Research notes on every problem, and a library of the papers behind them. Built from the open erdos repository.
Updated
The rational-rank defect at most two case
Statement
Let be a finite nonempty set in a finite-dimensional rational vector space, containing neither zero nor an antipodal pair. Let , and let be its maximum dissociated-subset size. Then
The cases were proved in [[research/erdos_963/corank_one|the preceding note]]. The proof below treats . Together with the intrinsic lift in that note, it establishes the target for positive real sets whose intrinsic ternary-relation rank satisfies .
This is a proved special case, not a solution of the unrestricted problem.
Maximum-volume coordinates
Choose a maximum-absolute-determinant rational basis , and use it as coordinate basis. All coordinate columns belong to , and every -column determinant has absolute value at most one. More generally, every square minor of coordinate columns has absolute value at most one: extend that minor to an -column determinant using the missing coordinate columns of . When the minor is nonzero, no repeated column is introduced.
The integral-coordinate columns satisfy
For completeness, such columns are ternary vectors. Distinct ones are not proportional, since nonzero proportional ternary vectors are equal or opposite. If two reduced to the same vector modulo two, extend them to a basis using ; its coordinate determinant would be a nonzero even integer, contradicting the maximum-volume bound. Their distinct reductions are therefore among the nonzero binary vectors.
We will prove that the nonintegral columns satisfy
Together with (2), this gives .
Individual sign changes preserve dissociation and the absolute minor bounds. They never identify columns because has no antipodal pair. We may therefore orient each nonintegral column to a specified representative of its signed residue class in .
Two elementary coordinate counts
Consider an oriented nonzero residue class . Suppose it has nonzero fractional coordinates and zero coordinates. In a fractional coordinate, a column has at most two possible values, differing by one. In the other coordinates its values belong to .
For a fixed support in these integral positions, at most two integer sign patterns occur, and if two occur they are opposites. Indeed, two patterns on the same support that are neither equal nor opposite agree at one nonzero coordinate and disagree at another. The corresponding -by- minor has absolute value two, which is impossible. The empty support has just one pattern. Thus any one oriented residue class has at most
columns. This conclusion does not use a dissociation hypothesis.
There is a better bound if every pair of its columns satisfies
If , only the difference can be integral. Every coordinate then has at most two consecutive values, giving at most columns. If , orient the first half-integral coordinate to . There are only fractional patterns, so the minor argument just given bounds the number of columns by . In either case, (5) implies
We also use the following box count. If four nonempty sets each have at most two consecutive values in each of coordinates, and at most two of the four sets vary at any coordinate, their total cardinality is at most
To prove it, let be the number of coordinates varying in set . Its size is at most , while and . Convexity of , or transferring one unit from a smaller positive to a larger one below , shows that the sum is maximized at . This proves (7).
Residue configurations of dissociation dimension at most two
Let be the residues of . A dissociated triple of residues would, together with , give dissociated columns: reduce a putative relation modulo , and then use the rational independence of . Hence
Let be the number of nonzero signed residue classes occurring. If , choose dissociated representatives of two distinct such classes. By (8), every other residue belongs to the ternary span of . Consequently
The following observation forces (5) for many classes. If a residue class has an occurring residue
choose any column of residue . For distinct columns oriented to residue , the set has elements and has a ternary relation. If the coefficient of were nonzero, reduction modulo would contradict (10). Thus the relation omits and must use both ; it gives (5). Such a class therefore obeys (6).
Zero, one, two, or three signed classes
If , there are no nonintegral columns. If , (4) is already stronger than (3).
Suppose , with representatives . If at least one of or holds, that choice supplies (10) for one of the two classes. Bound this class by (6) and the other by (4), proving (3).
The remaining two-class case has and . After orientation take . Then either or . The first would identify the two signed classes. Thus has order five, and the classes are precisely the two nonzero signed classes of . This exceptional case is handled below.
If , every class has another occurring class outside . All three therefore obey (6), proving (3).
The four-class case: an elementary abelian group of order nine
Suppose , so all four classes in (9) are distinct. The triple is dependent. Its relation must use all three residues, since their signed classes differ. If the coefficients of and are equal, the relation implies (the alternative is excluded). If those coefficients are opposite, it implies , which identifies two of the four signed classes. Therefore . Similarly . As are dissociated, they generate a copy of , and the four occurring signed classes are its four one-dimensional subspaces.
Every class obeys (10), so, after fixing orientations, it has at most two consecutive coordinate values. Any three distinct classes have a unique ternary relation up to sign on their three representatives. For arbitrary columns from these classes, the dependence of must use all three and has exactly those residue signs. Hence their correspondingly signed sum belongs to .
At a fixed coordinate, let the width of a class be zero or one according as it is constant or varying. Since the three columns can be chosen independently, the width of a signed sum from three classes is the sum of their three widths. The preceding ternary bound makes this sum at most two. Every three of the four widths therefore have sum at most two, so at most two classes vary at that coordinate. Apply (7) with to obtain , proving (3).
The two-class group of order five
Orient the two classes to and , where has order five; call their column sets and . They have the same fractional-coordinate positions and the same integral-coordinate positions. Fix a support in the latter. Within each of and , at most two integer sign patterns occur, by the minor bound preceding (4).
If at most three pattern groups occur across the two classes, their total size is at most . Otherwise there are four nonempty groups , with different integer patterns inside each class. Any and disagree in an integral coordinate of their common support, so is not ternary; their sum is not integral. Thus they cannot have a ternary relation with alone. For any , the dependence of must use all three columns. Its unique residue signs give
Similarly, for any , , and , the unique residue relation gives
In a fractional coordinate, each pattern group has width zero or one. Equations (11)--(12), with the columns chosen independently, say that every three of the four widths have sum at most two. Thus at most two groups vary in each of the fractional coordinates. Their integral coordinates are fixed within each group. Applying (7) with bounds their total size by .
There are choices of . Summing the bound gives
as required. This completes the proof of (1).
What this leaves open
The intrinsic lift preserves positivity, distinctness, and all ternary relations, and its rank satisfies . The present argument and the defect-one argument settle all configurations with . No argument for arbitrary intrinsic defect at least three is proved here.
The odd-slice strengthening
The following strengthening is also proved. Under the hypotheses of (1), suppose that has a homomorphism taking the value one on every column. Then
The intrinsic lift of a positive odd integer set retains this homomorphism, as shown in the defect-one note. Thus the odd-slice bound also holds whenever its intrinsic rank is at least .
Again only remains to prove. The integral columns have odd-weight reductions modulo two, so . We show
For columns in the same oriented residue class, is integral and . Consequently a box with varying consecutive-coordinate choices contains at most such columns. A constant box contains at most one. Write
This is a convex sequence. The same transfer argument as for (7) gives
A general nonzero signed residue class has at most columns in the present odd setting. If its residue is not of order two, fix an integral sign pattern in the argument for (4). Parity halves the fractional choices, giving at most columns in total. For an order-two residue, the number of half-integral coordinates is even: apply to . Hence . Orient its first half-integral coordinate positively. Parity halves the remaining fractional choices, giving fewer than columns. Finally, a class satisfying (5) has at most columns, either by the checkerboard restriction on its width-one box or by this order-two count.
These estimates prove (14) in every ordinary case of the preceding proof: one class uses the general bound; two nonexceptional classes use one general and one (5) bound; three classes all satisfy (5).
In the four-class case, at most two fibers vary in each coordinate, exactly as before. Equation (15), with , bounds their total size by . Here , since the subgroup of elements killed by three in has dimension over . Therefore , proving (14) in this case.
It remains to refine the order-five argument, retaining its notation. Fix an integral support . Each fixed-pattern group has at most columns by parity. Three or fewer groups therefore contribute at most . If four groups occur and , their width constraints and (15) give at most columns.
Suppose instead that four groups occur and . Necessarily . In each residue class the two integral sign patterns are opposites, by the minor argument. The difference of their integral-coordinate sums is even. Since the only fractional coordinate has two consecutive possibilities and the total coordinate-sum difference must be even, this coordinate is the same for both groups in each class. Denote these fixed values by for the residue and for the residue . Equations (11)--(12) give
The total coordinate sums of the integral vectors in (11) and (12) have parity one, since their -values are respectively and . Their integral-position contributions both have parity . Thus have the same parity . If both are even, they are zero, forcing , contrary to the nonzero fractional residue. Otherwise , and
One of is . The two opposite integral patterns in that class, together with its fixed fractional coordinate, give a -by- minor of absolute value , a contradiction. Hence four groups cannot occur when .
In every case the columns for one support contribute at most . Summing over its possible supports proves (14). Adding proves (13).