John von Neumann

John von Neumann

Mathematician, 1903–1957. Study focus: stored instructions, strategic randomization, and the difference between a model and a machine.

Von Neumann contributed to mathematical logic, quantum theory, game theory, numerical computation, and automata. For computer science, his importance includes the widely circulated 1945 EDVAC report and the collaborative computer project at the Institute for Advanced Study. Naming an architecture after him does not establish that he alone invented stored programs.

Historical setting

Born in Budapest, von Neumann studied mathematics and chemical engineering, completed his mathematics doctorate in 1926, and worked in Göttingen before moving to the United States. His 1928 minimax work and the 1944 book with Oskar Morgenstern helped establish mathematical game theory. His 1932 book developed a rigorous framework for quantum mechanics. These were distinct contributions, not applications of one universal recipe that made every subject tractable. MacTutor biography

Von Neumann joined the Institute for Advanced Study in 1933. During the war he worked on military research, including nuclear-weapons problems, and became involved with the Moore School's electronic-computing work. The EDVAC design emerged from discussions involving several people, including J. Presper Eckert, John Mauchly, Herman Goldstine, and Arthur Burks. The surviving report is attributed to von Neumann; attribution of the report and priority for every idea in it are different questions.

The 1945 First Draft describes central arithmetic, central control, memory, input, and output. The later IAS report inventory separately credits Burks, Goldstine, and von Neumann for logical design, and Bigelow, Pomerene, Slutz, and Ware for a physical implementation report. This is direct evidence that designing and building the system was collaborative. EDVAC report, IAS report records

What storing a program changes

An instruction can be encoded as a finite pattern, placed in addressable memory, fetched by a control mechanism, and interpreted as an operation. Instructions can therefore be loaded and transformed by other programs. A compiler's output can become another program's input without rewiring the machine.

Stored program and shared instruction/data memory are related but distinct properties. A machine can store instructions while providing separate instruction and data memories. Nor does finite physical hardware literally have unbounded Turing-machine storage. Universality is assessed using an idealized model with expandable resources, sufficient instructions, and an appropriate control structure.

Here is an original teaching machine, not the EDVAC instruction set. It has an accumulator A, a program counter PC, and integer-addressed memory. LOAD a copies M[a] into A; ADD a adds it; STORE a writes A; and HALT stops. After each non-halting instruction, PC advances by one.

Address  Content
0        LOAD 10
1        ADD 11
2        STORE 12
3        HALT
10       7
11       5
12       0

Starting with PC=0, the state transitions are (PC,A)=(1,7), then (2,12), then (3,12) with M[12]=12, then halted. Loading a different instruction sequence at addresses 0–3 changes behavior without replacing the arithmetic circuitry. Changing only M[10] changes the input, not the algorithm.

Shared storage also permits accidental overwriting of instructions. If the third instruction were STORE 3, the value 12 would replace the next instruction. What happens next depends on the machine's encoding: 12 might be valid code, invalid code, or something dangerous. Modern memory protection and executable-permission rules constrain such behavior. The logical possibility of programs as data is not an instruction to make every memory page writable and executable.

Memory bandwidth as a resource

The EDVAC report's decomposition remains useful even though modern processors include caches, pipelines, multiple execution units, and parallel cores. An architecture diagram is an abstraction, not a literal picture of every smartphone or microcontroller. Original design and its subdivisions

Consider a constructed workload that reads two 8-byte values and writes one 8-byte result for each addition. Assume no reuse, no extra cache traffic, and a sustained memory bandwidth of 24 billion bytes per second. Each addition requires 24 bytes, so memory alone limits throughput to one billion additions per second. Increasing arithmetic capacity from two to ten billion additions per second cannot overcome that particular limit. Reuse, compression, or changed data movement could. The calculation's assumptions matter more than attaching a famous name to the bottleneck.

Why optimal play may require randomness

For a finite two-player zero-sum game, one player's gain is the other's loss. A mixed strategy is a probability distribution over actions. The minimax theorem states equality between the best guaranteed expected payoff and the opposing player's smallest achievable upper bound when mixed strategies are allowed. It does not require perfect information, and it does not guarantee a saddle point among pure actions. Mixed strategies and minimax

In a constructed matching game, the row player receives +1 if the choices match and -1 otherwise:

	Column H	Column T
Row H	1	-1
Row T	-1	1

Either fixed row action can be defeated. If row chooses H with probability p, its expected payoffs against H and T are 2p-1 and 1-2p. Choosing p=1/2 makes both zero. Column can similarly guarantee that row gains no more than zero. That proves the value of this game is zero. No cell of the matrix alone expresses the solution.

This is a theorem about specified incentives and action spaces. It does not show that a real conflict is zero-sum, that participants know the payoff matrix, or that a numerical utility measure captures all ethical considerations.

Reading historical claims carefully

A design report establishes that an author described an architecture at a particular time. A project archive can establish who implemented it. Neither alone establishes that every later computer descended directly from it. Claims about an inventor's photographic memory, the cause of a terminal illness, or what would have happened without that person need independent evidence. They add little to understanding the technical achievements and are omitted here.

Exercises

1. Change the teaching program so it computes M[10]+M[11]+M[13] into M[12]. Trace it for 7, 5, and -4. Explain why the data addresses must remain distinct from the added instruction.
2. In the matching game, suppose row chooses H with probability 3/4. Find column's best response and row's expected payoff. Explain why a randomized strategy can provide a better worst-case guarantee.
3. A historian finds a sole-authored architectural report and a later implementation report with four different authors. What can each document establish? Give two additional kinds of evidence needed before claiming sole invention.

Compare your reasoning with the worked solutions

1. Use LOAD 10; ADD 11; ADD 13; STORE 12; HALT at addresses 0–4. The accumulator becomes 7, 12, 8, 8, then the machine halts; location 12 holds 8. The extra instruction occupies address 4, so it does not collide with the data at 10–13. If a program expands into a data location, merely preserving instruction order is insufficient: addresses and memory layout must also be updated.
2. Against H, row earns 2(3/4)-1=1/2. Against T it earns 1-2(3/4)=-1/2, so column chooses T. Row's worst-case expected return is -1/2, compared with zero at an equal mixture. Randomness here prevents exploitation of a predictable imbalance; it does not guarantee that an individual play wins.
3. The first document establishes the published description and named author, subject to checking its date and authenticity. The second establishes recorded implementation responsibility. To infer invention priority, examine earlier dated designs or notebooks and correspondence or meeting records showing how ideas were shared. Distinguish authorship, conceptual contribution, engineering realization, and dissemination instead of forcing them into one winner.