Thesis: Voynich research begins at the wrong end of a transformation chain. We possess a surviving representation and try to infer backwards toward the world, practices and distinctions that generated it. That is an inverse problem, and inverse problems can be non-unique, unstable and highly sensitive to assumptions.
This article develops the mathematical core underneath The Representation Trap and The Original Receiver Is Missing.
We Have the Output
Imagine a machine with an unknown internal process.
You feed something into it. The machine transforms the input through hidden stages. You do not see those stages. You see only the final output.
The Voynich Manuscript places us in a similar position.
LOST WORLD / PRACTICE → OBSERVATION → EXTRACTION → TOKENISATION → REPRESENTATION → MANUSCRIPT
The right side survives.
We are attempting to reason left.
MANUSCRIPT → ? TOKEN SYSTEM → ? EXTRACTION RULE → ? REPRESENTATIONAL PURPOSE → ? PRACTICE → ? WORLD
That reversal is not guaranteed to work uniquely.
Forward Problems Are Easier Than Inverse Problems
In a forward problem, we know a model and predict its output.
KNOWN SYSTEM + KNOWN INPUT → PREDICT OUTPUT
In an inverse problem, we know an output and try to infer the hidden system or input.
OBSERVED OUTPUT → INFER HIDDEN SYSTEM
Medical imaging contains inverse problems. Astronomy contains inverse problems. Seismology contains inverse problems. Archaeology contains inverse problems. Machine learning frequently contains inverse or latent-variable problems.
They are difficult because different hidden realities can produce similar observations.
Many Worlds Can Produce One Representation
Let R be an underlying reality and f be the representational process.
P = f(R)
If f discards information, multiple realities may produce the same or nearly the same P.
f(R₁) = P f(R₂) = P f(R₃) ≈ P
A simple example is a silhouette. Many three-dimensional objects can cast the same two-dimensional shadow from one angle. The shadow is real evidence, but it does not uniquely determine the object.
Voynich may contain similar projection loss. A recurring mark can be consistent with a letter, abbreviation, morpheme, syllable, number, operational marker or component of a larger unit. A circular diagram can be compatible with astronomy, calendar structure, medical correspondence, mnemonic organisation or another system that happens to use radial form.
The point is not that anything goes. The point is that fit alone is insufficient when the inverse is non-unique.
The Problem of Ill-Posedness
A classical well-posed problem ideally has a solution, a unique solution, and a solution that changes stably when the data change slightly.
Voynich interpretation threatens all three conditions.
Existence
We do not know whether the solution class we are searching contains the historical mechanism. If we search only among languages and ciphers while the system is partly mnemonic or operational, the correct model may be absent from the candidate set.
Uniqueness
Several models can reproduce the same statistical feature. Low conditional entropy, Zipf-like distributions, recurrence and page clustering are not unique signatures of one mechanism.
Stability
Small changes in transcription, glyph inventory, spacing assumptions or corpus segmentation can alter downstream statistics. A theory that changes dramatically under modest representation changes is unstable.
This is why inverse problems require regularisation, controls and explicit prior assumptions.
Our Priors Enter Before the Mathematics
Suppose we decide before analysis that visible gaps are word spaces.
Now our token inventory, word lengths, vocabulary growth, entropy calculations and language comparisons all inherit that decision.
The mathematics may be flawless.
The inverse model may still be wrong because the units were wrong.
ASSUMPTION → REPRESENTATION → MEASUREMENT → RESULT
Once the result looks sophisticated, the originating assumption can disappear from view.
This is one reason the EVA segmentation problem matters beyond transcription. It demonstrates how a representation can become the world a later model sees.
Regularisation: How We Stop the Solution From Becoming Anything
Inverse problems need constraints.
In Tangential Voynich, regularisation comes from several directions:
- Material chronology constrains historical possibilities.
- Codicology constrains page relationships.
- Paleography constrains production models.
- Transcription uncertainty constrains numerical precision.
- Matched historical controls constrain genre claims.
- Hostile artificial controls constrain “looks like language” claims.
- Held-out pages constrain overfitting.
- Predeclared rules constrain retrospective adjustment.
- Receiver reconstruction constrains what information may plausibly have been tacit.
The goal is not to force one answer early. It is to reduce the size of the solution space without smuggling the answer into the constraints.
Why a Beautiful Decoding Can Still Be a Bad Inverse Solution
A decoding can be locally impressive and globally weak.
If a researcher assigns meanings to signs and obtains plausible phrases, that establishes compatibility with selected material. It does not establish uniqueness.
A stronger inverse solution must predict held-out material, preserve mapping rules, explain exceptions, survive different transcriptions and outperform alternative generative mechanisms.
An inverse solution is strong when the observed representation becomes difficult to produce under competing hidden worlds.
The Null Space of Voynich
In linear inverse problems, a null space contains changes to the hidden input that do not alter the observed output.
Voynich has an analogous conceptual problem.
If several changes in underlying interpretation leave the measured features unchanged, those features cannot distinguish among them.
For example, if natural language, a verbose cipher and a copy-modify generator all produce similar character entropy, entropy alone lies partly inside a discriminative null space. It is real structure but weak for choosing among those models.
Research should move toward measurements outside that shared space.
This reframes progress. A famous feature becomes less valuable when more alternative mechanisms can reproduce it. A previously obscure feature becomes more valuable when competitors cannot.
Tangential Systems as Basis Changes
There is a mathematical analogy for what Tangential Voynich is doing.
The same object can be represented in different bases. A vector can have different coordinate descriptions without changing the underlying vector.
Tangential Voynich changes the observational basis.
VOYNICH → LINGUISTIC BASIS → CODICOLOGICAL BASIS → RAILWAY BASIS → MUSICAL BASIS → COMPILER BASIS → ECOLOGICAL BASIS
We do not claim that the object itself changes.
We ask which coordinates become easy to see under each basis.
Then we convert surviving observations back into neutral language.
If the same feature persists across basis changes, it becomes an invariant candidate.
The Danger of the Historically Nearest Basis
Historical proximity is necessary for provenance and genre testing. But it can be a poor environment for discovering observer bias because its vocabulary is too plausible.
Herbal, medical, astronomical and pharmaceutical models can explain many visible cues before making any hard prediction.
A deliberately distant system cannot hide behind plausibility.
If a railway model produces a useful graph prediction, the prediction must stand on structure because nobody can rescue it by claiming medieval railway culture.
This is the methodological advantage of absurd distance.
From Fit to Identifiability
Voynich theories often ask whether a model fits.
Inverse-problem thinking asks something stronger:
Is the hidden mechanism identifiable from the observations we actually possess?
If not, what additional observation would make it identifiable?
That turns research toward experiment design.
- Would higher-resolution imaging distinguish two glyph-generation models?
- Would material analysis separate production phases?
- Would a held-out quire distinguish competing token models?
- Would a known historical comparator discriminate visual function?
- Would a boundary experiment distinguish lexical spacing from internal segmentation?
- Would independent scorers reproduce a visual component family without semantic labels?
The best next test is the one that shrinks the equivalence class of possible hidden worlds.
World Return
Voynich is not merely a puzzle with an answer hidden behind a lock.
It may be an inverse problem with information loss.
That distinction changes our expectations.
A key opens a lock completely when the key is correct.
An inverse problem may permit several solutions until new observations constrain it.
So the task is not to become more imaginative about one solution.
It is to design measurements that make alternative solutions stop looking equivalent.
We do not need the theory that can explain Voynich after seeing it. We need measurements that make the wrong hidden worlds predictably fail.
That is the mathematical reason Tangential Voynich rotates representations.
Every rotation is another attempt to discover which parts of the solution space are real structure—and which parts are only the coordinates we chose.
Continue: Voynich | The Mathematics and Science · The Representation Trap · The Original Receiver Is Missing · Voynich Research Library