pure

← Home

The Protocol Argument

Abstraction does for intellectual traffic what a common protocol does for a communications network: it replaces an unmanageable collection of pairwise translations with a single interoperable language.

“Failure to abstract” is difficult to establish directly, because the counterfactual — how quickly ideas would have travelled under a better vocabulary — can't be observed. The strongest evidence comes from cases with at least one of these signatures: the same object or algorithm was independently reinvented in several fields; the same invariant acquired several unrelated names; important theory remained buried in special cases; or a later common abstraction immediately transferred proofs, algorithms, complexity results, or applications. By those standards, the following are unusually strong.

1. Matroid girth, compressed-sensing spark, Kruskal rank, and coding-theoretic stability were the same invariant

This is perhaps the cleanest exact example. Given a matrix \(X\), several communities independently defined closely related quantities:

The identities are exact:

\[ \operatorname{girth}(X) = \operatorname{spark}(X) = \operatorname{krank}(X)+1, \]

with a corresponding equivalence for \(k\)-stability. Lim and Comon observe that vector-matroid girth appeared only rarely in conventional linear-algebra discourse and was consequently reinvented under different names. Once the identification is made, results transfer immediately: hardness results for computing vector-matroid girth become hardness results for spark, Kruskal rank, and stability; uniqueness conditions for sparse recovery can be compared directly with uniqueness conditions for tensor decompositions.

\[ \boxed{ \text{One invariant} \quad\longrightarrow\quad \text{four literatures, four names, repeated theorems} } \]

The matroid abstraction acts as a lossless translation layer. It doesn't merely reveal a poetic resemblance; it proves that a theorem stated in one vocabulary has already answered a question being posed in another. For an essay defending pure abstraction, this may be the best opening example because the transfer is mathematically indisputable.

2. Factor graphs unified algorithms that several fields had separately invented

Coding theory had trellises, the Viterbi algorithm, BCJR decoding, Tanner graphs, turbo decoding, and LDPC decoding. Statistics and signal processing had hidden Markov models, the forward–backward algorithm, Kalman filtering, and smoothing. Artificial intelligence had Bayesian networks and belief propagation. These were not merely analogous techniques — at an appropriate level of abstraction, they were instances of essentially the same operation: factor a global function into local factors and pass messages that successively eliminate variables.

Hans-Andrea Loeliger's historical account notes that the BCJR algorithm remained relatively obscure and was independently reinvented several times, and describes the later recognition that coding theory's sum-product algorithm and artificial intelligence's belief propagation were the same fundamental procedure. The factor-graph formulation then made connections to hidden Markov models, Kalman filtering, recursive least squares, FFTs, and iterative receiver design explicit; some new receiver architectures became almost mechanically derivable once the common representation existed.

The information failure here was not absence of intelligence or results. It was absence of a canonical intermediate language. Each community stored the same computation in field-specific coordinates:

\[ \text{trellis},\qquad \text{Bayesian network},\qquad \text{state-space model},\qquad \text{parity-check graph}. \]

A paper indexed under one vocabulary was functionally invisible to researchers searching under another. Why this is a particularly strong example: the abstraction did more than rename existing algorithms. It exposed interchangeable modules, made combinations of algorithms natural, and allowed practical lessons from one field — such as the usefulness of iterative message passing on graphs with cycles — to migrate into another.

3. Reverse-mode automatic differentiation was repeatedly rediscovered

Reverse-mode automatic differentiation is the general computational principle underlying backpropagation. Given a scalar output computed by a potentially enormous program, it propagates derivatives backward through the program's computational graph. Its decisive property: the cost of obtaining the complete gradient can remain a small multiple of the cost of evaluating the original program, largely independent of the number of inputs.

The same basic method appeared separately in:

Andreas Griewank's historical study explicitly describes numerous independent incarnations and even says that he himself rediscovered the method. Seppo Linnainmaa had already represented a calculation as a computational graph and reversed its chain rule, but his work initially received little attention and wasn't presented primarily as a general “cheap gradient” method.

The missing abstraction was:

\[ \text{arbitrary computation} \longrightarrow \text{directed acyclic computational graph} \longrightarrow \text{reverse accumulation of local derivatives}. \]

Without that formulation, every field encountered a local story: sensitivity of an engineering system, adjoint equations for a differential model, training weights in a neural network, differentiating a computer program. With it, these become the same algorithm. This example is especially consequential because fragmentation did not merely duplicate terminology — it delayed the circulation of a powerful complexity fact: a high-dimensional gradient is not necessarily \(n\) times as expensive as the function whose gradient is required.

4. The EM algorithm existed as many disconnected special procedures before its general formulation

Before Dempster, Laird, and Rubin's 1977 paper, procedures now recognized as expectation–maximization had appeared in problems involving missing observations; censored or truncated data; finite mixtures; variance components; latent variables; and some Bayesian and robust-estimation problems. The authors explicitly noted that the algorithm had been proposed repeatedly in special circumstances. Their contribution was to formulate the common incomplete-data structure and prove general properties, including monotonic improvement of the likelihood under the EM update — theory they described as having remained buried in special-example literatures.

The discussion accompanying the paper is especially revealing: one discussant explained that convergence arguments that had been difficult in particular models became strikingly simple after the problem was generalized. The broad formulation also taught statisticians how to recognize EM in problems that did not superficially resemble the familiar missing-data examples.

This case undermines the common idea that abstraction necessarily makes a problem harder or less usable. Here:

\[ \text{greater generality} \quad\Rightarrow\quad \text{simpler proof} + \text{broader recognizability}. \]

The special cases contained too much irrelevant detail. Generalization removed the details and exposed the monotonicity mechanism that all the applications shared.

5. Mendelian genetics and biometry were treated as rivals until a more general model reconciled them

Around the beginning of the twentieth century, Mendelian researchers emphasized discrete hereditary factors and apparently discontinuous traits, while biometricians studied continuously varying traits using correlation, regression, and population statistics. The two approaches were frequently presented as conflicting accounts of heredity — but the opposition was mathematically unnecessary. A quantitative trait influenced by many Mendelian factors of individually small effect, together with environmental variation, can appear approximately continuous. Yule had already discussed compatibility early in the controversy, but the synthesis was not fully absorbed; Fisher's 1918 work gave a much more comprehensive account relating Mendelian inheritance to correlations among relatives and the statistical decomposition of variation.

The missing generalization was from:

\[ \text{one gene producing a conspicuous categorical difference} \]

to:

\[ \text{many discrete inherited factors jointly producing continuous variation}. \]

That generalization showed that the two fields had often been observing different scales of the same system. This case should be presented with care — the historical divide involved experimental evidence, personalities, institutions, and competing scientific programs, not mathematics alone, and modern historians regard the tidy “Mendelians versus biometricians” story as somewhat stylized. Even so, it remains an important example of a false interdisciplinary conflict sustained by failure to recognize that a broader model contained both local descriptions.

6. Iterative proportional fitting was rediscovered as Sinkhorn scaling, RAS, raking, and other methods

Consider the elementary-looking iteration that alternately rescales the rows and columns of a nonnegative matrix to obtain prescribed marginals. Versions of this procedure appeared under several names: iterative proportional fitting in statistics; raking in survey estimation; RAS in economics and input–output analysis; biproportional fitting; Sinkhorn or Sinkhorn–Knopp scaling in numerical analysis; alternating projection in information geometry; Fortet iterations and related procedures in transport and Schrödinger-bridge theory. Historical and technical accounts explicitly describe repeated independent rediscovery. The common abstraction is matrix scaling, or more generally alternating information projections minimizing a Kullback–Leibler or Bregman divergence subject to marginal constraints.

Once expressed this way, convergence theory, computational techniques, and modeling insights become transferable among contingency-table estimation, economics, optimal transport, network models, choice models, and machine learning. Recent work connecting choice models to matrix balancing illustrates the continuing dividend: familiar choice-estimation algorithms can be recognized as Sinkhorn procedures, and existing matrix-balancing theory can settle convergence questions that appear specialized in the choice-modeling vocabulary. This is an excellent example of search failure caused by naming — a statistician looking for “raking,” an economist looking for “RAS,” and an optimization researcher looking for “Bregman projections” could be studying the same iteration without finding one another's literature.

7. Wavelets, quadrature-mirror filter banks, and image pyramids developed in partially separate languages

Applied mathematicians studied wavelet bases and multiresolution analysis. Digital-signal-processing researchers studied two-channel perfect-reconstruction filter banks and quadrature-mirror filters. Computer-vision researchers studied multiscale image pyramids. The mathematical overlap was substantial: compactly supported orthonormal wavelets corresponded to particular perfect-reconstruction finite-impulse-response filter banks, and the coefficients used in wavelet constructions were already familiar to signal-processing researchers as quadrature-mirror-filter sequences. Mallat's multiresolution framework and pyramidal algorithm provided a particularly effective bridge among wavelet theory, filter-bank implementation, and computer-vision scale spaces.

The history here is more a case of partial duplication and delayed synthesis than total isolation. Nevertheless, the abstraction mattered. Multiresolution analysis separated three questions that had been entangled in local terminology:

\[ \begin{aligned} &\text{What nested approximation spaces exist?}\\ &\text{What basis functions represent their differences?}\\ &\text{What filters implement the transformation efficiently?} \end{aligned} \]

Once those were connected, regularity theory from analysis, perfect-reconstruction conditions from engineering, and fast pyramidal computations from image processing could reinforce one another.

8. Kriging, Gaussian-process regression, splines, and kernel machines remain divided by vocabulary

Geostatistics speaks of kriging, variograms, spatial covariance, and nuggets. Machine learning speaks of Gaussian-process regression, kernels, and observation noise. Statistics speaks of smoothing splines and penalized function estimation. Learning theory speaks of reproducing-kernel Hilbert spaces and support-vector machines. In many standard settings, kriging and Gaussian-process regression produce the same predictor from closely corresponding assumptions, yet comparative work explicitly reports that differences in terminology inhibit knowledge transfer and can conceal the underlying equivalence. “Kernel,” “covariance function,” and the variogram-related formulation are often alternative coordinates for closely connected objects, although the fields may estimate parameters differently or interpret the model differently.

A related literature shows how placing penalized splines inside the reproducing-kernel Hilbert-space framework connects classical statistical smoothing to kernel machines — a unification that transfers ideas about interpretation, structured effects, computation, model diagnosis, and regularization between statistics and machine learning, and can produce substantial computational simplifications. This is a useful contemporary example because the barrier hasn't completely disappeared: the mathematics may already be known, yet discipline-specific semantics still determine which papers researchers encounter and which methods they regard as relevant.

The appropriate caution is that these methods are not universally identical — different parameter-estimation procedures, prior interpretations, uncertainty treatments, and asymptotic traditions matter. The point is that a common covariance-kernel/RKHS abstraction tells us precisely which parts are the same and which parts are not.

9. Polynomial spectral factorization was rediscovered across operator theory, control, signal processing, and algebra

Closely related factorization results have appeared as: the matrix Fejér–Riesz theorem; Wiener–Hopf factorization; spectral factorization; polynomial matrix sum-of-squares factorization.

A modern synthesis notes that polynomial spectral factorization was rediscovered repeatedly and that symbolic and numerical approaches developed largely separately in different fields for decades. Connecting those literatures made it possible to combine algebraic existence theory, numerical algorithms, and computational-complexity analysis, including stronger bit-complexity results. This is a more specialized example, but it is unusually direct: the lost information was not just conceptual inspiration — symbolic mathematicians and numerical analysts possessed complementary parts of the solution while field-specific formulations prevented those parts from being assembled efficiently.

What these examples show

The central information-theoretic advantage of abstraction is not merely compression. It is the creation of a shared address space for knowledge. Suppose \(k\) disciplines possess local formulations \(F_1,F_2,\ldots,F_k\) of the same underlying structure \(S\). Without \(S\), useful transfer may require a separate translation between every pair of fields, \(F_i\longleftrightarrow F_j\). There are

\[ \binom{k}{2}=\frac{k(k-1)}2 \]

possible pairwise interfaces. In practice, most are never built. A common abstraction changes the architecture:

\[ F_i\longrightarrow S\longleftarrow F_j. \]

Each field needs only to establish how its objects realize \(S\). The translation burden becomes approximately linear rather than quadratic:

\[ O(k^2)\quad\longrightarrow\quad O(k). \]

That is a stylized coordination model rather than an empirical law, but it captures what happened in the examples above. Matroids became the hub for several rank-like invariants; factor graphs became the hub for message-passing algorithms; computational graphs became the hub for adjoint differentiation; latent-variable likelihood became the hub for EM; and matrix scaling became the hub for IPF, RAS, raking, and Sinkhorn iteration.

Five mechanisms of loss

These histories reveal five distinct mechanisms:

The strongest formulation

Not that abstraction merely makes knowledge more elegant or more future-proof. The stronger defensible claim: without a common abstraction, structurally identical results are stored under incompatible field-specific coordinates. They become difficult to search for, difficult to recognize, and costly to transfer. Each community may repeat the same discovery while remaining unaware that another community already possesses its proofs, algorithms, counterexamples, or complexity bounds. Rigorous abstraction constructs a canonical interface through which one result can propagate to every field that realizes the same structure.

Look, the case is bloody obvious

It's more efficient for those with skill and momentum to build the machinery before it is needed than to wait for the right person to need exactly the right result at exactly the right time. Abstraction is the extraction of invariants; generalization is the weakening of assumptions; rigor is what makes conclusions transportable. A concrete solution settles one problem; a rigorous abstract theorem settles every problem, known or unknown, that realizes the same structure. Its future usefulness isn't an accidental exception to its abstraction — it's a consequence of abstraction's ability to separate reusable structure from the circumstances in which that structure was first noticed. None of this denies that applied work supplies plenty of its own impetus. It just means waiting for it is the inefficient policy, not the safe one.

The stakes are not abstract

In 2021 the University of Leicester moved to make all eight of its permanent pure mathematics staff redundant, offering to replace them with three teaching-only positions; ten members of the School of Informatics working on theoretical and foundational topics faced the same threat. Administrators justified it by citing demand for AI and data science — the exact fields the Ledger keeps finding downstream of the mathematics being cut. The London Mathematical Society condemned the decision. Timothy Gowers campaigned publicly against it from January of that year. Alison Parker, an external examiner at the University of Leeds, resigned her position over it and donated her examiner's fee to the resulting legal fund. A petition, “Mathematics is not redundant,” drew 8,500 signatures. None of it changed the outcome: three staff were dismissed outright, three were moved into teaching-only roles, and the rest took early retirement or severance. The University and College Union called for an international academic boycott. (This account is drawn from a detailed, widely circulated contemporaneous record naming the people and organizations above, not checked here against primary university or LMS documents.)

I bring this up because the myth isn't only a historical curiosity about mathematicians being wrong in print. It has a live version, and the live version costs people their jobs. Nobody at Leicester argued, the way Hardy did, that a specific branch of mathematics would never see practical use. They argued that pure mathematics had no near-term institutional value next to AI — the same AI that runs on linear algebra, optimization, and probability, built on results that were once dismissed exactly the same way.

The Ledger runs on named dates and cited applications on purpose, aimed at readers who aren't naturally comfortable reasoning in the abstract. In my experience, the readers who are comfortable reasoning in the abstract are not likely to call for funding cuts to mathematics departments. The dates and citations are for everyone else.

The name is doing the opposite of its job

“Pure” sounds like an indulgence — a luxury version of mathematics, done for its own sake, maybe at the expense of something more useful. But “pure” is close to a synonym for efficient, or non-duplicative. Call the field Efficient Mathematics, or Non-Duplicative Thinking, and nobody questions the funding.

Meanwhile every branch of science keeps reinventing the same underlying structure and renaming it, field by field, oblivious to the fact that someone settled the general case decades or centuries earlier. A tidy example: the probability integral transform — push data through its own cumulative distribution function and it becomes uniform; pull a uniform back through an inverse CDF and it becomes anything you like — is one simple idea from classical statistics. It has been independently rediscovered and renamed as the chi-squared and Kolmogorov–Smirnov tests, inverse-transform sampling and Monte Carlo simulation, copulas, the time-rescaling theorem for point processes, histogram equalization in image processing, rank histograms and Talagrand diagrams in weather forecasting, normalizing flows in deep learning, and conformal prediction — most without citing the others. Full history: skaters.microprediction.org/heritage.html.

Abstraction is what avoids paying that cost twice. It maps the world the way a zip code maps a location: a short, general-purpose address that lets you skip re-deriving the whole route every time you need to refer to the same place.

None of this is an objection to renaming as such. An economist calling a Lagrange multiplier a shadow price has added real interpretive content. The failure mode is the rename that happens because nobody realized there was already a name.

A personal postscript

The nine cases above are cited literature, checked against named papers. What follows isn't — it's seven instances from my own work where the same pattern showed up, offered as testimony, not evidence.

Despite starting out in pure mathematics, and despite a career spent entirely in industry since — a fairly narrow industry, at that — I was guilty of exactly this. More than once I looked at some construct in a textbook and filed it as a useless toy. I was wrong every single time.

The most recent was Stokes' theorem. I did not expect it to matter to a paper I wrote this year, and it settled the argument the moment it showed up — a 19th-century result about boundaries and interiors, called on by name in a problem that had nothing to do with either until it did.

Information theory looked, for years, like a narrow specialty about channel capacity and compression — until the duality with growth-optimal betting became impossible to unsee. The bits you save compressing a source are the same bits you'd have won betting on it. Shannon and Kelly are the same theorem, wearing different clothes.

Savage's 1971 characterization of proper scoring rules reads, on first pass, like taxonomy filed in a basement: a catalog of which loss functions elicit honest probabilities and why. It turns out to be an engineering guide. Anything you want an autonomous agent to report truthfully about its own beliefs needs a proper scoring rule underneath it, and Savage worked out the admissible family decades before anyone had a reason to build one.

I once found a matrix identity in a textbook with no name attached to it at all — not a named theorem, not a named lemma, just a manipulation sitting in an exercise. It was the Schur complement, and it's the reason hierarchical risk parity and minimum-variance portfolio construction turn out to be two ends of one dial rather than two competing schools, once you augment each side's covariance block with the other's.

Differential geometry looked, from a distance, like a subject invented to chart curved terrain near black holes. The space of covariance matrices is itself curved, and treating a shrinkage estimate as a move along a geodesic on that curved space — rather than a straight-line average in flat coordinates — is the more natural operation, not a borrowed metaphor from general relativity.

And de Finetti's representation theorem — an exchangeable sequence is a mixture of independent, identically distributed ones — is a statement about coin flips that turns out to be the same idea behind pricing a CDO tranche (defaults are exchangeable conditional on a common factor) and behind why an epidemic can turn over well before the textbook herd-immunity threshold (the population is a mixture of transmission rates, not one number). The same mixture, twice, in fields that don't read each other's journals.

The earliest instance predates all of these. I reinvented iterative proportional fitting as a kid, staring at trifecta pools at the racetrack, with no idea it had a name — let alone that statisticians call it raking and economists call it RAS. I just wanted to know how to rescale a betting pool across outcomes so the payouts balanced, and the row-and-column iteration fell out on its own. That's case 6 above, decades before I knew it was case 6 above.

Continue to