How long can you trust a Starlink TLE? An empirical comparison of SGP4 and high-fidelity propagation against operator-updated truth across a megaconstellation

This empirical study of over 24,000 Starlink TLE pairs from April 2026 reveals that position errors follow a power-law growth over time, that high-fidelity propagation using public TLEs generally fails to outperform the standard SGP4 model due to operator-epoch residuals and model alignment biases, and that SGP4 staleness shows a statistically significant but uncalibrated correlation with solar flux at specific altitudes.

The Gist

Imagine you are trying to predict where a specific car will be on a highway in one hour, three hours, or a week. You have two tools to help you:

1. The "Official Map" (SGP4): This is a simple, fast, and widely used method that the government provides for free. It's like a standard GPS app that gives you a good estimate based on average traffic patterns.
2. The "Supercomputer Simulation" (High-Fidelity): This is a complex, physics-heavy simulation that accounts for every tiny detail: wind resistance, the exact shape of the car, the weight of the passengers, and even the gravitational pull of the moon. It's like a racing team's wind-tunnel simulation.

The paper asks a simple question: If you start with the same "Official Map" data, does the "Supercomputer Simulation" actually give you a better prediction of where the car will be?

The researchers studied thousands of SpaceX Starlink satellites (which are like a massive fleet of cars in low Earth orbit) to find out. Here is what they discovered, using simple analogies:

1. The "Freshness" Rule (How long can you trust the data?)

The paper found that the "Official Map" (SGP4) is surprisingly good, but only for a short time.

• The Analogy: Think of the satellite's position data like a weather forecast. If you check the forecast 4 hours after it was published, it's usually accurate. If you try to use that same forecast to predict the weather 7 days from now, it becomes useless.
• The Finding: For Starlink, the "Official Map" is reliable for about 4 to 6 hours. After that, the error starts to grow. By day 7, the satellite could be tens of kilometers away from where the map says it is. The researchers found that this error grows in a predictable pattern (like a power law), meaning they can mathematically estimate how "stale" the data is based on how long it's been since the last update.

2. The "Supercomputer" Surprise (Does more detail help?)

You might think the "Supercomputer Simulation" (High-Fidelity) would always win because it knows more physics. It didn't.

• The Analogy: Imagine you are trying to guess where a runner will be in 10 minutes.
  • Tool A (SGP4): You use a simple rule: "They run at 10 mph."
  • Tool B (Supercomputer): You use a complex model that accounts for wind, shoe friction, and muscle fatigue, but you have to guess the runner's starting speed based on a blurry photo.
  • The Result: Because the starting photo (the public data) was blurry, your complex model started with the wrong speed. The simple rule (SGP4) actually worked better because it was "calibrated" to the same blurry photo. The complex model tried to be too smart with the wrong starting point and ended up further off track.
• The Finding: For most satellites and most timeframes, the simple "Official Map" (SGP4) was more accurate than the complex simulation. The complex simulation only won in one specific case: for the newest, largest satellites (v2-mini) after a long time (3–7 days). In that specific scenario, the simple map was failing so badly that even a slightly flawed complex model could do better.

3. The "Direction" Problem (Where does the error happen?)

The paper looked at where the satellites were wrong.

• The Analogy: Imagine the satellite is a train on a track. The error almost never happens by the train going off the tracks (sideways) or flying into the sky (up/down). The error happens almost entirely because the train is early or late on the track.
• The Finding: The satellites were almost always in the right "lane" but were off by seconds or minutes in their timing. This is because the biggest source of error is atmospheric drag (air resistance), which slows the satellite down or speeds it up along its path.

4. The "Solar Weather" Connection

The researchers tried to see if solar activity (sunspots and solar flares) made the predictions worse.

• The Analogy: Think of the atmosphere like a sponge. When the sun is active, it heats the sponge, making it expand and get "thicker" (more dense). This makes the satellites feel more air resistance.
• The Finding: They found a hint that when the sun is more active, the predictions get slightly worse, but the data wasn't strong enough to prove it with 100% certainty. It's like seeing a pattern in the clouds but not having enough rain to confirm a storm is coming.

The Bottom Line for Everyday People

• Trust the simple map for the short term: If you need to know where a Starlink satellite is in the next few hours, the free, simple data (SGP4) is good enough.
• Don't overcomplicate it: Unless you have a perfect starting point (which the public doesn't have), using a super-complex physics model doesn't help. In fact, it often makes things worse because it amplifies small errors in the starting data.
• Watch out for the "New" Satellites: The newest, biggest satellites are harder to track with the simple map over long periods. For those specific ones, a complex model might eventually be better, but only after waiting a few days.

In short: The paper proves that for public data, "less is more." A simple, well-tuned model often beats a complex one if the starting information isn't perfect. The best strategy is to update your data frequently (every few hours) rather than trying to predict too far into the future.

Technical Summary

Technical Summary: Empirical Comparison of SGP4 and High-Fidelity Propagation for Starlink TLEs

Problem Statement
The proliferation of Low Earth Orbit (LEO) megaconstellations, particularly SpaceX's Starlink, has shifted the operational context for Two-Line Element (TLE) data. While TLEs remain the de facto standard for academic and commercial space situational awareness (SSA), their accuracy is traditionally assessed against older, sparse datasets with daily update cadences. The Starlink constellation publishes TLEs approximately six times per day, rendering the "staleness" question one of hours rather than weeks. A critical gap exists in understanding how public TLEs propagate over these short horizons and whether high-fidelity numerical propagators (e.g., GMAT) offer improved accuracy over the standard SGP4 analytic model when initialized from public TLEs, which inherently contain orbit determination (OD) residuals.

Methodology
The study conducts a large-scale empirical sweep of 24,641 TLE pairs across 501 Starlink satellites over a one-month window in April 2026. The dataset is stratified by three altitude shells (540, 550, and 560 km) and three platform generations (v1.0, v1.5, and v2-mini).

1. Data Construction: Starting TLEs are sampled once per satellite-day. "Truth" is defined as the operator's subsequently released TLE (the "next-TLE"), evaluated at its own epoch using SGP4. This "next-TLE-as-truth" methodology serves as a self-consistency check against the operator's internal OD, acknowledging that it includes the OD residual at the target epoch.
2. Propagation: Each starting TLE is propagated forward to the target epoch ($t_j$) using two methods:
   • SGP4: The standard analytic model.
   • High-Fidelity (Hi-Fi): NASA's General Mission Analysis Tool (GMAT) configured with EGM2008 (70×70) gravity, NRLMSISE-00 atmospheric drag, conical-shadow solar radiation pressure (SRP), and third-body gravity from the Sun and Moon.
3. Filtering: Pairs are filtered to exclude intervals containing semi-major axis jumps exceeding 100 m, ensuring the "no-maneuver" assumption holds for the comparison.
4. Analysis: Position errors ($\|\Delta r\|$) are computed against the proxy truth. The study fits a power-law model $\|\Delta r(\Delta t)\| \approx A \Delta t^k$ to characterize error growth. It also performs pair-wise comparisons to determine the fraction of cases where Hi-Fi outperforms SGP4 and regresses the error coefficient $A$ against solar flux ($F_{10.7}$).

Key Findings

1. Error Growth Dynamics (H1):

   • Position error follows a power law across all cohorts.
   • SGP4 v1.x: Exhibits sub-linear growth ($k \lesssim 1$) at 540 and 560 km, and is sub-linear at 550 km, suggesting a mix of mean-motion bias and constant acceleration.
   • Hi-Fi v1.x: Generally exhibits super-linear growth ($k > 1$), indicating that unmodeled constant in-track acceleration (likely drag-related) dominates the error budget.
   • v2-mini Cohort: Shows super-linear growth ($k \approx 1.3\text{--}1.4$) for both propagators, but the Hi-Fi arm grows more slowly than SGP4 at long horizons ($\Delta t \ge 3$ days).
   • Magnitude: At the 6-hour horizon, the median error is $\sim 1$ km (limited by the OD residual floor). By 7 days, SGP4 errors reach $\sim 38$ km, while Hi-Fi errors reach $\sim 76$ km in the pooled v1.x cohort.

2. Propagator Performance (H2):

   • General Result: A high-fidelity propagator initialized from a public TLE does not improve median position error over SGP4 at any sampled horizon (6 h, 1 d, 3 d, 7 d). SGP4 wins on approximately 65–75% of pairs.
   • Mechanisms: This negative result is attributed to three factors:
     1. OD Residual Dominance: At short horizons, the error is dominated by the operator's initial state uncertainty, which affects both propagators.
     2. Kernel Alignment: The "truth" is constructed using SGP4. SGP4 predictions benefit from structural alignment with this truth (e.g., shared WGS-72 constants, Brouwer mean elements), whereas Hi-Fi propagators suffer from a dynamical reference mismatch (TEME to MJ2000Eq conversion and pseudo-osculating initial states).
     3. Spacecraft Property Bias: Hi-Fi propagators rely on external estimates of ballistic coefficients ($C_D A$). Errors in these public estimates accumulate as integrated drag errors, whereas SGP4's $B^*$ is fitted directly to the tracking data used to generate the TLE.
   • The Exception: The v2-mini cohort at long horizons ($\Delta t \ge 3$ days) is the only regime where Hi-Fi overtakes SGP4 on a majority of pairs (winning $\sim 55\text{--}73\%$ of pairs at 7 days). This occurs because the operator's fitted $B^*$ for the newer v2-mini satellites appears less stable (super-linear error growth), allowing the explicit drag modeling of Hi-Fi to outperform the SGP4 mean-motion approximation.

3. Solar Modulation (H3):

   • An exploratory regression of the SGP4 staleness coefficient against $F_{10.7}$ shows a positive slope at 560 km that clears conventional significance, consistent with thermospheric density gradients. However, the 30-day window of moderate solar activity limits the statistical power, and the result is read as direction-consistent rather than a calibrated measurement.

Significance and Practical Implications
The paper provides a "staleness atlas" for public Starlink TLEs, offering practitioners specific error budgets ($A, k$) stratified by altitude and generation. Its primary contribution is a counter-intuitive operational finding: investing in high-fidelity propagation from public TLE inputs is generally not beneficial unless the user also possesses improved initial states (e.g., from internal OD or third-party tracking) or is specifically forecasting long-term trajectories for the v2-mini generation.

The study establishes that for the vast majority of use cases, the "truth-construction kernel alignment" of SGP4 and the accumulation of spacecraft-property uncertainty in Hi-Fi models negate the theoretical advantages of high-fidelity force models. The per-cell power-law parameters provided serve as a benchmark for future enhanced-propagator research (e.g., SGP4-XP or differentiable SGP4) and as a baseline for uncertainty propagation studies in conjunction analysis.

Limitations
The analysis is constrained to a single constellation (Starlink), a 30-day window of moderate solar activity, and the "next-TLE-as-truth" methodology, which bounds the absolute error interpretation by the operator's OD residuals. The v2-mini drag area is scaled from v1.x data rather than fitted, introducing a modeling assumption, though sensitivity tests suggest this does not alter the primary conclusions.