The war of signals

These articles are not designed to offer definitive answers or fixed positions. Instead, they are explorations—reflections grounded in history, data, and evolving thought. Our aim is to surface questions, provide context, and deepen understanding. We believe education thrives not in certainty, but in curiosity.

Prologue: A Whisper Beneath the Waves

In the mid-20th century, amid the geopolitical tension of the Cold War, the U.S. Navy made a stunning discovery. Deep below the ocean’s chaotic surface, an extraordinary acoustic corridor spanned entire ocean basins. It was a natural waveguide, later named the SOFAR channel (Sound Fixing and Ranging), where low-frequency sound waves could travel thousands of miles with barely any loss.

Whales had used it for millennia.

Now, humans would too.

To exploit this marine whisper channel, the U.S. Navy deployed vast hydrophone arrays—an underwater ears network called SOSUS (Sound Surveillance System)—across the Atlantic seafloor. But listening was not enough. They needed to detect, decide, and track. The faint rumble of a Soviet submarine could be easily buried in thermal noise, snapping shrimp, or the moaning of a blue whale.

At the heart of solving this was a deceptively simple formula that quietly became the skeleton key to modern detection systems:

$$\Lambda(x) = \frac{p(x|H_1)}{p(x|H_0)} \gtrless \eta$$

This likelihood ratio test became the central dogma of statistical signal processing—a way to decide whether the data $x$ suggested the presence of a signal ($H_1$) or just noise ($H_0$). And from the abyss of the SOFAR channel to the algorithms in your smartphone, this equation still governs the way machines listen.

I. The Equation is Born: Signal vs. Noise

The Cold War didn’t invent the problem of signal detection—it merely raised the stakes. In WWII, radar operators faced it too: is that blip an enemy plane, or is it a bird, a cloud, or just random noise?

But it wasn’t until mathematicians like Harry L. Van Trees systematized the field that a unified theory emerged. Drawing on Bayesian and Neyman-Pearson decision frameworks, the likelihood ratio test gave engineers a method to quantify belief under uncertainty. Here's how it works:

• Assume two competing explanations for observed data $x$:

  • $H_0$: noise only

  • $H_1$: signal plus noise

• Evaluate how likely the data is under each. The ratio $\Lambda(x)$ compares these two beliefs.

• Decide “signal” if $\Lambda(x) > \eta$, where $\eta$ is a threshold calibrated to control false alarms.

It was revolutionary. With this test, engineers could tune submarine detection systems to maximize the chance of hearing Soviet traffic without being fooled by snapping shrimp.

II. Listening Like a Whale: The SOFAR Channel

Nature had already solved the problem of long-distance underwater communication. Blue and fin whales, among the loudest animals on Earth, vocalize at low frequencies (10–40 Hz) precisely because those sounds can glide effortlessly through the SOFAR channel.

The U.S. Navy, led by underwater acoustics pioneers like Maurice Ewing and later strategists at ARPA (which would become DARPA), began exploiting this channel to hear faraway undersea activity. The goal: track enemy submarines across the Atlantic without ever seeing them.

But the signals were incredibly faint and the noise overwhelming. Under these conditions, human judgment falters. What you need is mathematics. You need the likelihood ratio.

With it, sonar technicians could automate the detection of “pings” in noisy environments, balancing:

• Probability of Detection: How often do we correctly detect a submarine?

• False Alarm Rate: How often do we mistake a whale or ambient noise for a sub?

This trade-off was visualized in the now-ubiquitous ROC curve, a direct child of the likelihood ratio framework.

III. From the Deep to the Sky: The Equation Takes Flight

The elegance of $\Lambda(x)$ is its universality. Once proven in sonar, it migrated to:

• Radar systems, tracking planes across cluttered skies.

• Communications, deciding whether a bit was 0 or 1 through a noisy channel.

• Space systems, where the Voyager probes and deep-space arrays decoded whispers from the edge of the solar system.

Harry L. Van Trees, in his magnum opus Detection, Estimation, and Modulation Theory, formalized the principles, extending them to time-varying signals, random processes, and dynamic systems. His work taught engineers how to shape $\Lambda(x)$ when the data came in streams, not snapshots—a critical development for tracking moving targets like missiles and satellites.

IV. Estimating More Than Presence: A Continuum of Inference

Detection is just the binary cousin of estimation, and here the narrative deepens.

Suppose we detect a submarine. Now we want to estimate its bearing and speed. In estimation theory, the goal shifts to extracting continuous parameters from noisy observations.

This is where the Kalman filter entered Cold War operations. Feeding sonar measurements into a recursive algorithm, the Kalman filter estimates a target’s path with minimum error variance, based on prior knowledge and new observations. It is a dynamic version of $\Lambda(x)$, reimagined for evolving systems.

Again, Van Trees' work stitched together the disciplines: detection, estimation, modulation—all united under a probabilistic framework.

V. The Digital Age: Echoes of the Formula

Today, the likelihood ratio still lives, quietly nested inside technologies most people never associate with submarines:

• Spam filters weigh the likelihood a message is spam vs. not.

• Medical diagnostics evaluate the probability of disease given symptoms.

• Speech recognition, fraud detection, neural decoding—all implement decision rules born from $\Lambda(x)$.

Even in machine learning, classifiers like logistic regression and Naive Bayes echo its logic. Deep inside the hidden layers of neural nets, the spirit of comparing competing models of the data still pulses.

VI. Epilogue: The Machine that Learned to Listen

The SOFAR channel is a metaphor for our world—signal hidden in an ocean of noise. And $\Lambda(x)$ is the formula that taught machines to listen as whales do: patiently, probabilistically, and with mathematical precision.

In Harry L. Van Trees’ hands, this formula matured into a full theory that underpins our modern ability to detect meaning in chaos. And while the Cold War is over, the challenges it posed—how to detect threats, extract information, and make decisions under uncertainty—remain more relevant than ever.

So next time your phone filters out background chatter during a call, or a satellite tracks a hurricane from orbit, remember: it’s all just the echo of that Cold War equation, still listening.

🔎 Further Reading & Source Notes

1. Harry L. Van Trees – Detection, Estimation, and Modulation Theory, Part I (1968)
   The foundational text for statistical signal processing. Written for engineers but revered by economists, radar scientists, and even modern AI researchers. This is the origin of the likelihood ratio as we know it.

2. Urick, R. J. – Principles of Underwater Sound (1983)
   The definitive book on how sound travels through water, including Cold War sonar systems and the SOFAR channel. A brilliant read on how the U.S. Navy learned to listen like a whale.

3. Neyman & Pearson (1933) – “On the Most Efficient Tests of Statistical Hypotheses”
   The original paper that inspired the likelihood ratio test. Dense, but a good reminder that every modern ML classifier owes a debt to early 20th-century statistics.

4. Baggeroer, Kuperman & Mikhalevsky – “Matched Field Methods in Ocean Acoustics” (IEEE, 1993)
   A technical but fascinating look at how to detect submarines using probability models and ocean acoustics. Especially relevant for understanding how the likelihood test works in the field.

5. Blackman & Popoli – Design and Analysis of Modern Tracking Systems (1999)
   Want to understand how radar and sonar systems follow moving targets? This is where estimation theory meets engineering elegance.