VOYAGE TWENTY · ASKING ABOUT OBJECTIVES AND CONSTRAINTS BEFORE OPTIMA

Who Chooses the Best Answer? — From Shortest Paths to AI

Begin with reflected paths in Alexandria and the fastest curve in Groningen, then minimize functions and errors in Lausanne and Paris. Compute constraints and plans in Chicago, Leningrad, Washington, Berkeley, and Santa Monica before reaching interior points in Murray Hill and learning in Toronto—separating ‘the calculation is correct’ from ‘the objective is right.’

QUESTION FOR THE ROUTE

How does optimization turn ‘what is feasible?’ and ‘what is better?’ into mathematics, and who must question the people, costs, and values omitted from that mathematics?

WHAT THE LINE DOES NOT CLAIM

This route does not equate optimization with universal efficiency, automated justice, or one engine of AI. It separates shortest distance from shortest time, local descent from global optimality, necessary from sufficient conditions, worst-case theory from practical performance, and Pareto efficiency from justice. The map line is an edited comparison across distinct problems and institutions, not one text’s proven transmission.

The camera rests on each city while you read, then eases through the runway between scenes. Select any marker or scene link to travel in either direction.

The same route, four questions

A lens never hides a scene or proves a cause. It changes which places you compare first, and the URL preserves your choice.

Whole route Translation networks Commercial demand Observation and institutions Patronage and power

Translation networks · READING QUESTION

What was lost, preserved, or newly created when an idea met another language and audience?

A shortest geometric path was translated into variation over functions, losses over observational error, inequality constraints, sequential decisions, and stochastic learning. Under the same word ‘minimum,’ the objects, assumptions, and proofs kept changing.

Start with:Alexandria Groningen Lausanne Paris Chapel Hill

Translation, copying, or commentary does not prove that one document traveled directly through the whole route.

Live map · 지도를 불러오는 중…

01 / 15 · c. 60 CE

Alexandria

01 · c. 60 CE
Alexandria · Composition
Seeing a Shortest Route in a Broken Path of Light — the Catoptrica
Translation networks · lens spotlight
Unfold a reflected path at a plane mirror and the equal-angle route becomes the shortest broken line. This geometric argument is an important shoulder for later variational principles. Yet the work survives only in a medieval Latin translation, was misattributed to Ptolemy, and is only doubtfully assigned to Heron. It does not say that light minimizes geometric distance in every medium and setting.
PAUSE AND ASK
Why is the equal-angle reflected path the shortest as the reflection point moves?
How the idea changed
Move from measuring one path to comparing a family of possible paths and selecting an extremum.
What this place made possible
The optical and geometric tradition associated with Alexandria organized mirrors, vision, and reflection as diagrammatic arguments, but the surviving Latin translation cannot fix an exact writing room.
How it moved
Ancient reflection geometry → medieval Latin translation and misattribution → qualified reassignment to Heron → later comparison with Fermat and variational principles
Do not overclaim
The Catoptrica is usually but doubtfully attributed to Heron. Its plane-mirror geometry is not projected backward into a universal claim that light minimizes distance in every medium or into a modern optimization algorithm.
Evidence sources
- Oxford Classical Dictionary — Heron of Alexandria
  Supports: The medieval Latin translation, misattribution to Ptolemy, and uncertainty of attribution to Heron
Stable link to this scene
AlexandriaGroningen
02 · 1696 CE
Groningen · Publication
Asking for a Curve Faster Than the Shortest Line — Bernoulli’s Challenge
Translation networks · lens spotlight
Johann Bernoulli challenged mathematicians to find the quickest descent under gravity between two points. The cycloid is longer than the straight segment but drops steeply first and gains speed earlier. The 1696 Acta Eruditorum challenge and a 1697 Groningen reprint moved the problem through European correspondence; the result assumes an ideal point mass, no friction, and uniform gravity.
PAUSE AND ASK
Why is the fastest descent between two points not the shortest straight line?
How the idea changed
See the power of the objective: replace distance with time and the optimal curve changes.
What this place made possible
While teaching in Groningen, Bernoulli used a Leipzig journal and a separate reprint to send the challenge across Europe and create a public arena for comparing solutions.
How it moved
Galilean descent problems → Bernoulli’s 1696 journal challenge and 1697 Groningen reprint → solutions by Newton, Leibniz, Jakob Bernoulli, and others → calculus of variations
Do not overclaim
The cycloid is fastest under ideal assumptions such as a frictionless point mass and uniform gravity. The problem was neither Johann’s isolated invention nor a set of mathematically identical solutions.
Evidence sources
- University of Groningen — The Brachistochrone Problem
  Supports: The 1696 Acta Eruditorum problem and its 1697 separate Groningen printing
Stable link to this scene
GroningenLausanne
03 · 1744 CE
Lausanne · Publication
Choosing an Entire Function Instead of a Number — Euler’s Calculus of Variations
Translation networks · lens spotlight
Euler’s Methodus inveniendi systematically studied curves that maximize or minimize quantities such as length, time, or action. The unknown was no longer just a finite list of numbers but an entire function. Lausanne and Geneva mark the publication network rather than Euler’s main workplace, and one book did not single-handedly invent every part of the calculus of variations.
PAUSE AND ASK
How can one choose the best object from an entire family of functions rather than a few numbers?
How the idea changed
Expand the unknown from numbers to curves and functions, then extremize a total quantity attached to each.
What this place made possible
The Bousquet publishing network in Lausanne and Geneva turned Euler’s long work across St Petersburg and Berlin into a systematic Latin book for an international readership.
How it moved
Brachistochrone solutions → Euler’s generalization to extrema of functions → 1744 Methodus inveniendi → Lagrange’s variational notation → Euler–Lagrange equation
Do not overclaim
The pin marks publication, not Euler’s sole workplace. The book is not compressed into one moment that invented every variational idea or completed the physical principle of least action.
Evidence sources
- Euler Archive — Methodus inveniendi lineas curvas (E065)
  Supports: The 1744 Lausanne and Geneva publication and its broad early systematization of the calculus of variations
Stable link to this scene
LausanneParis
04 · 1805 CE
Paris · Publication
Squaring Errors to Choose the Best-Fitting Orbit — Legendre
In a work on comet orbits, Legendre first published least squares: minimize the sum of squared differences between observations and a model. Many residuals became one objective for selecting parameters. Gauss later claimed earlier private use and published a theory in 1809, so first publication and claimed prior practice remain distinct. Squared error is not automatically the right loss for every dataset.
PAUSE AND ASK
How can ‘best fit’ among conflicting observations be defined as one number?
How the idea changed
Replace separate explanations of each error with one sum-of-squared-residuals objective for choosing model parameters.
What this place made possible
Parisian astronomy, geodesy, printing, and academy networks made a procedure for reconciling observations public and comparable inside a work on comet orbits.
How it moved
Overdetermined observations in astronomy and geodesy → Legendre’s 1805 publication → Gauss’s 1809 theory and claim of prior use → error models, regression, and statistical estimation
Do not overclaim
Legendre’s first publication is separated from Gauss’s claim of earlier private use. Squared loss is sensitive to large errors and is not a natural or neutral criterion for every dataset and purpose.
Evidence sources
- ETH Zürich e-rara — Nouvelles méthodes pour la détermination des orbites des comètes
  Supports: The 1805 Paris edition and its comet-orbit context
Stable link to this scene
ParisParis
05 · 1847 CE
Paris · Presentation
Walking in the Steepest Local Downhill Direction — Cauchy
Translation networks · lens spotlight
Cauchy proposed following the negative gradient while solving simultaneous equations, an ancestor of gradient descent. Direction alone is not enough: a tiny step can be slow, a large one can oscillate or diverge, and a nonconvex landscape can contain local minima, saddle points, and flat regions.
PAUSE AND ASK
Once a downhill direction is known, how far should one move?
How the idea changed
Replace one closed-form solution with repeated local gradient calculations approaching an approximation.
What this place made possible
The Paris Academy’s short proceedings format quickly recorded a numerical procedure for simultaneous equations and circulated it among analysts and calculators.
How it moved
Differentiation and extrema → iterative solution of simultaneous equations → Cauchy’s 1847 negative-gradient procedure → steepest descent and line search → modern optimization and learning
Do not overclaim
The negative gradient is a local descent direction, not a guarantee of a global optimum for an arbitrary nonconvex function. Step size, initialization, stopping, and numerical error change the result.
Evidence sources
- Lawrence Berkeley National Laboratory — A Brief History of Optimization
  Supports: Cauchy’s 1847 steepest-descent proposal and the historical and numerical role of step choice
Stable link to this scene
ParisLausanne
06 · 1896 CE
Lausanne · Teaching / position
Finding the Boundary Where Helping One Person Hurts Another — Pareto
In political-economy teaching and books produced at Lausanne, Pareto analyzed allocations involving several preferences. A state later called Pareto efficient cannot improve one person without worsening another. That property guarantees neither fairness, justice, equality, nor a unique solution. The chosen preferences, resources, and measures come first.
PAUSE AND ASK
Is a boundary where one improvement requires another loss automatically a fair answer?
How the idea changed
Replace one objective value with a frontier of trade-offs among several agents and multiple nondominated solutions.
What this place made possible
Political-economy teaching and publication at Lausanne linked mathematical choice and allocation to institutions, without turning the period’s economic and class assumptions into universal values.
How it moved
Marginal utility and general equilibrium → Pareto’s 1896–1897 course → Pareto criteria in welfare economics → multiobjective optimization and Pareto frontiers
Do not overclaim
Pareto efficiency means no further unanimous improvement under given preferences and resources. It guarantees neither justice, equality, nor a unique social optimum; a highly unequal allocation can be efficient.
Evidence sources
- WorldCat — Cours d’économie politique professé à l’Université de Lausanne
  Supports: The bibliographic and institutional context of the 1896–1897 Lausanne course
Stable link to this scene
LausanneChicago
07 · 1939 CE
Chicago · Composition
Writing Conditions for Minima under Inequality Constraints — Karush
William Karush’s University of Chicago master’s thesis treated multiplier conditions for optimization with inequality constraints. It remained little known until its precedence was recovered after Kuhn and Tucker’s 1950 work, producing the name KKT. The conditions do not automatically guarantee a global optimum: necessity needs regularity assumptions, and sufficiency needs additional structure such as convexity.
PAUSE AND ASK
At an inequality boundary, which conditions identify candidates for a minimum?
How the idea changed
Replace the zero-gradient test with a joint calculation of active constraints, multipliers, and complementarity.
What this place made possible
The University of Chicago master’s-thesis system recorded the result, while limited circulation and later wartime and disciplinary paths kept it outside the standard story for years.
How it moved
Lagrange multipliers → extension to inequalities → Karush’s 1939 master’s thesis → Kuhn and Tucker’s 1950 presentation → recovery of precedence and the name KKT
Do not overclaim
KKT is not an unconditional solver. Necessity needs regularity such as a constraint qualification, global sufficiency needs further assumptions such as convexity, and a nonconvex stationary point need not be optimal.
Evidence sources
- University of Chicago Library — Minima of functions of several variables with inequalities as side conditions
  Supports: The bibliographic record and title of Karush’s 1939 University of Chicago master’s thesis
Stable link to this scene
ChicagoSaint Petersburg
08 · 1939 CE
Saint Petersburg · Publication
Giving Scarce Factory Resources Shadow Values — Kantorovich
In Leningrad, Kantorovich published a booklet using linear inequalities and solution methods to plan production with limited machines, labor, and materials. Multipliers on scarce resources later supported the interpretation of shadow prices. The work remained largely unknown in the West for years, so it is not merged into one direct transmission line with later American linear programming.
PAUSE AND ASK
How should scarce machines, labor, and materials be allocated among production plans?
How the idea changed
Translate factory rules of thumb into a planning problem with a linear objective, inequalities, and marginal values for scarce resources.
What this place made possible
Leningrad University, industrial consulting, and Soviet planning made production allocation urgent and publishable, while linguistic and institutional barriers delayed international movement.
How it moved
Plywood-factory planning → Kantorovich’s 1939 booklet → limited reception across war and institutions → 1950s republication and economic interpretation → linear programming and shadow prices
Do not overclaim
Kantorovich and Dantzig reached core structures independently in different institutions. The Soviet work is not asserted to have directly produced the 1947 American simplex method, and the planning model did not capture every real-world value.
Evidence sources
- Nobel Prize — Leonid Vitaliyevich Kantorovich, Biographical
  Supports: The 1939 Leningrad University Press booklet and its delayed recognition in the West
Stable link to this scene
Saint PetersburgWashington DC
09 · 1947 CE
Washington DC · Main activity
Turning Military Plans into Moves between Vertices — Dantzig
While working on US Air Force planning, George Dantzig developed a linear-planning model and the simplex method, which moves among vertices of a feasible polytope. ‘Programming’ meant organizing plans, not primarily writing computer code. The method is powerful but not neutral with respect to its military and administrative objectives, and it cannot value harms or benefits omitted from the objective.
PAUSE AND ASK
When there are too many feasible plans, how can a good vertex be found without checking every point?
How the idea changed
Read a planning table as a high-dimensional polytope and improve the objective by moving among adjacent vertices.
What this place made possible
Postwar US Air Force planning organizations in Washington assembled deployment, training, logistics problems, and computing labor to test a general planning model.
How it moved
Wartime and postwar logistics planning → Dantzig’s 1947 linear model and simplex method → diet problem and early-machine tests → industrial, transport, and communication optimization
Do not overclaim
Programming here means planning rather than coding. Simplex is often powerful in practice, while some pivot rules have exponential worst cases, and neither its military objective nor omitted social costs are neutral.
Evidence sources
- INFORMS — Oral History of George Dantzig
  Supports: The 1947 planning model and simplex method, Pentagon work, and the period meaning of programming
Stable link to this scene
Washington DCBerkeley
10 · 1950 CE
Berkeley · Presentation
Publishing Stationarity Conditions for Constrained Nonlinear Problems — Kuhn and Tucker
At the Berkeley symposium on nonlinear programming, Kuhn and Tucker combined objective and constraint gradients, multipliers, and complementary slackness. Published in the 1951 proceedings, their work is read with Karush’s 1939 predecessor as the KKT conditions. Satisfying them does not prove global optimality for an arbitrary nonconvex problem, and constraint qualifications matter.
PAUSE AND ASK
How can one equation express the balance between objective and constraint gradients at a boundary?
How the idea changed
Translate feasible-region geometry into multipliers, active constraints, and complementarity, creating a common language for nonlinear programming.
What this place made possible
The 1950 Berkeley symposium gathered distinct optimization results in one venue and proceedings, giving nonlinear programming a name and an inspectable research community.
How it moved
Lagrange multipliers and Karush’s thesis → 1950 Berkeley presentation → 1951 symposium proceedings → extensions through constraint qualifications, duality, and convex optimization
Do not overclaim
Kuhn and Tucker’s public influence is recognized without erasing Karush’s predecessor. Satisfying KKT and being the unique global optimum of an arbitrary problem are different claims.
Evidence sources
- UC Berkeley Digital Collections — Nonlinear Programming
  Supports: The 1950 Berkeley symposium and bibliography of the Kuhn–Tucker paper published in 1951
Stable link to this scene
BerkeleyChapel Hill
11 · 1951 CE
Chapel Hill · Publication
Reducing Step Size while Estimating through Noise — Robbins and Monro
Translation networks · lens spotlight
Robbins and Monro published a stochastic-approximation process for locating a root from noisy observations of an otherwise inaccessible function. Properly diminishing steps can average random disturbance while continuing toward the target. It is an important ancestor of stochastic gradient methods, not the identical mini-batch neural-network algorithm used today.
PAUSE AND ASK
Can a target be found when exact function values are unavailable and only noisy observations remain?
How the idea changed
Replace the demand for a complete function table with repeated observations and diminishing steps that average uncertainty.
What this place made possible
The mathematical-statistics community at the University of North Carolina linked probability with iterative computation and supported publication in the Annals of Mathematical Statistics.
How it moved
Statistical estimation and sequential experiments → Robbins–Monro stochastic approximation in 1951 → stochastic root finding and control → stochastic gradient and online learning
Do not overclaim
The original problem is noisy root finding, not identical to modern mini-batch SGD. Convergence needs conditions on steps, noise, and the function and does not guarantee success after any finite run.
Evidence sources
- Project Euclid — A Stochastic Approximation Method
  Supports: The original 1951 Robbins–Monro stochastic-approximation paper and author affiliations
Stable link to this scene
Chapel HillSanta Monica
12 · 1953 CE
Santa Monica · Publication
Folding a Long Decision into the Values of Remaining Problems — Bellman
At RAND, Bellman organized multistage decisions through dynamic programming: after a current choice, the remaining tail must itself be optimal from the new state. ‘Programming’ again meant planning, and Cold War military systems analysis supplied institutional support. As the number of states grows, the curse of dimensionality can overwhelm the recursion.
PAUSE AND ASK
Can a long decision be folded into values of remaining subproblems instead of being solved afresh each time?
How the idea changed
Replace comparison of complete paths with reusable optimal values for each state, computed backward.
What this place made possible
RAND in Santa Monica supplied sustained salaries, military systems problems, and computing resources for turning sequential-decision theory into reports, books, and computation.
How it moved
Variational and control problems → multistage decisions at RAND → first dynamic-programming report in 1953 → principle of optimality → control, economics, and reinforcement learning
Do not overclaim
Programming again means planning. The state must model all information relevant to the future, and exact computation can become infeasible under the curse of dimensionality.
Evidence sources
- INFORMS — RAND Corporation
  Supports: Bellman’s first RAND dynamic-programming report in 1953 and its military-systems research environment
Stable link to this scene
Santa MonicaBerkeley
13 · 1972 CE
Berkeley · Publication
Grouping Problems Whose Optima May Be Hard to Find Quickly — Karp
Richard Karp connected many combinatorial problems by polynomial-time reductions, greatly expanding the family known to be NP-complete. This does not mean no optimum exists or every input is impossible to solve. It is a worst-case, conditional hardness claim while P versus NP remains open; approximation, heuristics, special structure, and small instances still matter.
PAUSE AND ASK
Can an exact optimum exist while the time needed to find it becomes unmanageable as input grows?
How the idea changed
Separate existence and verification from fast search, linking distinct problems by reductions to compare worst-case difficulty.
What this place made possible
Berkeley’s theoretical-computer-science community and expanding conference and journal networks made computational complexity a shared question in algorithm design.
How it moved
Turing computability → Cook’s 1971 NP-completeness → Karp’s 1972 reductions among twenty-one combinatorial problems → approximation, parameterization, and heuristics
Do not overclaim
NP-complete does not mean absolutely unsolvable. P versus NP remains open, the claim concerns worst cases and input growth, and special structure, approximation, and heuristics can work well on large real instances.
Evidence sources
- Simons Institute, UC Berkeley — Richard Karp
  Supports: The 1972 paper’s role in connecting many combinatorial problems as NP-complete and likely intractable
Stable link to this scene
BerkeleyMurray Hill
14 · 1984 CE
Murray Hill · Publication
Crossing the Interior Instead of Following the Polytope’s Edges — Karmarkar
At Bell Labs, Karmarkar published a polynomial-time linear-programming method that traveled through the feasible region’s interior. It opened a major era for interior-point methods, but Khachiyan’s 1979 ellipsoid method was the first polynomial-time LP algorithm. Worst-case theory and practical speed are different comparisons, and simplex remains valuable on many problems.
PAUSE AND ASK
Can one reach an optimum quickly by crossing a polytope’s interior instead of following its vertices?
How the idea changed
Use projective and barrier geometry distinct from simplex boundary moves to build a polynomial-time path for linear programming.
What this place made possible
Bell Labs in Murray Hill combined long-term basic research, telecommunications optimization, and large machines, allowing a new algorithm to be tested in theory and implementation.
How it moved
Simplex’s practical performance and worst cases → Khachiyan’s 1979 ellipsoid method → Karmarkar’s 1984 interior-point method → barrier and conic optimization
Do not overclaim
Karmarkar’s method was not the first polynomial-time LP algorithm. Theoretical complexity and practical performance differ, and interior-point methods did not replace simplex on every problem.
Evidence sources
- Springer — A new polynomial-time algorithm for linear programming
  Supports: Karmarkar’s original 1984 paper, its polynomial-time claim, and Bell Labs affiliation
Stable link to this scene
Murray HillToronto
15 · 2012 CE
Toronto · Publication
Learning Features by Lowering Loss across Millions of Examples — AlexNet
Krizhevsky, Sutskever, and Hinton combined GPUs, large-scale ImageNet data, convolutional networks, backpropagation, and stochastic optimization to sharply reduce image-classification error. AlexNet did not single-handedly invent gradient descent, backpropagation, or deep learning. Minimizing benchmark loss measures performance on chosen data and metrics, not general intelligence, fairness, or social value.
PAUSE AND ASK
Does lowering loss on millions of examples mean a system understands the world best?
How the idea changed
Move from hand-designed features to a computational regime where large data, GPUs, and backpropagation jointly learn representations and parameters.
What this place made possible
Long-running neural-network research at Toronto, public ImageNet data, NVIDIA GPUs, and an international conference created a stage for testing a combination of algorithm, data, and hardware.
How it moved
Earlier neural networks and backpropagation → ImageNet data and GPU parallelism → AlexNet in 2012 → expansion of large-scale deep learning → questions of objective mismatch, bias, and energy
Do not overclaim
AlexNet did not invent gradient descent, backpropagation, or deep learning; it combined three authors’ work with multiple infrastructures. Minimizing benchmark error is not the same objective as general intelligence, fairness, truth, or social good.
Evidence sources
- NeurIPS — ImageNet Classification with Deep Convolutional Neural Networks
  Supports: The original AlexNet paper’s authors, affiliations, data scale, GPUs, and classification performance
Stable link to this scene

FOUR QUESTIONS BEFORE ‘THE BEST’

Shortest, lowest, feasible—and best for whom?

Each control runs a small model in the browser and remains in the shared URL. The calculation can be exact inside the model while the model’s objective and omissions remain open to criticism.

THE OBJECTIVE CHANGES THE WINNER

The shortest slide is not the fastest slide

A straight line minimizes geometric length. Under uniform gravity, a cycloid drops steeply first and reaches the same endpoint sooner. Optimization begins only after we say what ‘better’ means.

selected winner

straight

time saved by cycloid

8.7%

Straightlength 1.414 · time 0.639

Cycloidlength 2.141 · time 0.583

Normalized simulation: unit horizontal and vertical drop, uniform gravity, no friction, point mass released from rest.

An optimum is conditional: objective + constraints + data + algorithm + stopping rule. Changing any one can change the answer.

TOUCH THE MATHEMATICS