We're 99.9% sure this pattern is true, but no one can prove it

Scientific concepts / mathematical phenomena (ideas and discoveries)

Twin primes and their expected frequency

• Twin primes conjecture (unproved): there are infinitely many prime pairs of the form ((p, p+2)).
• Prime gaps trend: the average gap between consecutive primes near size (N) grows like (\ln(N)), so prime twins become rarer.

• Heuristic expectation (Hardy–Littlewood):

  • By the Prime Number Theorem, the probability a large number near (N) is prime is about (1/\ln(N)).

  • Heuristically, the probability both (N) and (N+2) are prime is about [ \approx \frac{1}{(\ln N)^2}. ]

  • Summing/integrating these probabilities predicts the growth of the number of twin primes up to (N), matching computations closely.

  • Key limitation: the Hardy–Littlewood prediction is a heuristic, not a proof—twin primes might still be finite.

Sieve methods and “main term vs error term”

• Sieve of Eratosthenes (classic tool):
  • Iteratively remove multiples of primes.
  • To find primes up to (N), it suffices to sieve using primes up to about (\sqrt{N}).
• Inclusion–exclusion:
  • Use alternating add/subtract steps to account for overlaps among “removed” sets.
• Brun’s twin-prime sieve (1930s):
  • Adjusts the sieve so that when sieving by a prime (p), it removes (N) whenever either:
    • (N) would be divisible by (p), or
    • (N+2) would be divisible by (p).
  • Produces an “almost twin” counting method.
• Why Brun’s full-strength approach fails to prove infinitely many true twin primes:
  • Rounding/control errors blow up for twin sieves, because inclusion–exclusion creates many terms.
  • The error term overwhelms the main term, preventing a rigorous positive lower bound.

From “twin primes” to “almost twin primes”

• Brun’s breakthrough result:
  • By weakening the sieve (not sieving all the way to (\sqrt{X}); instead only to a smaller power of (X)) and controlling errors, he proved:
  • There are infinitely many pairs of integers two apart where each integer has at most 9 prime factors (counted with multiplicity).
• Refinements of Brun’s method:
  • Improved bounds for the number of prime factors:
    • 9 → 7 → 3
• Chen Jingrun’s theorem (1973):
  • Proved infinitely many primes (P) such that (P+2) has at most two prime factors (so it is either prime or an “almost prime” in the semiprime-type sense used in the statement).

Bounded gaps between primes (GPY → Zhang → Maynard)

A different route aims to prove primes occur with bounded distance infinitely often, not necessarily with gap (2).

Average gap and relative improvement

• Typical gap between consecutive primes is about (\ln x).
• Over time, results showed guaranteed small gaps occur on a shrinking “relative to the average” scale.
• Goldston–Pintz–Yıldırım (2005):
  • Proved gaps are frequently much smaller than the average—the ratio to the average can be made arbitrarily small infinitely often (but not yet a fixed absolute bound).

GPY method (2005): a “stencil” + weighting strategy

• Core setup: place a finite pattern (a stencil) of positions along the number line; attempt to force at least two primes among those positions.
• Need averaging: since exact prime locations are unknown, GPY uses an average count of primes hit by the stencil over long intervals, using distribution results for primes.
• Key technique: a weighted averaging machine
  • starting positions get lower weight if the stencil would force too many residue classes that are likely composite (e.g., forced evenness or divisibility by small primes).
• Main obstacle:
  • requires strong information about primes in arithmetic progressions, depending on a level of distribution parameter (often described via (\theta)).
• “Impossible wall” near the half level:
  • GPY could not cross a barrier around (1/2)—they could get “almost” enough, but not exceed the threshold required for bounded gaps.

Yitang Zhang’s breakthrough (2013)

• Zhang changed the approach to distribution in arithmetic progressions:
  • used special moduli built from small prime factors,
  • reorganized error terms so most cancel,
  • achieved a tiny improvement past the (1/2) threshold (described as pushing beyond it by about (1/584)).
• Result: proved a bounded gap between primes infinitely often.
  • In the narrative, the bound is described via a scale (stencil size) corresponding to about 70 million.

Polymath and ongoing improvements

• Polymath project (T. Tao):
  • collaborative refinements of Zhang/GPY-style methods,
  • repeatedly reduced the best known prime-gap bound.
  • ultimately mentioned improvement to 4,680 (as a world-record level in the video timeline).

Maynard’s “orthogonal” approach and multi-prime windows

• James Maynard (2013–2014):
  • developed a different method (in this narrative) that avoids treating the “half barrier” as a fundamental limit.
  • proves results like: 3 primes in a bounded window, with bounds depending on how many primes are targeted.
  • quickly improved the prime-gap record to around 600.
• Tao and Green connection: Maynard eventually joins forces with the Polymath collaboration (2014).

Current unconditional bound mentioned

• The video states the best unconditional bound reached is 246:
  • i.e., infinitely many prime pairs differing by at most 246.
• Fields Medal: Maynard received the Fields Medal (2022) for prime-gap work.

Conditional results via distribution conjectures (Elliott–Halberstam)

• Conditional on strong hypotheses about prime distribution in arithmetic progressions:
  • Under the Elliott–Halberstam conjecture (or stronger variants), the bounded gap record could be reduced further:
    • down to 12 (Maynard, 2013, conditional),
    • down to 6 (Polymath, a year later, conditional).
• Unconditional (no assumptions): the story keeps 246 as the current bound.

---

Methodologies / logical structures (as bullet points)

Hardy–Littlewood heuristic framework (twin prime counting)

• Use the Prime Number Theorem to approximate prime likelihood near (N).
• Treat “prime events” as approximately independent (noted explicitly as false, hence heuristic).
• Estimate:
  • (\Pr(N \text{ prime}) \approx 1/\ln N),
  • (\Pr(N+2 \text{ prime}) \approx 1/\ln(N+2)\approx 1/\ln N).
• Multiply to get twin probability (\approx 1/(\ln N)^2).
• Sum/integrate over (N) up to the target limit to predict twin prime counts.
• Include a constant correction factor accounting for non-independence.

Brun-style sieve and inclusion–exclusion (conceptual)

• Begin with the ordinary sieve removal rule using primes as factors.
• For twin primes, remove numbers where either:
  • (N) is divisible by a sieve prime (p), or
  • (N+2) is divisible by (p).
• Use inclusion–exclusion to count “removed” and account for overlaps across primes.
• Encounter rapid growth in:
  • rounding and error terms as the number of sieving primes increases.
• Mitigate by:
  • weakening sieve extent (only sieve up to a smaller power),
  • controlling error growth to get an infinite “almost twin” result.

GPY/Maynard-style bounded gap strategy (stencil + averaging)

• Choose a finite set of offsets: a stencil spanning a window of width (H).
• Slide the stencil across large intervals and count primes it hits.
• Replace unknown exact prime positions with averaged expected counts using prime distribution in arithmetic progressions.
• Add weights to de-emphasize placements structurally unlikely to produce primes (e.g., forced divisibility by small primes).
• Show the averaged weighted count exceeds a threshold (\Rightarrow) there must exist placements with multiple primes (\Rightarrow) bounded gaps infinitely often.
• Bottleneck: strength of distribution in arithmetic progressions (level of distribution / (\theta)).

Zhang’s refinement

• Keep the GPY scheme but:
  • restrict/modulate the choice of moduli to special forms from small prime factors,
  • reorganize errors so cancellations occur,
  • obtain a small improvement beyond the (1/2) barrier.
• Enough to yield an explicit (though large) bounded gap.

---

Researchers / sources featured (named or strongly implied)

• Derek (narrator/speaker; likely tied to the channel context)
• Edmund Landau (cited as characterizing a problem as “unattackable”)
• Hardy (G. H. Hardy)
• Littlewood (J. E. Littlewood)
• Terry Tao
• Viggo Brun
• Chen Jingrun (1973 result)
• Goldston (in GPY)
• Pintz (in GPY)
• Yıldırım (in GPY)
• Andrew Granville
• Kannan Soundararajan
• GPY authors: (Goldston, Pintz, Yıldırım)
• Yitang Zhang
• James Maynard
• Roger Heath-Brown (advisor mentioned)
• Ben Green (mentioned via communication)
• Terence Tao Polymath group (Tao-led collaboration)
• Elliott–Halberstam conjecture (named)
• Scientific American (mentioned in the 2013 context)
• Roger Bannister (4-minute mile analogy)
• John Landy (4-minute mile analogy)