Why a literature survey comes before a theorem
A theorem is a statement a thesis proves; a literature survey is how the researcher finds out which statement is worth proving. Before any attempt at a proof, the survey has to establish three things: what is already known and proved about the structures in question, what definitions and notation the field has settled on, and which related questions are open, recently resolved, or known to be false. Skipping this step risks reproving a known result, using a definition that does not match the rest of the literature, or chasing a conjecture that a counterexample already rules out.
The survey also supplies the proof techniques most likely to work. If closely related results were proved by induction on a parameter, by a reduction to a known theorem, or by a probabilistic argument, that is strong evidence about where to start, even though it is no guarantee of success.
Finding a gap, and asking whether it is the right question
The most common outcome of a survey is a gap: a case the existing theorems do not cover, a hypothesis that has not been relaxed, or a structure to which a known proof technique has not yet been applied. Sandberg and Alvesson (2011) call this gap-spotting, and contrast it with problematisation, which questions the assumptions behind the existing results rather than filling a visible hole between them. Both are legitimate starting points for a thesis, but they carry different risks: a spotted gap may turn out to be narrow or uninteresting once examined closely, while a problematised assumption may turn out to be load-bearing for the whole area, so disturbing it changes more than intended.
Before committing to a gap, check three things in the literature: that no unpublished or little-cited source already closes it, that the statement is not a direct corollary of a known theorem under a change of variables, and that small cases do not already provide a counterexample. A morning spent on these checks is cheaper than months spent on a proof of something false or already known.
Running the survey: a concept-centric method
A reference-by-reference survey, in which each source is summarised in turn, tends to lose track of how results relate to one another. Webster and Watson (2002) recommend organising a review around concepts instead: identify the handful of objects, properties or operators central to your question, and for each one, note every definition, theorem and open problem that uses it. This produces a map of the area rather than a list of papers, and the gaps in the map are usually the gaps in the theory.
For doctoral work in mathematics and theoretical computer science, where the literature is large and technical, a systematic protocol keeps the survey reproducible and defensible to an examiner. Kitchenham and Charters (2007) set out a staged process of search, inclusion criteria and extraction originally written for software engineering, and Page et al. (2021) give the PRISMA reporting standard for documenting how many sources were found, screened and included at each stage. Both were written for empirical reviews, but the discipline of writing down the search terms, the inclusion criteria and the reasons for exclusion transfers directly to a mathematical literature survey.
- 1Define the concept or structure of interest and fix notation
- 2Search and screen the literature with recorded inclusion criteria
- 3Map known results, definitions and proof techniques by concept
- 4Identify a gap, unexplored case or questionable assumption
- 5Check small cases and known corollaries before committing
- 6State the conjecture precisely, with hypotheses made explicit
From a surveyed gap to a stated conjecture
A gap is not yet a conjecture. Turning it into one means writing a precise statement: the exact hypotheses, the exact structure they apply to, and the exact conclusion claimed. Vague statements such as "a similar result should hold in the non-commutative case" are not ready to attempt; a conjecture should be stated so that a single counterexample would refute it and a single correct argument would prove it.
Pólya (1954) describes the informal reasoning, induction from special cases and analogy with results in related structures, by which a plausible statement is reached before any proof is attempted. Checking small or special cases by hand or by computation is not a substitute for a proof, but it is good evidence that a conjecture deserves the effort: verify the smallest non-trivial case, then the next one, and look for a pattern in how the argument would need to change as the case grows.
Real theorem development does not move neatly from conjecture to finished proof. Lakatos (1976) documents, through the history of Euler's formula for polyhedra, how a conjecture found in the literature is often restated and its hypotheses tightened as counterexamples are found, sometimes several times before a workable statement is reached. Expect to revise your stated conjecture as you test it, and keep a record of each counterexample and what it forced you to change.
A worked analogue: building theory from a literature-defined gap
Outside pure mathematics, researchers building new theory from a focused literature face the same problem in a different form, and the methods they use are instructive. Cash (2018) sets out a staged method for design research in which a systematic review of a bounded literature is used to derive explicit propositions, each one stated so that it could be supported or contradicted by further evidence, and then tested in further studies. The parallel for a mathematical thesis is direct: a surveyed, bounded body of results yields one or more explicit propositions, and each proposition is tested, here by proof attempt and by checking cases, before being written up as a theorem or abandoned.
A more literal example of combining known literature to produce a new mathematical statement is literature-based discovery. Swanson (1986) showed that two disconnected literatures can share an implicit, previously unstated link, call it B, such that one literature establishes a relation between A and B and the other establishes a relation between B and C, without either literature stating the resulting A-to-C connection. The same logical structure appears in mathematics whenever a result proved in one subfield, expressed in another subfield's notation, reveals a connection that neither subfield's specialists had stated outright.
Automated conjecture generation: what it changes, and what it does not
Algorithmic tools now search within the known literature on certain classes of mathematical objects to propose new conjectures. The Ramanujan Machine, for example, systematically generates candidate continued-fraction expressions for fundamental constants, flags the ones that match to high numerical precision, and leaves formal proof to mathematicians; several of its conjectures were later proved (Raayoni et al., 2021). This is literature-grounded conjecture generation in a strict sense: the search space, the constants and the expression families, is defined entirely by what the literature already studies.
For a doctoral researcher, the practical lesson is that such tools narrow a literature survey, they do not replace it. A proposed conjecture still needs to be checked against the existing literature for prior statement or disproof, situated within known results, and, crucially, proved before it can be called a theorem. Numerical agreement to many digits, however striking, is evidence, not proof.
What to report, and common mistakes
| Element | What to record from the survey |
|---|---|
| Search and inclusion | Search terms, databases and dates searched, inclusion and exclusion criteria |
| Concept map | Central objects and properties, and which sources define or use each one |
| Known results | Exact statements and hypotheses of the closest prior theorems, with precise citations |
| The gap | Why it is a gap and not already a corollary, counterexample or duplicate of known work |
| Evidence for the conjecture | Special cases checked, analogies drawn, and where each came from |
| Stated conjecture | Exact hypotheses and conclusion, versioned if restated after a counterexample |
- Treating "no one seems to have proved this" as evidence that the statement is true, rather than checking it against small cases first.
- Surveying papers rather than concepts, so related results scattered across subfields are missed.
- Stating a conjecture too loosely to be falsified by a single counterexample.
- Citing a result for a claim it does not actually make, rather than quoting its exact hypotheses.
- Treating output from a conjecture-generating tool as proved, or omitting that the tool was used.
- Abandoning a survey-identified gap at the first counterexample instead of tightening the hypotheses.
Key takeaways
- A literature survey for theorem development has a narrower aim than a general reading list: the exact proved statements, definitions and open problems in the area (Webster & Watson, 2002).
- Gap-spotting and problematisation are different ways a survey can generate a question, and carry different risks (Sandberg & Alvesson, 2011).
- A systematic, documented search process makes a mathematical survey reproducible and defensible to an examiner (Kitchenham & Charters, 2007; Page et al., 2021).
- A conjecture should be checked against special cases and known corollaries, and stated precisely enough to be falsified, before serious proof effort begins (Pólya, 1954; Lakatos, 1976).
- Literature-based discovery and algorithmic conjecture tools narrow a survey; they do not remove the need for one, and their output still needs proof (Swanson, 1986; Raayoni et al., 2021).
- Explicit, testable propositions derived from a bounded literature are a transferable method from theory-driven design research (Cash, 2018).
