The Billion-Dollar Error Term

The general struggle for existence of animate beings is not a struggle for raw materials—these, for organisms, are air, water and soil, all abundantly available—nor for energy which exists in plenty in any body in the form of heat, but a struggle for low entropy, which becomes available through the transition of energy from the hot sun to the cold earth.
— Ludwig Boltzmann, 'The Second Law of Thermodynamics' (1886)

In the winter of 1712, Thomas Newcomen set his atmospheric engine to work pumping water from a coal mine in Dudley Castle, Staffordshire. The engine consumed coal at a rate that would have appalled any later engineer — burning roughly one hundred pounds of fuel to produce work that a modern turbine could deliver with three. But the mine owners paid willingly, because the coal was already there, extracted from the same seams the engine was draining, and the alternative was flooded shafts and abandoned pits. The engine's inefficiency was invisible to the account books. What the books recorded was output: coal lifted from depths that hand-pumps could not reach, sold at prices that covered the engine's appetite and left a profit. The gap between the energy consumed and the work performed — a gap that would narrow over three centuries from ninety-nine percent waste to under forty percent — appeared nowhere in the mine's ledger. It was, in the language that economists would later develop, part of the residual: the portion of output growth that the measured inputs could not explain.

If the ledger-and-law distinction marks one boundary, growth accounting exposes another. Modern models assign a large share of output growth to an error term, and that term is doing real explanatory work. The gap runs between output as a number and output as a physical process.

Abramovitz and the Residual

In 1956, Moses Abramovitz set out to decompose American economic growth from 1869 to 1953 into its constituent sources.(Abramovitz 1956) The accounting framework he used was straightforward: add up the contributions of labor and capital, and whatever output remains unexplained is attributed to a residual. Over that period, net output per capita had roughly quadrupled. What Abramovitz found was that the measured inputs, more workers, more machines, more hours, explained less than half of that increase. The rest came from somewhere else. He called it, with characteristic dryness, "a measure of our ignorance."(Abramovitz 1956, 11)

The phrase was honest, but the discipline kept the residual because it traveled well. It allowed comparison without inventorying the physical particulars of each economy, its engines, grids, fuels, machines, and constraints. The cost of that portability was explanatory blur. The residual became the place where mechanism could be implied rather than specified.

Robert Solow formalized the decomposition a year later, in a 1957 paper that became one of the most cited in economics.(Solow 1957) He derived growth accounting from a neoclassical production function and gave the residual a name that endured: "total factor productivity," or TFP.

The Solow residual became the workhorse of growth economics. Cross-country comparisons, policy evaluations, and long-run projections all depended on it. When a country grew faster than its accumulation of labor and capital could explain, the difference was attributed to TFP growth. When a country stagnated despite investing heavily, the difference was attributed to TFP decline. The residual could absorb good news and bad news alike, and it did not require the analyst to specify what, physically, was changing.

Economists have never believed the residual is a single substance. It contains measurement error, changing utilization, imperfect price deflators, shifts in labor quality, sectoral reallocation, management practices, learning, and innovation. Entire literatures attempt to unpack it. But precisely because it is a bundle, it allows energy and material throughput to disappear into the label.

How large is the residual in practice? For the United States between 1948 and 1998, Hulten's survey of the growth-accounting literature finds that TFP growth accounts for roughly fifty to sixty percent of the increase in output per worker.(Hulten 2001) The residual is not uniformly large everywhere: Alwyn Young's decomposition of East Asian growth during the high-growth decades found that measured input accumulation explained most of the output gains, leaving a smaller residual, which itself is evidence that the residual's size varies with the energy and institutional substrate.(Young 1995) But in the canonical case of the industrialized West, the error term carries most of the phenomenon. When it does, the framework is missing something structural.

The Solow framework achieved its reach through abstraction. Economists wanted a model that could handle many countries and many periods without requiring detailed knowledge of each country's physical infrastructure, and the residual delivered exactly that. By treating technology as a residual, the model remained agnostic about the mechanisms of improvement. Portability came at a cost.

The abstraction did not arrive all at once. Consider what happened when Solow wrote his production function: Y = A·f(K, L). Capital entered as a dollar-denominated stock and labor as hours worked. A power station and a shopping mall contributed the same way—as "K." A coal miner and a management consultant contributed the same way—as "L." The aggregation was deliberate. It purchased generality: the same equation could describe the United States in 1950 and India in 1980 without requiring an inventory of turbines, transmission lines, or fuel stocks. But generality has a price. When capital is measured in dollars rather than in machines-that-convert-energy, the production function cannot distinguish between an economy that invested in combined-cycle gas turbines and one that invested in suburban office parks. Both show up as capital deepening. Only one changes what the economy can physically do.

What the Residual Absorbs

Consider what the residual might be absorbing. Stand on the floor of the Shidongkou power station in Shanghai, and the residual becomes physical. The original plant, built in 1992 with subcritical coal technology, converted roughly thirty-three percent of the coal's energy into electricity. The remainder exited through the stack as waste heat, a plume visible for miles. When the plant was retrofitted with ultra-supercritical boilers in the 2010s — operating at higher temperatures and pressures, with better metallurgy in the turbine blades and more efficient heat exchangers — the conversion efficiency rose above forty-four percent. The same coal, the same labor, the same capital stock (adjusted for the retrofit cost). Eleven percentage points more of the energy in each ton of coal now became electricity instead of atmospheric heat. That improvement appears in China's national accounts as higher total factor productivity. It is not a mystery. It is better engineering applied to a thermodynamic process, measurable in joules.

A factory installs a more efficient boiler. It can now produce the same output with less fuel, or more output with the same fuel. That improvement will show up in the data as higher productivity, more output per unit of measured input. But the improvement is not magic, and it is not "ideas" in some disembodied sense. It is a physical change in the conversion efficiency of a machine, the fraction of the energy in the fuel that ends up doing useful work rather than escaping as waste heat.

Resistance in the wires is lower and combustion more complete. The turbine blades are shaped to extract more momentum from the steam. The heat exchanger transfers more energy before the exhaust escapes. These are engineering facts, measurable in joules and watts. Newcomen's first atmospheric engine converted less than one percent of the coal's energy into useful work. Watt's improved design reached two to three percent. A modern combined-cycle gas turbine converts over sixty percent.(Smil 2017) The curve moved in steps, punctuated by decades of stagnation and bursts of improvement, but the direction is unmistakable. That trajectory sits inside the Solow residual, unnamed.

The Corliss Engine

At the 1876 Philadelphia Centennial Exhibition, visitors from thirty-seven countries stood before a single machine that made the trajectory physically visible. George Corliss's double-acting duplex engine, forty-five feet tall, its thirty-foot flywheel weighing fifty-six tons, drove eight main distribution shafts through some thirteen thousand feet of overhead line shafting, transmitting fourteen hundred horsepower to roughly eight hundred machines across thirteen acres of Machinery Hall.(Bryant 1991, 207–208) It ran for six months without failure. The engine was not an "idea." It was a device that converted coal into coordinated mechanical motion at a scale and reliability that would have been inconceivable to Newcomen. What visitors witnessed was conversion efficiency made visible, and what the Solow residual hides is precisely this: not metaphysical progress but mechanical fact.

The same logic applies at larger scales. Consider the Ruhr Valley. In 1850 its coal mines produced roughly two million tonnes a year, much of it hauled by barge along the Rhine. By 1900 the figure was sixty million tonnes, and the Ruhr had become the industrial heartland of continental Europe: Krupp's steelworks at Essen, Thyssen's blast furnaces at Duisburg, a canal network purpose-built to move fuel to furnace. Every one of those facilities appears in the national accounts as "capital." Every one of them was, at bottom, energy-conversion infrastructure, machinery for turning the chemical energy stored in Carboniferous-era coal into motion, heat, and steel. The Solow framework records the accumulation; it does not record the substrate that made the accumulation productive.

The pattern recurs at every fuel transition. Coal is denser and more storable than wood; oil is liquid and therefore easier to move through pipes and tanks; electrification transmits power over long distances and applies it precisely where and when it is needed. Each transition required new systems: mines, rail lines, pipelines, refineries, transmission grids. All built over decades, financed and maintained like any other capital stock, but with constraints and failure modes that are not well summarized by the word "capital." Each of these transitions shows up in growth accounting as a rise in TFP, because output increases faster than the measured inputs of labor and capital. The transitions are legible. They are changes in the energy substrate.

Technology as Conversion

The modern habit is to treat growth as a triumph of ideas, and the habit is not entirely wrong. Ideas do matter, and the diffusion of useful knowledge has been one of the great engines of prosperity. But the record reads more like a succession of improvements in energy conversion than like a smooth accumulation of disembodied insights. The transitions that show up as discontinuities in the growth data, the British Industrial Revolution, the electrification of manufacturing, the postwar boom in the United States, are also transitions in how societies captured, converted, and deployed energy. Ideas about how to do things more efficiently are, in practice, often ideas about how to extract more useful work from a given quantity of fuel. The value of those ideas depends on having the fuel to apply them to.

"Technology," in this framing, is often the name we give to conversion efficiency once we have forgotten its physical origin.

A steam engine is a technology, but it is also a device for turning the chemical energy in coal into mechanical motion, with a measurable efficiency that improved from Newcomen to Watt to Corliss. An integrated circuit is a technology, but it is also a device for switching electrical states at a certain energy cost per operation, and the progress of computing has been, among other things, a story of reducing that cost by many orders of magnitude. When we say that technology explains growth, we are often saying, without quite realizing it, that improvements in energy conversion explain a substantial share of that growth.

Whether one calls this ignorance or abstraction, the omission is systematic. The framework counts labor and capital in value terms while treating the energy-conversion machinery that makes them productive as an implied background condition. What changes when that condition is modeled explicitly, when energy enters not as an expenditure line but as useful work?