and Hartman (1968). The authors demonstrate that a properly derived DSKF yields exactly the same state and covariance updates as Stochastic Cloning, without requiring state augmentation.

The paper presents two equivalent formulations of the DSKF, derived by expressing the measurement innovation in terms of either the prior state (forward-time formulation) or the current predicted state (backward-time formulation):

Forward-Time Formulation: Expresses the measurement model in terms of the prior state estimate. This approach avoids explicitly inverting the State Transition Matrix (STM) but requires propagating the measurement model backward through the dynamics.
Backward-Time Formulation: Expresses the measurement solely as a function of the current predicted state. This mirrors the structure of the traditional Kalman filter but requires the evaluation of the inverse STM.

The authors provide a rigorous mathematical proof by induction demonstrating the equivalence between Stochastic Cloning and the DSKF (specifically the forward-time formulation). They show that the Kalman gain and the updated state error covariance derived via SC are mathematically identical to those derived via DSKF. Furthermore, the Appendix proves the equivalence between the forward-time and backward-time DSKF formulations.

The paper also explores simplifications for cases where the delayed-state measurement depends only on a subset of the prior state dimensions (sparse prior-state dependence). By exploiting this sparsity, the authors derive reduced-dimension versions of both SC and DSKF that operate only on the relevant state submatrices, thereby lowering computational overhead.

Key Results

Mathematical Equivalence: The paper establishes that Stochastic Cloning and the Delayed-State Kalman Filter produce identical state estimates and covariance updates. The perceived inability of the standard Kalman filter to handle delayed-state correlations is attributed to the specific formulation used, not a fundamental limitation of the filter itself.
Complexity Analysis: The authors perform a detailed analysis of arithmetic complexity (floating-point operations) and memory usage.
- Both SC and DSKF exhibit cubic scaling in arithmetic complexity and quadratic scaling in memory usage with respect to the system dimensions.
- However, the Forward-Time DSKF consistently requires fewer floating-point operations (flops) than Stochastic Cloning.
- In terms of memory, the Forward-Time DSKF offers reduced storage costs compared to SC, particularly when the state dimension ( $n$ ) is large relative to the measurement dimension ( $m$ ).
Sparse Dependence: When measurements depend on only a subset of prior states, both methods can be simplified. The DSKF formulations naturally accommodate these simplifications by operating on reduced submatrices, further enhancing efficiency.

Significance and Claims
The primary significance of this work is to clarify a common misconception within the estimation community: that Kalman filter variants are inherently unable to handle correlated delayed-state measurements without state augmentation. The authors assert that the DSKF is a valid, well-established alternative that achieves the same results as Stochastic Cloning without the artificial inflation of the state vector.

The paper does not claim that one method is universally superior in all implementation contexts, noting that performance can depend on specific hardware and software frameworks. However, it argues that the DSKF offers a more unified framework for handling correlations and can provide tangible reductions in arithmetic and memory costs for certain problem dimensions. The work serves to remind the community that revisiting the fundamental Kalman filter formulation to incorporate cross-correlation terms is a viable and efficient alternative to state augmentation.

1. The "Ghost" vs. The "Smart Calculator"

2. The Result: Same Destination, Lighter Load

3. Two Ways to Build the "Smart Calculator"

4. Why This Matters (According to the Paper)

Summary Analogy

More like this