Every diffusion model rests on a single, almost suspiciously simple idea:
Gradually destroy the structure in your data by adding noise until nothing is left but pure randomness — then learn to reverse that process, one small step at a time.
If you can learn to undo noise, you can start from pure noise and generate data. That’s it. Everything else — DDPM, score-based SDEs, probability-flow ODEs, flow matching — is a different lens on this same forward/reverse pair.
We corrupt a data point \(x_0\) over a continuous time \(t \in [0, 1]\) with a variance-preserving schedule:
\[x_t = \sqrt{\bar\alpha_t}\, x_0 + \sqrt{1 - \bar\alpha_t}\, \varepsilon, \qquad \varepsilon \sim \mathcal{N}(0, I),\]where \(\bar\alpha_t\) decays smoothly from \(1\) (all signal) to \(\approx 0\) (all noise). At \(t=0\) we have the data; as \(t \to 1\) every point relaxes into the same standard Gaussian ball, regardless of where it started.
Drag the slider below to see this happen — watch structured data (a spiral, two moons, or blobs) dissolve into Gaussian noise, and notice how the schedule \(\bar\alpha_t\) controls the pace:
A few things worth noticing as you play:
From this single forward/reverse picture we’ll build up, with code and an interactive widget for each step:
This first note is the conceptual on-ramp; the implementation-heavy sections are on their way (see below).
]]>A Gaussian in \(\mathbb{R}^d\) is fully described by a mean vector \(\mu\) and a covariance matrix \(\Sigma\):
\[x \sim \mathcal{N}(\mu, \Sigma).\]\(\mu\) says where the cloud sits; \(\Sigma\) says how it’s stretched and tilted. The diagonal of \(\Sigma\) sets the spread along each axis, and the off-diagonal sets the correlation — which rotates the cloud.
Here’s the move that everything depends on. Instead of sampling \(x\) directly, we sample standard noise \(\varepsilon \sim \mathcal{N}(0, I)\) and transform it:
\[x = \mu + L\,\varepsilon, \qquad \text{where } L L^\top = \Sigma.\](\(L\) is the Cholesky factor of \(\Sigma\); in the isotropic case it’s just \(x = \mu + \sigma\,\varepsilon\).) Because this map is differentiable in \(\mu\) and \(\Sigma\), gradients can flow through the act of sampling. That single fact is what makes VAEs, and the entire diffusion training objective, trainable by backprop.
Play with it below — the grey cloud is the fixed base noise \(\varepsilon\); the coloured cloud is \(x = \mu + L\varepsilon\). Notice how the same noise points get reshaped as you drag the sliders, and how \(\rho\) tilts the covariance ellipses:
Things worth internalizing:
Next up in the foundations: conditioning and marginalizing (how Gaussians combine), then measuring distance between distributions with KL — the last pieces before we noise our first dataset.
]]>Every day, more than 100,000 aircraft operate worldwide, transporting people and goods across continents and oceans. At any given moment, thousands of aircraft, piloted by humans or increasingly by software, share the same sky. Despite this immense scale and complexity, mid-air collisions are extraordinarily rare. This safety record is not accidental. It is the result of decades of engineering, procedures, training and coordination embodied into one of the most complex socio-technical systems ever built: Air Traffic Control (ATC).
Yet this success often masks how fragile and demanding the system truly is. This post is intended as a conceptual overview rather than a technical survey.
At a high level, Air Traffic Control has a simple mandate:
To ensure that aircraft remain safely separated while moving efficiently from origin to destination.
In practice, fulfilling this mandate requires far more than issuing instructions to pilots. ATC continuously orchestrates the motion of thousands of independent aircraft in shared airspace, balancing safety, efficiency, uncertainty, and human decision-making in real-time.
Contrary to popular belief, ATC is not a reactive system that responds to imminent danger. It is fundamentally predictive: controllers and automated tools constantly anticipate where aircraft will be minutes into the future and intervene before conflicts materialize.
The foremost responsibility of ATC is to prevent aircraft from coming dangerously close to one another. This is enforced through minimum separation standards, such as maintaining several nautical miles horizontally or thousands of feet vertically between aircraft.
What makes this challenging is that separation is not assessed based on current positions alone. Controllers must project aircraft trajectories forward in time, accounting for speed, heading, climbing rates, and anticipated maneuvers. A conflict, in ATC terms, is therefore a future event, not a present one.
Safety alone is not sufficient. ATC must also ensure that traffic flows smoothly through the airspace. This includes sequencing aircraft for landing, managing merges at busy waypoints and preventing congestion from cascading across regions.
Many of these decisions are strategic rather than tactical, made tens of minutes or even hours in advance. Delays, reroutes, and ground holds are often applied proactively to preserve stability downstream.
Despite increasing automation, humans remain central to ATC operations. Controllers synthesize radar data, procedures, weather information, and experience to make judgements under time pressure. Pilots execute instructions while managing aircraft performance and onboard systems.
ATC is therefore not just a technical system but a human-in-the-loop control system, where workload, trust and interpretability are as critical as algorithmic correctness.
Given these responsibilities, it is natural to ask why ATC is so hard to automate or scale. The difficulty arises from three fundamental properties of the problem.
Aircraft do not move on fixed tracks. They operate in continuous 3D space, with continuously varying speed, heading and altitude. Even small deviations can propagate over long distances and time horizons.
As a result, ATC cannot rely on discrete planning or simple enumeration. Instead, it must reason over an effectively infinite set of possible trajectories.
Aircraft do not interact in isolation. Resolving a conflict between two aircraft can affect many others: slowing one aircraft may delay those behind it; diverting an aircraft laterally can create new conflicts elsewhere.
This coupling means that local decisions often have global consequences. ATC is therefore best understood as a multi-agent system with tightly coupled dynamics, rather than a collection of independent pairwise problems.
Every ATC decision is made with imperfect information. Weather forecasts are uncertain, aircraft performance varies, pilot response times differ, and surveillance data is noisy. Yet safety constraints must be respected at all times.
To manage this uncertainty, ATC relies on conservative buffers and procedural margins, often trading efficiency for robustness.
Historically, ATC has addressed this complexity through structure. Airspace is divided into sectors, traffic flows are organized along standard routes, and procedures define how aircraft climb, descend, merge, and land.
Human controllers oversee limited regions of airspace, handling off responsibility as aircraft move between sectors. This division of labor has proven extraordinarily effective, enabling safe operations at a global scale.
The effectiveness of traditional ATC rests on an implicit assumption: that traffic density remains manageable by human cognition and procedural control.
That assumption is increasingly challenged. Commercial air traffic continues to grow, while new entrants, such as drones, and electric Vertical Takeoff and Landing (eVTOL) aircraft promise orders of magnitude more vehicles operating at lower altitudes.
In these emerging environments, the number of agents, frequency of interactions, and variability of behavior may exceed what centralized, human-centric control can safely manage.
These trends do not imply that classical ATC has failed. Rather, they suggest that its underlying principles may not scale indefinitely.
New airspaces may require approaches that are:
Meeting these requirements has led researchers to explore ideas beyond traditional aviation, drawing from robotics, control theory, artificial intelligence, and even physics.
One particular promising direction treats air traffic not as a collection of independent aircraft, but as a collective motion system, akin to particles moving within a structured flow. Instead of resolving conflicts after they arise, the airspace itself can be designed so that conflicts are naturally avoided.
This perspective motivates the next post in this series:
Fluid Dynamics Meets Air Traffic Control: A Novel Approach to Conflict Resolution
where we will explore how ideas from fluid dynamics can inform scalable, conflict-free traffic management in the future airspace.
Air Traffic Control is one of the great engineering achievements of modern society. Its success rests on careful prediction, structured procedures, and human expertise operating under uncertainty.
As aviation enters an era of higher density and greater autonomy, understanding the intricacies of ATC is a necessary step toward reimagining how the sky can be safely and efficiently shared by all.
If you wish to cite this article in academic work, please use the following format:
APA
Zongo, A. (2025). Aircraft Traffic Control: Managing Order in a Crowded Sky. Retrieved from https://alex-zongo.github.io/posts/2025/12/aircraft-traffic-control/. Archived on Zenodo: https://doi.org/10.5281/zenodo.17970035
BibTex
@misc{zongo2025aircraft,
title={Aircraft Traffic Control: Managing Order in a Crowded Sky},
author={Zongo, Alex},
year={2025},
doi={10.5281/zenodo.17970035},
url={https://alex-zongo.github.io/posts/2025/12/aircraft-traffic-control/},
note={Accessed: YYYY-MM-DD}
}