WorksMathematics
New lower bounds for six small hypergraph Ramsey numbers
Abstract
Let denote the least such that every colouring of the -subsets of an -set with two colours has points all of whose -subsets have the first colour, or points all of whose -subsets have the second. We give explicit colourings proving , , , , and . The first three improve the bounds 35, 63 and 88 listed in Radziszowski's dynamic survey. The last three improve the listed bounds 67, 133 and 72, and for the two 4-uniform numbers also the values 159 and 317 of a recent data deposit. The colourings of 3-subsets are invariant under groups of affine maps of , for , 69 and 97, in the last two cases with one further fixed point. Those of 4-subsets live on the projective line over , for and , and are invariant under or , so each is given by a table of entries. For 5-subsets we show that, for every prime , invariance under forces six points all of whose 5-subsets have the same colour; our colouring is invariant under all affine maps of . By known constructions the bounds give and , and , and . The colourings, and programs that verify them, accompany the paper as ancillary files.
Introduction
For integers and , the hypergraph Ramsey number is the least such that every colouring of the -subsets of an -set with the colours has, for some , a set of points all of whose -subsets have colour . We write when . With two colours, which we call 0 and 1, a colouring of the -subsets of an -set with no points whose -subsets all have colour 0 and no points whose -subsets all have colour 1 is a -colouring of order ; it proves . We call the 3-subsets triples.
Very little is known about these numbers for . The only nontrivial value known is , computed by McKay and Radziszowski [2]. Radziszowski’s dynamic survey [1, Section 7.1] lists the lower bounds , due to Dybizbański [3], and and , also due to Dybizbański [4]; these come from explicit colourings. Among the bounds that the survey obtains from recurrence relations and related methods [11, 12] are , and . Two data deposits of 2026 report further bounds: [6], and and [7]. The last two come from colourings of the projective line that are invariant under ; from the second of them and a recurrence, [7] also states , without a colouring.
Theorem 1. , and .
Theorem 2. and .
Theorem 3. .
The proofs consist of explicit colourings. Those for Theorem 1, of orders 37, 70 and 98, are invariant under groups of affine maps of and are described in Section 2. Those for Theorem 2, of orders 200 and 600, are colourings of the 4-subsets of a projective line that are invariant under its group or (Section 3). The colouring for Theorem 3, of order 83, is invariant under all affine maps of (Section 4). Section 5 explains how the colourings were verified and how a reader can verify them again. Table 1 compares the bounds.
Table 1. Previous and new lower bounds. The first six rows are Theorems 1, 2 and 3; the last four are Corollaries 4 and 5, obtained from them by known constructions. The values marked † are not stated in [7]: we derive them from its bound by the recurrences of Section 6.
Two known constructions turn these bounds into further ones (Section 6). The survey notes that and a product construction of Bruce, Budden and Hiller [5] give ; the same construction applies to our colouring of order 98. And the recurrences of Song, Ye and Liu [11], from which the survey derives its bounds for , apply to our bound for .
Corollary 4. for every . In particular .
Corollary 5. , , and .
The colourings of triples
Let and let be a unit of of multiplicative order . We write
for the group of en affine permutations of generated by the translations and the multiplication by . We also let act on , with every map fixing the point . A colouring of the triples of the point set is -invariant if every map of the group preserves the colour of every triple. Such a colouring is constant on the orbits of the group on triples, so it is described by naming the orbits that receive colour 0.
We define three colourings.
is a -invariant colouring of the triples of .
is a -invariant colouring of the triples of .
is a -invariant colouring of the triples of .
In each case colour 0 is the union of the orbits of the triples listed in Appendix § 10, and every other triple has colour 1. Table 2 gives the parameters.
| Points | |||
| Multiplier (its order) | (3) | (11) | (24) |
| Order of the group | 111 | 759 | 2328 |
| Orbits of triples | 78 | 74 | 72 |
| by size | |||
| Orbits in colour 0 | 32 | 27 | 36 |
| by size | |||
| Triples in colour 0 | 3182 | 19 803 | 76 048 |
| Triples in colour 1 | 4588 | 34 937 | 76 048 |
Table 2. The three colourings of triples. A term stands for orbits of size .
Proposition 6. is a -colouring of order 37, is a -colouring of order 70, and is a -colouring of order 98.
Computational verification. By the exhaustive search described in § 5.
Theorem 1 follows at once from Proposition 6.
Remark 7 (a colour-reversing symmetry). The map of , which fixes , exchanges the two colours of on every triple. Since 2 generates the squares modulo 97 and 4 generates the fourth powers, the multiplications by fourth powers preserve and the multiplications by the other squares reverse it. In particular the two colour classes of are isomorphic, and each has triples.
Remark 8. is not invariant under the projective map of the line : that map preserves the colour of only 74 512 of the 152 096 triples. So is not invariant under in its action on this line.
Colourings of the projective line
Let be a prime and let be the projective line over the field with elements, a set of points. The group acts on by the maps with . For four distinct points of the cross-ratio is
where the two factors that contain , if any, are cancelled against each other; for instance . The cross-ratio lies in and is unchanged by every map of . Reordering the four points replaces a value by one of
and we call this set of at most six values the class of . Since is sharply 3-transitive on , two 4-subsets of lie in one orbit of the group exactly when their cross-ratios lie in one class. Hence a -invariant colouring of the 4-subsets of is the same thing as a function that is constant on classes: the colour of is . There are about classes, so such a colouring is a short table.
Colourings of the 4-subsets of that are invariant under the subgroup were used by Özdemir [7], with for and for . We searched the -invariant colourings, for at larger and for directly, and define three colourings by their functions .
is a colouring of the 4-subsets of ; its takes the value 0 on 15 of the 33 classes.
is a colouring of the 4-subsets of ; its takes the value 0 on 41 of the 84 classes.
is a colouring of the 4-subsets of ; its takes the value 0 on 51 of the 100 classes.
Appendix § 11 lists these classes. has 26 059 632 of its 57 211 376 4-subsets in colour 0; has 1 273 572 750 of 2 573 031 125; has 2 740 245 300 of 5 346 164 850.
Proposition 9. is a -colouring of order 194. and are -colourings of orders 500 and 600.
Computational verification. By the exhaustive searches described in § 5.
The second bound of Theorem 2 follows from ; the first is proved below, with a smaller group. We record as well because its verification uses no symmetry at all.
Appendix § 12 records the primes at which our searches found colourings of these kinds, and those at which the solver reported that the family contains none.
The smaller group. The subgroup of consists of the maps for which is a nonzero square. It is the group used in [7]. Each orbit of on the 4-subsets of is one orbit of the subgroup or splits into two, so the subgroup admits more invariant colourings. Let , so that is not a square. For distinct the map
sends , , to , , and any fourth point to , and its determinant is
If is a square, lies in . If not, the map does, and it sends , , , to , , and . So every 4-subset of lies in the orbit of a set , and a -invariant colouring is known from the function on for which is the colour of : for distinct and a fourth point of ,
The next lemma says which functions occur. We call the set of values that it describes the narrow class of .
Lemma 10. Let and . The values for which lies in the orbit of under are , and if and are both squares, and all the values in the class of otherwise.
Proof. Write . The map has determinant and permutes , , cyclically; it and its square send to and to , so these values always occur. The maps and send to itself and have determinants and ; with their product and the identity they form a group that permutes the four points of transitively.
Suppose that or is not a square, so that some lies outside . If lies in the class of , some sends to , and so does ; one of the two lies in . Suppose now that both are squares, so that lies in , and let send to . Choosing with , we obtain a map of that sends to itself. The six maps of that do this are , and , of determinant , and , and , of determinant , which is not a square. So is one of the first three, and is one of , and .
Hence, for , the -invariant colourings of the 4-subsets of are given through (1) by the functions that are constant on narrow classes, and the orbits of the group on 4-subsets correspond to the narrow classes.
We define as the -invariant colouring of the 4-subsets of whose function takes the value on the 20 narrow classes listed in Appendix B, out of 43, and the value on the others. It has 27 909 750 of its 64 684 950 4-subsets in colour 0.
Proposition 11. is a -colouring of order 200.
Computational verification. By the exhaustive search described in § 5.
So , which is the first bound of Theorem 2. The colouring gives with the larger group.
A colouring of 5-subsets
Colourings with the full symmetry of a projective line cannot be used for , for the following reason.
Lemma 12. Let a group act on a set , let , and let be a set of points of . If some subgroup of maps to itself and permutes its points transitively, then in every -invariant colouring of the -subsets of all -subsets of have the same colour. In particular there is then no -invariant -colouring of .
Proof. The -subsets of are the complements in of its points, so the subgroup permutes them transitively, and an invariant colouring gives them one colour.
For every prime the lemma applies to acting on , with . For the group permutes the six points of transitively. For the maps with permute the six points transitively. For consider the subgroup that maps to itself. It has six elements. Each of its three elements of order two fixes at most two points of , and its two elements of order three, which are inverse to each other, fix the same points, at most two. So at most eight points are fixed by a non-identity element, and some point of has an orbit of six points, on which the subgroup acts transitively. In each case Lemma 12 shows that no -invariant -colouring of exists.
Let be the group of all maps of with , for a prime . It has elements. If , the lemma applies to as well: the six solutions of in are permuted transitively by the multiplications by these same six numbers. If , it does not apply: a group that permutes six points transitively has an order divisible by 6, and 3 does not divide . This is the case for , where .
The group is sharply 2-transitive: for two points the map sends to 0 and to 1. So every 5-subset of can be moved to one that contains 0 and 1, and an -invariant colouring of the 5-subsets is determined by its colours on the subsets that contain 0 and 1.
We define as the -invariant colouring of the 5-subsets of whose colours on the subsets through 0 and 1 are given in the ancillary file anc/R665_83_points_short.json. It has 14 500 183 of its 29 034 396 5-subsets in colour 0, and the 5-subsets fall into 4276 orbits of .
Proposition 13. is a -colouring of order 83.
Computational verification. By the exhaustive search described in § 5.
Theorem 3 follows.
Verification
The check. Let be the size of the subsets that are coloured, a colour and a size. A set of points all of whose -subsets have colour is called one-coloured. Our programs search for a one-coloured set of points by backtracking: a one-coloured set is extended by a point larger than all of its points, and the new point is accepted only if every -subset that consists of it and points of the set has colour . The search is exhaustive, because every initial segment of a one-coloured set, in increasing order of its points, is one-coloured. A colouring is accepted as a -colouring only if the search finds no one-coloured set of points in colour 0 and none of points in colour 1. For the colourings of triples the sets so excluded number and for ; and for ; and for each colour of .
Independent programs for the colourings of triples. Each of , and was checked by several independently written programs, none of which shares code with the search that produced the colourings. One of them, anc/verify.py, accompanies this paper. It uses only the Python standard library and reads only a plain list of the triples with their colours, so it depends on no ordering convention. It was also run on altered colourings in which a one-coloured set had been planted, and reported it. On an ordinary laptop anc/verify.py needs a few seconds for each colouring.
From the paper to the files. The script anc/rebuild_from_rule.py rebuilds each colouring from its group and the representatives of Appendix § 10, and compares the result with the stored colouring. So the printed data and the ancillary files describe the same three objects. § 8 gives SHA-256 digests.
The colourings of 4-subsets. For , and the colour of a 4-subset is computed from the table and the cross-ratio of its points taken in increasing order, with as the last point. The program anc/verify_crossratio.c treats this as a plain function on 4-subsets and runs the search described above, with bit vectors, for 4-subsets in place of triples. It assumes nothing about the table and uses no symmetry.
For it examined all 83 870 080 one-coloured sets of 5 points in colour 1; none extends to 6 points, and no 5 points are one-coloured in colour 0.
For it examined 6 411 351 600 one-coloured sets of 5 points in colour 0 and 6 647 428 500 in colour 1; none extends to 6 points.
For the program first confirmed, on each of the 5 346 164 850 4-subsets, that the map of , which fixes , preserves the colour. Every set of at least two points contains a point other than and so has a translate through the point 0; the program then examined the 147 041 921 and 121 788 680 one-coloured sets of 5 points through 0 in the two colours, and none extends to 6 points.
Each run began with controls: a colouring known to be valid had to pass, and tables altered on purpose had to fail, among them one that does not preserve, which the program must refuse. On , the analogous colouring for , the search through 0 and the search without symmetry gave the same answer.
was checked a second time in a different way: the colouring was written out in full, one colour for each of its 57 211 376 4-subsets, and the program anc/verify_rank.c examined every one of the sets of 5 points and of the sets of 6 points. The two programs also agree on the number of 4-subsets in colour 0.
A second program for the tables . The program anc/verify_projective.cpp was written independently of the first. It checks that a table is constant on the classes of cross-ratio values, so that the colouring is invariant under , and then examines every set of the size in question that contains , and . This is a complete check, because is 3-transitive on : every one-coloured set of at least three points has a one-coloured image that contains these three points. For the program examined sets of 5 points and sets of 6 points; for , sets of 6 points; and for , sets of 6 points. None of them is one-coloured in a forbidden colour. The script anc/check_projective.py compiles the program and runs it on the three tables.
The colouring . was written out in full, one colour for each of its 4-subsets, and anc/verify_rank.c examined every one of the sets of 5 points and of the sets of 6 points. This check uses no symmetry and no property of the group. The script anc/rebuild_psl_rule.py confirms that the function is constant on narrow classes, rebuilds all the colours from and (1), and compares the SHA-256 digest of the result with that of the colouring written out in full; so the description in Section 3 and the file that was checked are the same colouring. It also examines the sets of 5 points and the sets of 6 points that contain , and , and finds none that is one-coloured in a forbidden colour. Every set of at least four points of has an image under that contains these three points, so this is a second complete check, by another program and another method.
The colouring of 5-subsets. was written out in full from its short form, and anc/verify_rank.c examined every one of the sets of 6 points, for each colour. The search program’s own check, written independently of it and using no symmetry, agrees. The script anc/rebuild_short_form.py rebuilds the full colouring from the short form and compares its SHA-256 digest.
Consequences
The following construction is the one of Bruce, Budden and Hiller [5] that the survey [1] uses; we include the short argument for completeness.
Lemma 14. Let . Suppose that the triples of an -set can be coloured with colours so that no 5 points have all their triples in one colour. Then the triples of can be coloured with colours with the same property.
Proof. Let be the given colouring, with colours . Colour a triple
of as follows. If are distinct, gets the colour . If , then are distinct and gets the colour . Otherwise exactly two of are equal, and gets the new colour .
Let be a set of 5 points of whose triples all have the same colour . Suppose first that . If two points of have the same first coordinate and a third point has a different one, these three points form a triple of colour . So the first coordinates of the points of are either all distinct or all equal. In the first case they form 5 points of whose triples all have colour under ; in the second case the second coordinates do. Both contradict the assumption on . Suppose now that . If three points of have distinct first coordinates, their triple
::: has a colour at most . Otherwise the 5 points of have at most two distinct first coordinates, so three of them share one, and their triple has a colour at most . In both cases we have a contradiction.
Proof of Corollary 4. By Proposition 6, the triples of a 98-set can be coloured with 2 colours so that no 5 points have all their triples in one colour. Applying Lemma 14 times gives such a colouring with colours on points. For this is points.
Proof of Corollary 5. The survey [1 Section 7.1] quotes two recurrences of Song, Ye and Liu [11]. We use them only in the following ranges:
Together with the symmetry and Theorem 2, the first gives . The second, applied to , , and in turn, gives , , and . The values 661, 3961 and 194 041 that the survey lists for , and arise in the same way from . From the bound of [7] the same steps give 1581, 9481 and 464 521; these are the values marked † in Table 1.
Remarks on the search
Method. The colourings of triples were found by exact search over invariant colourings. For a chosen group , with or without a fixed point, there is one Boolean unknown for each orbit of triples. The group contains all translations, and every set of at least three points contains a point of , so it can be translated to a set that contains the point 0; hence it suffices to forbid the one-coloured sets of the relevant sizes that contain 0. The resulting formulas were given to the SAT solvers Glucose [9] and CaDiCaL [10] through the PySAT toolkit [8]. For we required in addition that exchange the colours (Remark 7), which leaves 36 unknowns; the solver answered in 14 seconds. has 74 unknowns and took six minutes.
The tables of § 3 were found in the same way, with one unknown for each class of cross-ratio values: since is 3-transitive, it suffices to forbid the one-coloured sets that contain 0, 1 and . For there is one unknown for each narrow class, and the sets forbidden are those that contain , 0 and one of two further fixed points, a square and a non-square. For the SAT solver answered at once. For and for , which has 4276 unknowns, we used a local search of the probSAT type [13] instead; it found in about steps and in about . A local search can only find a colouring; what it finds is checked as in § 5.
What the search did not find. The following are reports of a SAT solver on the encoded families, or of a local search where that is said; no proof certificate was recorded. They describe the limits of these particular families, establish no upper bound, and say nothing about other colourings.
For : on with two fixed points (99 points), the solver returned UNSAT for the -invariant family and for the -invariant family. For the larger -invariant family the search ended without an answer.
For : on with two fixed points (71 points), the solver returned UNSAT for the -invariant family.
For : on with one fixed point (38 points), the solver returned UNSAT for the -invariant family. The searches we ran on and ended without an answer.
For : on (702 points), neither the local search nor the SAT solver gave an answer in the time allowed.
For : the solver returned UNSAT for the -invariant family at every prime with , and for the -invariant family at every prime with (Appendix C).
For : on with one fixed point (84 points), both with all multipliers and with the squares as multipliers, and on with all multipliers (89 points), the local search found no colouring. A local search cannot show that none exists.
Data availability
The colourings accompany this paper as ancillary files, which are listed in Table 3.
| File | Content |
anc/R453_37_points.json, .csv | |
anc/R463_70_points.json, .csv | |
anc/R553_98_points.json, .csv | |
anc/R564_194_points_rule.json | |
anc/R564_200_points_rule.json | : the table |
anc/R564_200_points.txt.gz | , written out in full |
anc/R664_500_points_rule.json | |
anc/R664_600_points_rule.json | |
anc/R665_83_points_short.json | |
anc/verify.py | the check of Section 5, for triples |
anc/rebuild_from_rule.py | rebuilds a colouring from Appendix A |
anc/verify_crossratio.c | the check for , , |
anc/verify_projective.cpp | the second check for the same three |
anc/check_projective.py | compiles and runs it |
anc/verify_rank.c | examines every set, for a colouring written out in full |
anc/rebuild_short_form.py | writes out in full |
anc/rebuild_psl_rule.py | rebuilds from and compares |
anc/README.md | formats and usage |
Table 3. The ancillary files.
Triples. Each of , and comes in two forms: a JSON file that holds the colouring as a string of the characters 0 and 1, one for each triple of the points in lexicographic order, together with its group and orbit representatives; and a CSV file with one row for each triple, giving its three points and its colour. The point is the last point, numbered 69 in and 97 in .
4-subsets. Each of , and is a JSON file with its table as a string of characters (the places 0 and 1 are not used) and the list of classes of colour 0. The points are and , which is the last point, numbered . comes in two forms: a JSON file of the same kind, with its table and the list of narrow classes of colour 0; and a text file, compressed with gzip, with one character for each 4-subset of the points in lexicographic order, the point being numbered 199.
5-subsets. is a JSON file with the colours of the 85 320 subsets through 0 and 1, in lexicographic order.
SHA-256 digests of the colouring strings of , , , and , and of the tables of , and :
:
995fb6bb93ca2b0be5809d981cefbaa72712ce8b1463809dd66376af9f6c0e16:
f7c37ea5e0b72ad3773b13c0a291697f336998170931f208f6588352ce31c662:
53e042c9f44d870cdccd4d86b5e45003fadf981d7757580f863632a945f5e17d:
477e5506f17a2db541c806f9c57ca20c2e1d9adde4918dd01c29075d3092dbe7:
621a303b0e947fb3e52799288ba7f8ef883f3260f6a7e260c843eb902001602d:
3ed928e24dabb537c6d26c5da0251af9562e40f9be4c9a2f82b7ff55a97931a1:
c4f4b30a65c395b1d7e8fd014a6cd17a31acb70672ef7294da95a934cf4a7a6f:
b196450e13f22883d9e9c974c3f203b0a958db0560110c2cde73c5bd803daadb
Acknowledgments
The author thanks Stanisław P. Radziszowski for maintaining the dynamic survey on which this work relies. Computational searches were conducted using the author’s Solon research software. AI-based tools were used in the searches and in the preparation of this manuscript; the author is responsible for all of its content. The reported lower bounds are independently verifiable from the ancillary files.
Orbit representatives
For each colouring we list one triple from each orbit in colour 0; the triple listed is the lexicographically first of its orbit. Colour 0 is the union of the orbits of these triples under the group named, and every other triple has colour 1.
: on , 32 orbits. , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , .
: on , 27 orbits. , , , , , , , , , , , , , , , , , , , , , , , , , , .
: on , 36 orbits. , , , , , , , , , , , , , , , , , , , , , , , ,
Classes of cross-ratio values
For each colouring of Section 3 we list the classes on which takes the value 0, each by its smallest element as an integer in . On every other class is 1. For the function is and the classes are the narrow classes of Section 3.
: , 15 of 33 classes. 2, 3, 5, 6, 9, 11, 12, 13, 15, 20, 21, 22, 31, 40, 68.
: , the group , 20 of 43 narrow classes. The function takes the value 0 on the narrow classes of the following values , so that colour 0 is the union of the orbits of the 4-subsets : , 3, 7, 11, 12, 13, 15, 17, 24, 27, 29, 35, 37, 38, 41, 55, 68, 73, 81, 107.
: , 41 of 84 classes. 3, 5, 6, 9, 10, 11, 12, 14, 15, 17, 19, 20, 24, 27, 32, 34, 39, 43, 47, 56, 63, 67, 71, 72, 75, 77, 82, 92, 110, 115, 116, 117, 119, 124, 144, 168, 172, 177, 191, 223, 224.
: , 51 of 100 classes. 3, 4, 5, 10, 12, 14, 15, 16, 18, 20, 22, 23, 24, 33, 34, 35, 37, 38, 39, 42, 43, 49, 53, 55, 65, 66, 67, 69, 72, 79, 81, 84, 85, 90, 95, 103, 104, 106, 112, 115, 127, 138, 139, 143, 146, 147, 174, 197, 199, 241, 258.
Where the searches found colourings
This appendix records the primes at which our searches found colourings of that are invariant under the group named. Unless it is said otherwise, the answers are those of a SAT solver, which decided each of these small formulas in both directions. No proof certificate was recorded for the answers UNSAT. They concern only the invariant family named and establish no upper bound.
(5, 6; 4), the group . Among the 42 primes with , the solver found a colouring for , 103, 107, 109, 113, 127, 131, 137, 139, 151, 157, 163, 179, 191 and 193, and returned UNSAT for the other 27. For it reported that the colouring is unique.
(5, 6; 4), the group . Among the 30 primes with , the solver found a colouring for , 167, 179, 191, 193 and 199, and returned UNSAT for the other 24. For and the larger group admits none.
(6, 6; 4), the group . Colourings were found for every prime with , and for , 199, 251, 317, 331, 337, 347, 349, 359, 397, 449, 499 and 599; the one for was found by local search. For neither the local search nor the solver gave an answer in the time allowed.
(5, 5; 4), the group PGL(2, ). Among the primes with , the solver found a colouring for , 13, 17, 23 and 47, each time unique up to exchanging the two colours, and returned UNSAT for the others. The case is a colouring of order 48; it agrees with the bound of [7], and we make no claim for it.
References
References
- [1]S. P. Radziszowski, Small Ramsey numbers, Electron. J. Combin., Dynamic Survey DS1, revision #18 (24 April 2026), https://doi.org/10.37236/21.
- [2]B. D. McKay and S. P. Radziszowski, The first classical Ramsey number for hypergraphs is computed, Proceedings of the Second Annual ACM–SIAM Symposium on Discrete Algorithms (SODA ’91), San Francisco, 1991, 304–308.DOI
- [3]J. Dybizbański, A lower bound on the hypergraph Ramsey number R(4, 5; 3), Contrib. Discrete Math. 13 (2018), no. 2, 112–115.
- [4]J. Dybizbański, personal communication to S. P. Radziszowski (2018), cited in [1]; addendum to [3] at https://inf.ug.edu.pl/ramsey.
- [5]M. Bruce, M. Budden, and J. Hiller, Lexicographic products of r-uniform hypergraphs and some applications to hypergraph Ramsey theory, Australas. J. Combin. 70 (2018), 390–401.
- [6]H. Yang, Three improved lower bounds for hypergraph Ramsey numbers, data deposit, Zenodo, 2026, https://doi.org/10.5281/zenodo.24301648.
- [7]K. Özdemir, Lower bounds for the hypergraph Ramsey numbers R(5, 5; 4) and R(5, 6; 4) from PSL(2, q)-invariant colourings, data deposit, Zenodo, 27 October 2026, https://doi.org/10.5281/zenodo.23689461.
- [8]A. Ignatiev, A. Morgado, and J. Marques-Silva, PuSAT: A Python toolkit for prototyping with SAT oracles, in: Theory and Applications of Satisfiability Testing – SAT 2018, Lecture Notes in Comput. Sci. 10929, Springer, 2018, 428–437.DOI
- [9]G. Audemard and L. Simon, Predicting learnt clauses quality in modern SAT solvers, in: Proceedings of the 21st International Joint Conference on Artificial Intelligence (IJCAI 2009), 399–404.
- [10]A. Biere, K. Fazekas, M. Fleury, and M. Heisinger, CaDiCaL, Kissat, Paracooba, Plingeling and Treengeling entering the SAT Competition 2020, in: Proceedings of SAT Competition 2020 – Solver and Benchmark Descriptions, Department of Computer Science Report Series B, vol. B-2020-1, University of Helsinki, 2020, 51–53.
- [11]E. Song, W. Ye, and Y. Liu, New lower bounds for Ramsey number R(p, q; 4), Discrete Math. 145 (1995), 343–346.DOI
- [12]S. C. Liu, Lower bounds for small Ramsey numbers on hypergraphs, preprint, 2019, arXiv:1906.00132.arxiv.org/abs/1906.00132
- [13]A. Balint and U. Schöning, Choosing probability distributions for stochastic local search and the role of make versus break, in: Theory and Applications of Satisfiability Testing – SAT 2012, Lecture Notes in Comput. Sci. 7317, Springer, 2012, 16–29.DOI