Chapter 18 How science corrects error
9 minute read
Why did the Industrial Revolution ignite in 18th century Europe rather than in Song dynasty China, imperial Rome, or the Abbasid Caliphate? Each of those earlier civilizations possessed brilliant engineers, bustling commercial cities, and grand libraries. What set early modern Europe apart was an epistemic engine forged a century earlier, between 1543 and 1687. Across the generations from Nicolaus Copernicus, Galileo Galilei, and Johannes Kepler to Francis Bacon and Isaac Newton, humanity invented a self-correcting algorithm for discovering causal laws: the scientific method.
To understand how a species of storytelling primates learned to decode quarks, galaxies, and genomes, we must look inside the logical and mathematical machinery of scientific error elimination.
18.1Falsification and the web of belief
In 1739, the Scottish philosopher David Hume uncovered a devastating logical fracture at the foundation of empirical knowledge, known ever since as the problem of induction. Suppose a naturalist observes 10 million swans across the rivers of Europe, and every single one of them is white. No finite collection of positive observations can ever deductively prove the universal law that all swans are white. Logic offers no guaranteed bridge from the observed past to the unobserved future. A single ship landing in Western Australia can encounter a black swan on the very next observation and shatter the generalization. If enumerative induction cannot guarantee the truth of any universal law, what gives scientific theories their rational authority?
Nearly two centuries later, in his 1934 book The Logic of Scientific Discovery, the philosopher Karl Popper resolved Hume's paradox. He pointed out a profound deductive asymmetry between verification and falsification. No pile of confirming instances can ever prove a universal theory true. Yet a single genuine counterexample deductively proves the theory false through the classical logical rule of modus tollens, .
From this logical asymmetry, Popper drew his famous line of demarcation between genuine science and pseudoscience. What makes a theory scientific is not that it accumulates confirming anecdotes. An astrologer or a conspiracy theorist can always interpret any daily event as fresh confirmation of a vague doctrine. A theory is scientific if and only if it is falsifiable. It must stick its neck out by making sharp, quantitative predictions that strictly forbid specific empirical outcomes.
Recall from Chapter 4 how general relativity bends grazing starlight, a prediction put to a decisive test during the total solar eclipse of 29 May 1919. While Newtonian gravity implied that the Sun's mass should bend grazing starlight by 0.875 arcseconds, Einstein's curved spacetime field equations predicted a deflection of exactly twice that amount:
When the astronomer Arthur Eddington photographed the Hyades star cluster around the eclipsed Sun from the island of Príncipe, Einstein's theory risked immediate destruction. Had the starlight shifted by zero or by the Newtonian half-angle, general relativity would have been refuted on the spot. Science advances not by piling up inductive certainties, but through an evolutionary struggle of bold conjectures and severe empirical refutations. Like natural selection in the biosphere, experiment ruthlessly weeds out false hypotheses. Only fitter approximations to reality survive.
Yet as Pierre Duhem, Willard Van Orman Quine, and Thomas Kuhn demonstrated, real laboratory science is subtler than textbook modus tollens. A core physical theory never steps before the tribunal of observation alone. To derive a concrete prediction about a telescope plate or a voltmeter needle, a scientist must combine the core theory with a bundle of auxiliary assumptions. Those auxiliary hypotheses describe instrument calibration, atmospheric refraction, sample purity, and the absence of unseen disturbing forces.
When a careful experiment produces a recalcitrant result that contradicts the prediction, deductive logic tells us only that at least one strand in the combined bundle is false. Logic alone never points a finger at the guilty premise. It cannot tell us whether to discard the sacred core theory or to blame one of the humbler auxiliary assumptions surrounding it.
Two famous chapters in 19th century astronomy illustrate both halves of this dilemma. By the 1840s, precise telescopic tracking showed that the planet Uranus was straying systematically from the orbit prescribed by Isaac Newton's law of universal gravitation. Did astronomers declare Newton falsified and abandon classical mechanics? Not at all. Working independently in Paris and Cambridge, Urbain Le Verrier and John Couch Adams kept Newton's core law intact. Instead, they rejected the auxiliary assumption that the solar system contained only seven planets. Treating the orbital deviations as an inverse problem, they calculated the mass and sky coordinates of an unseen eighth planet pulling on Uranus from beyond. On the night of 23 September 1846, Johann Galle aimed the Berlin Observatory refractor at Le Verrier's coordinates and discovered Neptune within one degree of the predicted point. An apparent refutation of Newtonian physics turned into its most spectacular triumph.
Flush with victory, Le Verrier soon turned to a second orbital mystery at the opposite edge of the solar system. Mercury's elliptical orbit rotated around the Sun faster than Newtonian perturbations from the known planets could explain, leaving an anomalous perihelion advance of 43 arcseconds per century. Once again, Le Verrier protected Newton's core law by blaming the planetary census, postulating a small, scorching inner world named Vulcan orbiting inside Mercury's path. For decades, eclipse chasers scanned the solar glare for Vulcan, yet the planet was never there. This time the auxiliary patch failed. The 43-arcsecond anomaly stood waiting until 1915, when Einstein showed that it flowed naturally from the curvature of spacetime near the Sun without any hidden planet at all.
How do scientists know when defending a core theory through auxiliary adjustments is sound normal science, as Kuhn called it? And how do they know when that defense has decayed into stubborn excuses that call for a scientific revolution? In 1970, the philosopher Imre Lakatos synthesized Popper and Kuhn through his methodology of scientific research programmes. Every mature research programme consists of an inviolable hard core of foundational laws surrounded by a flexible protective belt of auxiliary models and initial conditions. Modifying the protective belt to absorb an anomaly is scientifically progressive only when the revised auxiliary model predicts novel, previously unexpected empirical facts. Those novel predictions must then be confirmed by independent observation, just as Le Verrier's model predicted the visible disk of Neptune. Conversely, theorists sometimes tack on after-the-fact adjustments that rescue a model from each new failure without predicting any fresh phenomena. When that happens, the research programme is degenerating and ready to be overthrown by a bolder rival.
18.2Bayesian reasoning and the social machinery of truth
In the second half of the 20th century, mathematicians and philosophers united Popper, Duhem, Quine, and Lakatos inside a single calculus: Bayesian probability theory. Building on the earlier work of Thomas Bayes, Pierre-Simon Laplace, and Harold Jeffreys, the physicist Richard Cox proved a foundational theorem in 1946. He showed that any system of plausible reasoning that assigns numerical degrees of belief to hypotheses without internal contradiction must obey the mathematical rules of probability. Independently, Frank Ramsey and Bruno de Finetti proved the same necessity through their Dutch book theorems. Any bettor whose credence levels violate the axioms of probability can be tricked into a sequence of wagers that guarantees a sure loss under every possible outcome.
To reason coherently under uncertainty, an epistemic agent must update the prior probability of each rival hypothesis given background knowledge. When new empirical evidence arrives, the agent updates those beliefs according to Bayes' theorem:
Every major insight in the philosophy of science emerges directly from the structure of this equation. First, Popper's deductive falsification appears as the noiseless limit of Bayesian updating. Whenever a hypothesis strictly forbids an observed outcome, the likelihood term in the numerator vanishes, driving the posterior credibility immediately to zero. In realistic settings, each hypothesis is a composite bundle of a core law and an auxiliary model. Repeated failures to observe predicted auxiliary rescues steadily drain probability from degenerating patches until a rival core theory overtakes them.
Second, Bayes' theorem explains mathematically why bold, precise theories defeat vague, flexible ones, giving a rigorous proof of William of Ockham's razor. Following Harold Jeffreys and David MacKay, compare two competing models of a physical quantity over a broad prior range. A simple parameter-free theory makes a sharp prediction centered at a specific value with narrow measurement uncertainty. A rival complex theory introduces an adjustable parameter that allows it to fit almost any conceivable outcome across the entire prior interval.
Because the prior predictive probability density over all outcomes must integrate to unity, hedging carries a steep price. By spreading its predictive density uniformly across the full prior width, the flexible theory assigns only a tiny density—the reciprocal of that width—to the actual measurement. Equivalently, marginalizing its free parameter multiplies its peak likelihood by the dimensionless Occam factor, the ratio of posterior measurement uncertainty to prior parameter width. Meanwhile, the austere parameter-free theory concentrates its predictive density inside the narrow measurement window around the predicted value, achieving a high probability density equal to the reciprocal of the uncertainty. When the measurement lands on target, the ratio of marginal likelihood densities—the Bayes factor—penalizes the over-parameterized model by the inverse of the Occam factor, . Simpler theories win not because nature has a mystical fondness for brevity, but because a theory that can explain everything predicts nothing in particular.
We can push this connection one step deeper by wedding Bayes' theorem to the algorithmic information theory of Ray Solomonoff and Andrey Kolmogorov from Book I. On a prefix-free universal Turing machine—where Kraft's inequality guarantees that exponential program weights sum to at most one—Solomonoff's universal prior weights each hypothesis by . Here the exponent is the Kolmogorov complexity from Chapter 3, the length in bits of the shortest program that generates the hypothesis. Taking the negative base-two logarithm of Bayes' theorem then yields Jorma Rissanen's principle of minimum description length:
The first term on the right measures the algorithmic complexity in bits of the theoretical laws themselves. Meanwhile, the second term measures the bits required to list the residual experimental anomalies that the theory fails to explain. A bloated theory with an extra clause for every data point has a huge model complexity. Conversely, an overly simplistic dogma leaves a massive unexplained residual. The best scientific theory minimizes their sum. Science is the collective algorithmic compression of the cosmos. It distills terabytes of raw sensory observations into a handful of compact differential equations and symmetry principles.
One final question remains. If Bayesian updating and algorithmic compression require ruthless honesty toward counterexamples, how do mortal human beings actually carry them out? Left to themselves, individual scientists are every bit as prone to confirmation bias, pride, tribalism, and fondness for their own pet hypotheses as anyone else. Newton feuded bitterly with Gottfried Wilhelm Leibniz over the calculus. Max Planck once remarked that a new scientific truth triumphs not by convincing its opponents, but because its opponents eventually die and a new generation grows up with it.
As the sociologist Robert Merton and the philosopher David Hull showed, science transcends individual bias through the design of its social institutions. Norms of open publication, shared raw data, blind peer review, and independent experimental replication pit the ambitions of rival researchers directly against one another. In the credit economy of science, a researcher wins very little prestige by merely repeating a famous professor's consensus opinion. The highest honors, prizes, and historical immortality go to the experimenter who uncovers a stubborn anomaly or the theorist who replaces a reigning paradigm with a deeper law. By aligning personal ambition with public falsification, the institutional architecture of science turns competitive, fallible human minds into an emergent collective intelligence. Soon that epistemic engine would uncover the physical laws of information itself, allowing humanity to distill mathematical logic into silicon and build thinking machines.