WorksMathematics
A lower bound 0.11096505 for Romanoff's constant
Abstract
Let d be the lower density of the integers of the form p + 2^a with p prime and a >= 1. We prove d >= 0.11096505, unconditionally. The previous best unconditional bound is d >= 0.110825, due to Tomar. The proof is Tomar's method, with the same analytic inputs, applied to a different modulus of the same period 27720; the bound is certified in exact rational arithmetic by Tomar's checking program, with changes that only make the computation faster and that we verify against the original program.
Result
For a positive integer let be the number of pairs with prime, and , and
Romanoff proved [6]. Explicit lower bounds are 0.0868 (Chen and Sun [1]), 0.0936275 (Habsieger and Sivak-Fischler [3]), 0.107648 (Elsholtz and Schlage-Puchta [2], recomputed as 0.10788 by Johnston and Trudgian [4]) and 0.110825 (Tomar [8]). In Tao’s list of optimization constants [7] this is the lower bound for .
Theorem 1. We have .
The theorem is unconditional: it uses exactly the analytic inputs of Theorem 1.1 of [8], namely the per-class sieve bound with the constant 8 quoted by Johnston and Trudgian from [2 (28)], the prime number theorem in arithmetic progressions to a fixed modulus, Pintz’s bound for the series [5], and Pintz’s inequality for nonnegative integers. It does not use the conditional input of Theorem 1.2 of [8].
The method
We recall the method of [8], which refines [2]; we refer to [8] for the proofs. Let be an odd squarefree integer and the order of 2 modulo . For a residue class mod let and . In each class the first moment of is given by the prime number theorem in progressions, and the second moment is bounded by the per-class sieve bound, applied to the pairs grouped by , together with upper bounds () for the singular series. The part of the singular series that is not computed is bounded only through its total mass , and the worst distribution of this mass over the divisors of is taken. With Pintz’s inequality in each class this gives ([8 Proposition 4.2])
where , and are explicit constants depending on , the are arbitrary integers, and runs through the vertices of the simplex .
The bound depends on the modulus . Tomar uses
with period , and obtains 0.110825. We use
that is, without 241 and with 89 and 127, for which and . The period is again .
The computation
The computation uses the structure of [8 Appendix A] and the certificate format r45a-u3f of its checking program. The base primes are 3, 5, 7, 11, 13, 17, 31 (period 120; their nonempty patterns fall into 822 rotation orbits). The core primes are 41, 19, 73, 37 (period 360). The remaining primes are treated as independent components: 43, 29, 127 for the factor 7 of the period (orders 14, 28, 7) and 23, 89 for the factor 11 (orders 11, 11). The program checks that the orders of 2 modulo these primes come from proven complete factorizations, that this decomposition is valid, that the multiplicities of the combined patterns (380 of them) add up to the product of the independent primes, and that the total weight equals .
The singular series is bounded as in [8 Section 5], from the complete factorizations of for and the known factorizations for . The program gives and a tail mass . The Pintz indices range over . The value at the vertex is 0.1110554 . . .; the minimum over the 97 vertices is at , and the program checks in exact rational arithmetic that it is at least 0.11096505. This proves Theorem 1.
Changes to the checking program
Tomar’s program [8] would need about 9.6 hours on two cores for this certificate. We changed three things, all inside the u3f part of the program, and an ancillary file lists them exactly.
In each base orbit, identical rows of the core enumeration (the same mask on ) are merged and their weights added. Every quantity computed from a row depends on the row only through its mask and is linear in its weight, so the sums do not change; the program checks that the total weight is unchanged.
The weighted sums over the rows are computed in 64-bit integers, with an asserted bound against overflow, instead of in floating point; the original bound that makes the computed exact is kept.
The base orbits can be processed in slices, with the exact partial sums written to a file after each orbit, and the slices are recombined exactly; the recombination repeats all the checks of the original program and checks that the slices cover every orbit exactly once.
With these changes the computation took about three hours on two cores.
Checks
On Tomar’s test certificate
test_cert_t1_u3f, the changed program gives 0.107200250, the value of Tomar’s log, both in one run and in two slices recombined.Tomar’s original program, changed only so that it processes a slice of orbits and writes its exact sums, and the changed program were run on the certificate of this note for the slices of orbits 0–1 (the most expensive), 410–413 and 815–821 (the least expensive). The exact sums agree in all 2717, 1747 and 292 integers respectively.
The original program accepts the certificate of this note: all its checks up to the evaluation of the orbits pass.
The original program has not been run end to end on this certificate, and the computation has not been formalized in Lean (Tomar’s bound 0.110825 is also proved in Lean [8]).
Ancillary files
The directory anc/ contains the following files.
The certificate
cert_E_27720.json.The changed program
check_r45u3f_patched.pyand the list of its changes.The exact partial sums
partial-sums_0_822.json, and the logs of the run and of the recombination.Tomar’s factorization data
fac300_41e4cdd6.jsonandfac1200_fdb_af082adc.json.For the comparison with the original program:
chk_orig_tranche.py, its list of changes, the slice sums andcmp.py.The Apache License 2.0, under which Tomar’s program is distributed.
To reproduce, with Python 3, python-flint and numpy:
TRANCHE=0:822 SORTIE=s.json python3 -I check_r45u3f_patched.py \
fac300_41e4cdd6.json cert_E_27720.json 3000000 \
fac1200_fdb_af082adc.json 1200 --workers 2
RECOMBINE=s.json python3 -I check_r45u3f_patched.py \
fac300_41e4cdd6.json cert_E_27720.json 3000000 \
fac1200_fdb_af082adc.json 1200 --claim 0.11096505
Acknowledgements and use of AI
The method and the checking program are those of T. S. Tomar [8], building on Elsholtz and Schlage-Puchta [2], Johnston and Trudgian [4] and Pintz [5]. The mathematics, the computations and the writing were produced with Claude (Anthropic) under the author’s direction. The result has not yet been examined by a professional mathematician.
Independent researcher, Meaux, France
URL: https://orcid.org/0009-0009-1526-1178
References
- [1]Y.-G. Chen and X.-G. Sun, On Romanoff’s constant, J. Number Theory 106 (2004), 275–284.DOI
- [2]C. Elsholtz and J.-C. Schlage-Puchta, On Romanov’s constant, Math. Z. 288 (2018), 713–724.DOI
- [3]L. Hasieber and J. Sivak-Fischler, An effective version of the Bombieri–Vinogradov theorem, and applications to Chen’s theorem and to sums of primes and powers of two, Arch. Math. (Basel) 95 (2010), 557–566.
- [4]D. R. Johnston and T. Trudgian, An update on the Linnik–Goldbach problem, arXiv:2605.17825v2 (2026).arxiv.org/abs/2605.17825
- [5]J. Pintz, A note on Romanov’s constant, Acta Math. Hungar. 112 (2006), 1–14.DOI
- [6]N. P. Romanoff, Über einige Sätze der additiven Zahlentheorie, Math. Ann. 109 (1934), 668–678.
- [7]T. Tao et al., Optimization constants in mathematics, constant C₄₅, https://teorth.github.io/optimizationproblems/constants/45a.html.
- [8]T. S. Tomar, New lower bounds for Romanoff’s constant, version 11 (2026), doi:10.5281/zenodo.23715955; programs at https://github.com/tejasvi/romanoff-constant, tag v11.DOI