2. Notes
1. General rule: all subfields of QE Objects should be Objects (e.g. use ASListStructure.List, and not LIST OF Object); this will facilitate offlining and onlining of these subfields.
2. Basic task for software: determine the different types of Objects to be manipulated; basically, each one is automatically offlined when created (but also left online).
3. In general, the software is able to guarantee that it only works with objects from the database for one qe problem at a time.
4. At top level, there is a computation manager who determines what tasks can be performed at the present time, gets them distributed out over Dorados (via ComputeServer) or Suns (?), and scans (or is notified of events in) the LB DB to determine when new tasks can be scheduled.
Should be able to give its state (e.g. "waiting for completion of 2-space cells") upon demand.
3. Cad implementation
3.1. New operations
Request for any AlgebraicSurfaces display op starts up the tool if not already started. Or have it running all the time, and each display request creates a new surfaceDisplay viewer.
AlgebraClasses has proc to add an op to a structures property list. Or MakeSequenceStructure proc takes procs to add to prop list as args.
Cad ops
induced cad
Alternate display
showTriangulationAll (valid only for cad's of 1-, 2- and 3-space)
showTriangulationSections
showTriangulation0Cells
showTriangulation1Cells
showTriangulation2Cells
rayTraceInputPolynomials (valid only for cad's of 3-space)
show cells
show statistics
Cad's should have names, statistics
Store only basis signatures with cells. Both clustering during cad construction, and clustering for disambiguation of signatures, are done with basis signatures (if basis signatures with ordinary projections don't suffice to get unambiguous clusters, then we have to go back and refine cad with respect to augmented projections).
Cluster ops
select containing cad
rayTrace (valid only for 2-cluster in 3-space)
showTriangulation (valid only for clusters in cad's of 1-, 2- and 3-space)
Cell ops
select containing cad
rayTrace (valid only for 2-section in 3-space)
showTriangulation (valid only for cells in cad's of 1-, 2- and 3-space)
Polynomial sequence ops, for base coeffs that are ints or bigrats
rayTraceZeros (valid only 3-variate polys)
decomposeZeros (forks Vax computation; pop up menu for options (including covering set parameters); inserts cad in Active MessageSet)
showZerosTriangulation (valid only for 1-, 2- and 3-variate polys)
New Polynomial ops, for base coeffs that are ints or bigrats
rayTrace (valid only for 3-variate poly)
decomposeZeros (forks Vax computation; pop up menu for options (including covering set parameters); inserts cad in Active MessageSet)
showZeroTriangulation (valid only for 1-, 2- and 3-variate poly)
Sequences
Basic display is one line: "sequence of n shortPNames"
Alternate display: Slider bar onto popup menu
popup menu entries are shortPNames of sequence elements
Insert creates a blank entry in popup menu; then select item to be inserted as an operand, it is type checked, inserted.
Delete modifies pop up menu
Selection of Insert or Delete forces Alternate display
3.2. Adjacencies and Clusters, Graph ops, read adjacencies, compute clusters
We have made a basic decision to go with basis signatures.
First step is the current "cells with signature" op. This assumes that every cell in the cad has a basis signature, then this is a binary op that takes a legal (i.e. correct length) basis signature (perhaps with * elements) and returns a sequence of cells with that sig.
Subr for boundary of 2-sector (i, j) in E2
I3=lelt(I,3); I2=lelt(I,2);
Pass I#, I3, and I2 to a routine
For each 2-sector (i, j, k) in I#
Subr for boundary of 2-sector (i, j) in E2:
L1 ← U1 ← 0;
L1L ← L1R ← U1L ← U1R ← 0;
for each 1-cell (a,b) adjacent to 2-sector (i, j)
if a = i then if b = j-1 then L1 ← j-1 else if b = j+1 then U1 ← j+1 else ERROR;
if L1#0 then for each 0-cell (a,b) adjacent to 1-section (i, L1)
if a = i-1 then L1L ← b else if a = i+1 then L1R ← b else ERROR;
if U1#0 then for each 0-cell (a,b) adjacent to 1-section (i, U1)
if a = i-1 then U1L ← b else if a = i+1 then U1R ← b else ERROR;
Return[ LIST(L1, LIST(L1L, L1R), U1, LIST(U1L, U1R) ) ].
ExtractClusters:
PROC [cad: Cad, numComponents:
CARDINAL]
RETURNS [clusters:
SEQ.Sequence];
components assumed to be of cad, and not some induced cad thereof.
Components: Menus.MenuProc = {
User specifies polynomials with respect to which to cluster by (*,*,@,*,*,@) tuples (@ is new FormulaOperators.Op).
edgePredicate: GraphOps.EdgePredicate ~ {
cell1: QET.Cell ← NARROW[v.rep];
cell2: QET.Cell ← NARROW[w.rep];
RETURN[cell1.basisSignature.structure.class.equal[cell1.basisSignature, cell2.basisSignature] ];
};
inducedCad: QET.Cad ← QEIO.GetInducedCad[cad, dimension];
numComponents: CARDINAL ← GraphOps.DetermineComponents[cad.abstractGraph, Undirected, edgePredicate];
clusters: QET.SignedRegionSeq ← Cads.ExtractClusters[inducedCad, numComponents];
QEIO.WriteSignedRegionSeq[clusters, out];
};
MaximalSignedRegions: PROC [dimension: CARDINAL, cad: Cad, primary: BOOL];
Depth-First Search connected components computation in the dimension-space cad graph, filling in the (IF primary THEN primaryRegion ELSE secondaryRegion) fields of cells.
ExtractSignedRegions: PROC [dimension: CARDINAL, cad: Cad, primary: BOOL] RETURNS [SEQ.Sequence];
Extract from the dimension-space cad graph, the (IF primary THEN primaryRegion ELSE secondaryRegion) cell-field-defined regions.
AdjacenciesOverAdjacency:
PROC [dimension:
CARDINAL, adjacency: Adjacency, cad: Cad]
RETURNS [newAdjacencies:
LIST
OF Adjacency];
For i = 2 or 3 only, returns all interstack adjacencies in i-space over an adjacency in dimension-1 space. cad is i-space cad. newAdjacencies = NIL if no adjacencies in dimension space can be built, either because needed auxiliary E^(i-1) cell-cell adjacencies, or non-NIL sample points, cannot be found.
Converts extended to primitive sample points as needed.
May build adjacencies over appropriate auxiliary E^(i-1) cell-cell adjacencies if necessary, and/or query i-space cad for existing adjacency information.
3.3. Cad statistics records
SmartInfo about the cad (0-cell bounding box, singularities, turning points).
3.4. New cell Ops: read covering set, display covering set
3.5. AlgebraStructures Rootfinder, new cell op: generate covering set
3.6. Compute and add adjacencies to an existing cad in Cedar
Do we want to separate adjacency computation from cad computation? If so, set up a SAC-2 routine to do it; allow ComputeServer to queue up a succession of jobs, to be done in background.
Step two: implement the computation in Cedar. 2D case first.
3.7. Cad databases
Reading such files is slow. Need command to read some more cells of this cad and add them to the database. Cypress database?
Once have distributed computations, we want to be able to do incremental updates of the database.
3.8. RefineCad operation (Real algebraic variety Combiner)
This is a great example for implementing cad computation related stuff in Cedar by forking off appropriate subcommands to remote SAC-2 systems.
RefineCad:
PROC [inCad: Cad, refiningPoly: Polynomial, cellsToRefine: Sequence]
RETURNS [outCad: Cad];
cellsToRefine is a set of cell indices. outCad is a refinement of inCad that is coarsest possible with respect to the property that each of the cells in cellsToRefine has been decomposed into refiningPoly-invariant cells.
Implemented by rerunning the whole cad construction algorithm. Consider 1-space. Assume we have saved the isolated roots of the univariate basis that define the cells, and assume that there is only one cell c to refine, and suppose that the 1-space projection of c is a 1-cell, say the interval (a,b). We compute new projections all the way down by combining refiningPoly with existing polys. When we get to 1-space, we probably have some univariate polynomials that we didn't have before. We ask if any of these new polynomials have roots in (a,b). If so, then we break up (a,b) into new cells, and rebuild the cad from there. Finally we assign new cell indices where necessary.
3.9. Projection Phase
CoarsestSqFreeBasis:
PROC [inputPolynomials:
SEQ.Sequence, localizationFormula:
QFF.Formula ←
NIL]
RETURNS [contents, basis:
SEQ.Sequence];
Projection:
PROC [basis:
SEQ.Sequence]
RETURNS [basisProjection:
SEQ.Sequence];
The requirements are:
(1) we have a certain set K of polynomials in E^r which we need to make delineable on each cell and each signature-cluster in E^(r-1).
(2) The collection of cells we end up with in E^r must have the boundary property: the boundary of any cell must equal the union of some collection of lower-dimensional cells.
Question: why does the boundary property really matter?
Possible answer: when you have the boundary property, it doesn't matter where you probe when slicing an adjacency. To be more precise, the boundary property is required for the "decoding" phase of our (0,1) adjacency determination in 3-space to work - we have to know that the 0-cells we see in our 2-space projection are images of 0-cells in the 3-space decomposition.
Given these, our approach is:
(1) Compute contents and basis(pp) = K of inputs, pull off initial sequences of non-constant coeffs of all basis elts, compute basis for contents union initial sequences, do a signature-invariant decomp D of E^(r-1) to get regions on which each element of K has constant degree. Note that if any element of K is identically zero at some point of E^r-1, D will have maximal regions for this property.
(2) using the appropriate reducta of elements of K (ignore identically zero elements), compute projection sets for each reducta pattern that occurs on a positive-dimensional regions, compute basis. For each positive-dimensional region, use this basis to break it up to be signature-invariant with respect to its projection basis. We end up with a new decomposition D'.
(3) Assume now r <= 3; maybe works for general r.
(4) On each region in E^(r-1), evaluate the appropriate non-identically zero reducta of elements of K to get a set of algebraic polys; compute its basis. If any element of K is
identically zero on this region, determine the appropriate algebraic polynomials whose roots we need to take account of to insure boundary property, compute their basis, merge the two bases. Isolate roots of the final basis to determine the stack over this region.
Simpler version, to get a single basis set in each E^i, which we want in order to be able to say very simply what clusters we have (namely: invariant with respect to the bases in all dimensions):
(1) Compute P1 = contents and K = basis(pp) of inputs.
Compute P2 = initial sequences of non-constant coeffs of all basis elts,
Compute each Pijk, where suppose K has three elements, 0 <= i < length of initial sequences of non-constant coeffs of K1, 0 <= j < etc., Qijk is projection computed using ith, jth, and kth reducta of respective elements of K, and Pijk is the result of simplifying Qijk with respect to the side relations that the appropriate initial non-constant coeffs of the elements of K are equal to zero.
(Possible precomputation for each Pijk: do a heuristic test (e.g. Groebner basis to check emptiness of complex variety), or cad, to check that real varieties stay nonempty as you take progressively larger sequences of initial non-constant coeffs; if such a variety is empty, then you don't need to continue. It could happen that appropriate tails of initial sequences of elements of K could be thrown out of P2).
Set P = P1 un P2 un (all Pijk's).
Compute P-invariant decomposition of E^(r-1).
(2) Assume now r <= 3; maybe works for general r.
(3) Extension phase: On each region in E^(r-1), evaluate the appropriate non-identically zero reducta of elements of K to get a set of algebraic polys; compute its basis. If any element of K is identically zero on this region, determine the appropriate algebraic polynomials whose roots we need to take account of to insure boundary property, compute their basis, merge the two bases. Isolate roots of the final basis to determine the stack over this region. Record K-signature of each element of the stack.
3.10. Base Phase
BasePhase:
PROC [cad: Cad];
cad is of r>=1 space; follow chain of induced cad's down to 1-space, contruct cells there. Fill in basisSignature's = primarySignature's of cells. If 1-space cad.localizationFormula # NIL, then discard cells which don't satisfy it.
3.11. Extension Phase
Information to save to pass to adjacency algorithms:
E1 to E2: For each 0-cell in E1: isolating intervals (perhaps expanded to be strictly open) and corresponding algebraic basis elements over it; for each element F of integral bivariate basis, a sequence of length m of positive integers, such that F has m real roots over it, and ith element of sequence is section (or element) of overall stack which is ith real root of F.
Use this list to modify cell indices of section-section adjacencies produced by adjacency algorithms.
E2 to E3: For each 0-cell and 1-cell in E2: isolating intervals (perhaps expanded to be strictly open) and corresponding algebraic basis elements over it; for each element F of integral trivariate basis, a sequence of length m of positive integers, such that F has m real roots over it, and ith element of sequence is section of overall stack which is ith real root of F.
RegionExtendCadToCad:
PROC [cad: Cad];
Extend cad of dimension-1 space to dimension space, using the largest signed regions (wrt primarySignatures) that can be constructed in the graph for dimension-1 space.
Let i = dimension.
Assume that in E^(i-1), with primaryPolynomials ← CompleteFactors(inputPolynomials), primarySignatures of cells are set.
1. MaximalSignedRegions[i-1, cad, TRUE]; regions ← ExtractSignedRegions[i-1, cad, TRUE]; sort regions in order of decreasing dimension (for i-1 = 2, put degenerate 0-dimensional regions after nondegenerate).
2. cad.cells ← NIL; Go down sorted list of regions, for each:
2a. currentCells ← ExtendRegionToCells[i, currentRegion, cad]; newAdjacencies ← NIL;
If currentRegion is a 0-cell, then as each element of (trivariate) basis is evaluated at sample point, check for identically zero; if so, then add that basis element to identicallyZeroBasisElements field of the 0-cell. If identicallyZeroBasisElements#NIL when all basis elements have been evaluated, i.e. if this is a degenerate 0-cell, then do the extra work required to determine the elements of the stack over this cell (RES sets, etc.), otherwise just isolate the roots of the algebraic basis of evaluated (trivariate integral) basis.
2b. Do this step only if i = 2 or 3. If dimension of currentRegion < i-1, then determine adjacent regions of larger dimension. Sort these into order of increasing dimension. Go down this list, and for each region on it, find all cell-cell adjacencies between currentRegion and this region. Sort these adjacencies into decreasing sum of cell dimensions. For the next such adjacency, newAdjacencies ← AdjacenciesOverAdjacency[i, adjacency, cad]. If newAdjacencies = NIL, i.e. if needed auxiliary E^(i-1) cell-cell adjacencies, or non-NIL sample points, cannot be found, then go to the next cell-cell adjacency between current region and this region. If needed auxiliary adjacencies and/or non-NIL sample points cannot be found for any cell-cell adjacency between current region and this region, give up on building adjacencies in E^i between stacks over current region and this region.
Note that there is no need to worry about whether regions have the boundary property; Note the specification that we try to find an adjacency (that we can build adjacencies over) between each pair of adjacent regions; the lower dimensional cell in such an adjacency necessarily belongs to the portion of the lower dimensional region contained in the boundary of the larger dimensional region.
2c. Insert cells of currentCells into cad.cells in lexicographical order. IF newAdjacencies # NIL, then enter its adjacencies into the appropriate cells of cad.cells.
3. If i-space cad.localizationFormula # NIL, then delete all cells of cad.cells which don't satisfy it, and all adjacencies involving those cells .
4. For i = 1, 2 or 3, if no localization, we now have a graph for the cad of E^i in which a maximal signed regions computation will yield (signed) components (in 1 and 2-space, we can have all adjacencies, in 3-space, probably only a proper subset of all adjacencies.)
ExtendRegionToCells:
PROC [dimension:
CARDINAL, region: SignedRegion, cad: Cad]
RETURNS [newCells:
SEQ.Sequence];
Extend region in dimension-1 space to cells in dimension-space. cad is dimension-space cad.
Fill in basisSignatures, and, if desired, (with primaryPolynomials ← CompleteFactors(inputPolynomials)), primarySignatures, of newCells. Also fill in induced adjacencies, and intrastack adjacencies, among newCells.
(Let i = dimension). Method: Evaluate polys at base sample point, irreducible basis of algebraic polys, isolate roots, construct extended sample points, eval gsfd of i-variate polynomial at endpoints of isolating interval, or i-variate polynomial itself at easy primitive sector sample point, when we want its sign.
Stack: TYPE = REF StackRec;
StackRec: TYPE = RECORD [
base: Cell, -- in E^(i-1)
baseTruthValue: BOOL, -- used in quantifier elimination
elements: CellSeq -- of cells in E^i
];
ExtendCellToStack:
PROC [cell: Cells.Cell, cad: Cad]
RETURNS [StacksInEi:
SEQ.Sequence];
Return a StackSeq of length one consisting of the stack over cell in E^(i-1).
ExtendStacksToStacks:
PROC [i:
CARDINAL, StacksInEiM1:
SEQ.Sequence, cad: Cad]
RETURNS [StacksInEi:
SEQ.Sequence];
StacksInEi returned in lexicographical order by base
RatApproxSamplePoint: PROC [cell: Cells.Cell, ratApproxBound: BR.BigRat] RETURNS [PTS.Point];
Return a rational approximation to cell sample point, each coordinate accurate to within ratApproxBound. Should have the property that always produces the same RatPoint for given RealFieldPoint and ratApproxBound inputs.
3.12. QE phase
QEProblem: TYPE = RECORD [
inputFormula: QFF.Formula,
localizationFormula: QFF.Formula,
equivalentQFFormula: QFF.Formula ← NIL,
freeSpaceSignatureTable: SignatureTable ← NIL,
cad: Cad ← NIL -- constructed up through k-space, projection data to r-space
statistics: QEStatisticsRec, -- to be added later
];
MakeQEProblem:
PROC [inputFormula:
QFF.Formula, localizationFormula:
QFF.QFFormula, inputVariables:
VARS.VariableSeq, minPolyVariable:
VARS.VariableSeq ←
AN.defaultMinPolyVariable, fieldElementVariable:
VARS.VariableSeq ←
RF.defaultFieldElementVariable]
RETURNS [
QET.QEProblem];
EliminateQuantifiers:
PROC [
QET.QEProblem];
1. Extract polynomials of formula and localizationFormula; cad ← ProjectionPhase[polynomials, inputVariables, localizationFormula].
2. Cad.BasePhase[cad];
3. For i in [2,...,k] do Cad.RegionExtendCadToCad[i, cad].
4. Cad.MaximalSignedRegions[k, cad, TRUE]; regions ← Cad.ExtractSignedRegions[k, cad, TRUE]; Sort regions into order of decreasing dimension (for k = 3, put degenerate 0-dimensional regions after nondegenerate), since qe for larger dimensional regions should go faster, since they are likely to have lower degree sample points.
5. For each region of regions, using representative cell = cellInEk (which has extended or primitive sample point), do:
5.1 StacksInEk+1 ← Cad.ExtendCellToStacks[k+1, cellInEk, cad];
5.2 For i in [k+2,...,r] do StacksInEi ← Cad.ExtendStacksToStacks[i, StacksInEiM1, cad].
5.3. EliminateInnermostQuantifier[StacksInEr, formula];
5.4. For i decreasing in [k+1,...,r-1] do EliminateInnerQuantifier[i, StacksInEi, StacksInEiM1, formula];
5.5 If StacksInEk+1.baseTruthValue then insert this signed component into SignatureTable.
6. DisambiguateSignatureTable, extract signatures corresponding to regions in solution set, bundle them up as a SignatureSeq
7. Apply SimplifySignatureSeq; Return SignatureSeqToFormula and the final SignatureTable.
EliminateInnermostQuantifier:
PROC [StacksInEr:
QET.StackSeq, formula:
QFF.Formula];
Go through StacksInEr determing truth value of formula on each element of each stack, looking at innermost quantifier, and setting baseTruthValue accordingly.
EliminateInnerQuantifier:
PROC [i:
CARDINAL, StacksInEi, StacksInEiM1:
QET.StackSeq, formula:
QFF.Formula];
Assume baseTruthValue's of StacksInEi already set, and for each stack S in StacksInEiM1, find the stacks in StacksInEi whose bases are the elements of S. Looking at the baseTruthValue's of these stacks in StacksInEi, and the appropriate quantifier in formula, set the baseTruthValue of S.
3.13. Simplifier
The basic strategy is to assume cells in r-space invariant with respect to bases in all dimensions. No effort to separate complete factors of input polynomials from factors of projection polynomials. If we have unambiguous signatures at this point, i.e. if for each signature, either every cluster with that signature is in solution set, or no cluster with that signature is in solution set, then we just have to simplify. If not, we have to refine the cad with respect to augmented projections and get clusters invariant with respect to these new larger bases in each dimension. Having done this we know we'll get unambiguous clusters.
SignatureTable: TYPE = REF SignatureTableRec;
SignatureTableBoundsType: TYPE = [1..65534];
SignatureTableRec: TYPE = RECORD [
cad: Cad, -- access to the cad to which the cells of entry regions belong
polynomials: POL.PolynomialSeq, -- irreducible (<= r)-variate rational polynomials, all of which occur in the bases of some (possibly 0th) induced cad (and hence are invariant on all cells of the r-space cad), with respect to whose signs the signatures of the entry regions are defined.
entries: SEQUENCE lengthPlus1: SignatureTableBoundsType OF SignatureTableEntry
];
SignatureTableEntry: TYPE = REF SignatureTableEntryRec;
SignatureTableEntryRec: TYPE = RECORD[
signature: Signature,
unambiguous: BOOL,
regions: SignedRegionSeq -- each of which has the signature of this TableEntry
];
DisambiguateSignatureTable:
PROC [inTable:
QET.SignatureTable]
RETURNS [outTable:
QET.SignatureTable];
Assume that initially inTable.polynomials ← inTable.cad.primaryPolynomials.
outTable ← copy(inTable).
While there is an ambiguous entry in outTable, do for each such ambiguous entry:
1. Select some polynomial P from the basis of some (proper) induced cad, such that P is not an element of outTable.polynomials (use hashing).
2. Augment the secondarySignature's of all cells of all regions of this entry by sign(P).
3. For each region of this entry, form the subgraph it induces, and construct signed components in this subgraph wrt secondarySignature.
4. Collect the cumulative data: if the sign(s) of P on all newly formed components which are contained in the solution set differ from the sign(s) of P on all newly formed components which are external to the solution set, then we have succeeded in disambiguating this entry with P.
5. Delete this entry, and create two new appropriate unambiguous entries, with appropriately augmented signatures.
6. For all unambiguous entries, augment their signatures with "don't care" in the P position, and augment appropriately (i.e. with the correct P sign) the secondarySignature's of all constituent cells of all entry regions.
7. For each ambiguous entry, carry out steps 2,3,4, and see if we have succeeding in disambiguating that entry with P. If so, then carry out step 5 for this entry. If not, then split this entry into up to three or more (each of which may be ambiguous or unambiguous, but at least one of which must be ambiguous if we are here) entries, corresponding to splitting each of the prior entry regions into up to three or more new regions by the sign of P.
SimplifySignatureSeq:
PROC [
QET.SignatureSeq]
RETURNS [
QET.SignatureSeq];
First determine signatures which do not occur, and use them when helpful.
Sort into lexicographical order and combine.
SignatureSeqToFormula:
PROC [
QET.SignatureSeq]
RETURNS [
QFF.QFFormula];
Convert to a formula.
3.14. CadCoodTransform
Actually is an op on a set of polys so that we end up with a decomposition which is cylindrical with respect to two coordinate systems.
3.15. Application of Cedar cad algorithm
1. Demonstrate topologically reliable display of algebraic and parametric curves and surfaces, and display of fat plane and space curves.
2. Application of distributed cad algorithm to qe problems involving roots of quartic polynomials (full Kahan ellipse, root configuration-coefficient classification for quartics)
3. Study of parametric curves and surfaces: spline analyzer, spline combiner.
3.16. Expert system for directing problem solving by cad computation
Has the knowledge base of info about root configuration from evaluation of coefficients.
Tries heuristics, e.g. detecting common zeros by Grobner basis computation.
Knows what computations to do to build a sufficient database for solving the qe or decision problem.