Relative squared distances to a conic

International Journal of Computer Graphics & Animation (IJCGA) Vol.5, No.1, January 2015
DOI: 10.5121/ijcga.2015.5102 17
LEFT DETb b=
DETb DETb
PPinside F DETb b b= ⊕
LEFT CCW DETb b b= ⊕
LEFT DETb b=
RELATIVE SQUARED DISTANCES TO A CONIC
Valere Huypens, Belgium
Abstract
The midpoint method or technique is a “measurement” and as each measurement it has a tolerance, but
worst of all it can be invalid, called Out-of-Control or OoC. The core of all midpoint methods is the accu-
rate measurement of the difference of the squared distances of two points to the “polar” of their midpoint
with respect to the conic. When this measurement is valid, it also measures the difference of the squared
distances of these points to the conic, although it may be inaccurate, called Out-of-Accuracy or OoA. The
primary condition is the necessary and sufficient condition that a measurement is valid. It is comletely
new and it can be checked ultra fast and before the actual measurement starts. .
Modeling an incremental algorithm, shows that the curve must be subdivided into “piecewise monotonic”
sections, the start point must be optimal, and it explains that the 2D-incremental method can find, locally,
the global Least Square Distance. Locally means that there are at most three candidate points for a given
monotonic direction; therefore the 2D-midpoint method has, locally, at most three measurements.
When all the possible measurements are invalid, the midpoint method cannot be applied, and in that case
the ultra fast “OoC-rule” selects the candidate point. This guarantees, for the first time, a 100% stable,
ultra-fast, berserkless midpoint algorithm, which can be easily transformed to hardware. The new algo-
rithm is on average (26.5±5)% faster than Mathematica, using the same resolution and tested using 42
different conics. Both programs are completely written in Mathematica and only ContourPlot[] has been
replaced with a module to generate the grid-points, drawn with Mathematica’s
Graphics[Line{gridpoints}] function. .
Index Terms .
Midpoint method, two-point method, incremental curve algorithms, squared Euclidean distance, Mathe-
matica, conic, QSIC, generation of CNC-grid points, Bresenham .
1. POINT LATTICE — DIRECTED POLAR — PROPERTIES OF CONICS (FIG.1., FIG.2.)
bx = 0, by = 1)
FC-FB = -2 SLxy ( |YM| + |XM| )
PD(xA + Sx∆, yA + Sy∆)
Sx = -1, Sy = +1)
PB PB
PC
PC
PA
-
-
-
-
+
+
+
+
PD(xA + Sx∆, yA + Sy∆) PD(xA + Sx∆, yA + Sy∆)
PD(xA + Sx∆, yA + Sy∆)
Sx = -1, Sy = -1) Sx = +1, Sy = -1)
Sx = +1, Sy = +1)
T1T2
T3 T4
When bLEFT = 1
bx = 1, by = 1)
bx = 0, by = 0) bx = 1, by = 0)
SLxy = +1
SLxy = +1 SLxy = -1
SLxy = -1
Fig. 1. The monotonic vectors Ti Fig. 2. The extreme vectors Ei

International Journal of Computer Graphics & Animation (IJCGA) Vol.5, No.1, January 2015 5
18
This paper only considers general conics defined in a point lattice, called a grid. The unit cells
are squares and the minimal distance between the grid points is a rational number ∆ . In practi-
cal algorithms, this grid distance equals one, but using ∆ clarifies and generalizes the midpoint
algorithm.
We use bold characters for vectors. The three vectors i , j, and k , codirectional with x, y, and
z, are the mutual orthogonal unit vectors of a right-handed Cartesian coordinate system.
Notation
The capital letter “S” symbolizes a sign function, and the miniscule letter “b” stands for a
Boolean function. FR ImplicitRegion[F[x,y]==0,{x,y}]≅ (17). The conditional expression “re-
sult = IF( condition, value1, value2)”, means that if the condition is true, then the result equals
value1, else the result equals value2.
We define the Boolean-function INDEXb as INDEXb ≅ IF( INDEX > 0, 1, 0 ). So, “1” means
“True” and “0” means “False”, compatible with Mathematica’s definitions: 1==Boole[True]
and 0==Boole[False].
Most of the time, the value INDEX==0, is prefiltered out, but sometimes, to indicate what we
mean, we write INDEX 0b IF( INDEX 0, 1, 0 )≥ ≥≅ .
The S-function INDEXS has two values 1± , because the value 0 will be filtered out; therefore we
define the S-function INDEXS as INDEX INDEXS IF( b 1, 1, 1 )== + −≅ .
The logical operator ⊕ means xor. As logical negation, we use the bar above the Boolean var-
iable, hence INDEX INDEX INDEXb b Not[b ]¬≅ ≅ .
The Candidate Points — the Monotonic Direction — the Boolean bLxy = bx by bLEFT
The real extreme tangent points and / or the intersection points with a bounding frame are pre-
calculated, and rounded to the nearest grid point. These points segmentize the conic in mono-
tonic segments, organized clockwise or counterclockwise. The start and endpoints of each i-
segment, define the monotonic direction i E S E S E S(x x ) (y y )− = − + −E P P i j≅ . The conditions
E S(x x 0)− == and E S(y y 0)− == will be filtered out, such that the monotonic direction x y(S ,S )
can be defined as,
x E Sb IF(x x ) 0,1,0)− >≅ (1), y E Sb IF(y y ) 0,1,0)− >≅ (2), or
x E SS IF(x x ) 0,1, 1)− > −≅ (1), y E SS IF(y y ) 0,1, 1)− > −≅ (2).
Fig. 1 shows four cells with the monotonic vectors T1, T2, T3 and T4. The general monotonic
vector equals i x yS S∆ + ∆T i j≅ (3), with monotonic direction x y(S ,S ) . Point AP is the actual,
optimal, best selected grid point (Section 4.), and the candidate points, concurring with the
monotonic vector iT , are the 4-connected grid points BP , CP or the 8-connected grid points
BP , CP and DP . For all cases, BP corresponds with a x-move and CP with a y-move.

19
The conic AF( )P with residue AF is defined as :
[ ] [ ]
A A
A A A A A A A A A A A
A
X A D I x
F F( ) W x y 1 Y x y 1 D B J y
W I J M 1
     
     +      
          
P P G≅ ≅ γ ≅ ≅ (4),
2 2
A A A A A A AF Ax By 2Dx y 2Ix 2Jy M+ + + + +≅ (5).
The discriminant is 2
DIS AB D−≅ . The determinant of a non-degenerated conic is
A D I
DET D B J 0
I J M
≠≅ , hence the Boolean DETb of the global sign of the coefficients of the
non-degenerated conic is DETb IF(DET 0, 1, 0)= > .
An exception to our notation rules is the Boolean LEFTb which equals one when F 0< on our
left, when traversing the conic, else it equals zero, hence LEFTb IF(F(x, y) 0, 1, 0)= < (12), but
LEFT LEFTS IF( b 1, 1, 1 )= == + − (12).
It is important that Lxy x y LEFTb b b b⊕ ⊕≅ (6) and Lxy LxyS IF( b 1, 1, 1 )= == + − (7) are constants
in a monotonic segment.
The points in the candidate cell have the following properties:
C A y B A x D A iS S+ ∆ + ∆ +P P j P P i P P T≅ ≅ ≅ (8), ( )C B x yS S− = − + ∆P P i j (9),
M MX Y
2 2
+ +
⇒ + =B C B C
M M
P P G G
P G i j≅ ≅ (10), B C
M
W W
W
2
+
= (11).
The purpose of the algorithm is to select the optimal candidate point as fast as possible.
Optimal means that the global least square distance to the actual monotonic conic segment is
minimal.
The Directed Polar of a Point with respect to the Conic
For general equations, we assume that the index Z refers to an arbitrary point Z Z Zx y= +P i j
with gradient Z Z ZX Y= +G i j , then we have:
• the polar of an arbitrary point ZP is the line Z Z Z Z ZW 0 ( ) F 0+ = = − + =P G P P Gγ γ (13)
• the “sense of” the directed polar of ZP is the vector Zp LEFT Z LEFT Z ZS ( ) S ( Y X )= × = − +T k G i j
(14), with magnitude Zp Z=T G .
Proof: From (12) and because the gradient points, by definition, in the direction of the greatest
rate of increase of F(x,y). Cross multiplying (14) with ×k proves (15).
From now on we assume that the polar always points in the sense of the movement, with other
words in the monotonic direction, such that the use of “sense of” can be avoided. This property
of the polar will be used in section 3.
• The gradient at the pole of the directed polar ZpT is ZZ Z Z LEFT pX Y S ( )+ = ×G i j T k≅ (15),
• The directed distance from the point BP to the polar of ZP is B Z Z
BZ Z2
Z
W
G
+
=
P G
r G
γ
(16).

20
Some important properties of conics [2, App. A.1-A.7]
The residues MF , CF and BF equal the residues of the conic in MP , CP and BP .
The essential keypoints of conics are:
1. Inflexion points do not exist;
2. Every point, except the center, has a unique polar;
3. The fundamental “switching” property of the polar is 1 2 2 2 1 1W W+ == +P G P Gγ γ (18);
4. The conic can be divided into separate monotonic pieces. Therefore, we say that the conic
can be subdivided into at most four monotonic quadrants;
5. The arithmetic mean equation is 2 1 2 1 2 1 2 1F F ( ) ( )
F
2 2 4
+ + − − 
= + 
 
P P P P G Gγ
(19),
the control factor Mλ equals by definition
x y 2C B C B
M
A B- 2S S D( - ) ( - )
4 4
+
λ = = ∆
P P G Gγ
(20),
hence, M
C B
a M M
F F
F F
2
+
= + λ≅ (21);
6. The incremental equation is 2 1
2 1 2 1
( )
F F 2( )
2
+
− = −
G G
P P γ (22), hence,
C B
C B C B C B MF F 2( ) ( ) 2( )
2
+
− = − = −
G G
P P P P Gγ γ (23);
7. The “relative or simple midpoint measurement” is ( )2 2
CM BM M C B2
M
1
F F F
G
− = −r r (24),
CMr is the directed distance (16) from the point CP to the polar of MP ,
BMr is the directed distance (16) from the point BP to the polar of MP .
2. INTRODUCTION
Overview of the paper and Problem Statements
1. Section 4,”DTLTI-SYSTEM”: describes the optimal conditions for every incremental al-
gorithm. The starting point must be optimal, the candidate points must belong to a mono-
tonic segment, and the optimal candidate point can be selected “locally”, if Bellman’s
principle of optimality holds. The dual of Bellman’s principle of optimality holds for a 2D-
incremental algorithm but not for 6-connected 3D incremental algorithm.
The optimal criterion is the minimal “Global Least Square Distance” to the curve.
2. The measurement of the distance of a point BP to the conic, using an advanced tool, such as
Mathematica’s [ ]BRegionDistance FR,P (17) is simple, but invalid when the “tangent”
(non strictly speaking the “gradient”) at the footpoint of PB is not conform with the mono-
tonic direction Ti. Checking the measurement is generally complex, but even if you find an
advanced checking tool, a new problem arises when there does not exists a valid measure-
ment. The incremental methods use as measurement the midpoint method or the arithmetic
mean method, also called the two-point method. Nowadays there is an agreement, that, for
conics, the midpoint method is better than the two-point method. As the arithmetic mean
equation (21) points out, both methods are related with a constant Mλ and both can be in-
valid. In that case we say that the measurement is Out-of-Control (OoC). When we say that
the midpoint method is “better”, it does not mean that the two-point method is invalid, but
it means that the two-point method may be inaccurate and the midpoint method may be
less inaccurate, or Out-of-Accuracy (OoA).

21
We also show that if the grid distance is sufficient small, that inaccuracy is not as im-
portant, because the measurement can be within tolerance. So the real problem is:
2.1.How can we easily and fast detect that the measurement is invalid ?
2.1.1.Section 3, “Algebraic OoC-condition”: the monotonic condition and the polar
of the midpoint of two candidate points, can be used to check if the “polar” of
the midpoint is conform with the monotonic direction Ti. We demand that the
polars of the surrounding points of the midpoint of the footpoints, all have the
same monotonic condition. .
2.1.2.Section 3, “Primary condition of a measurement”: the primary condition for a
measurement and the gradient of the midpoint of two candidate points, predict if
the measurement of the minimal distance of these points to the conic is valid.
The 2nd
condition and (see section 8.2 “Comparison with Van Aken & Novak”)
the 3rd
condition for a measurement are directly consequences of the primary
conditions, but the 2nd
and 3rd
conditions are not sufficient! .
2.2. How can we continue, when all the measurements are invalid ? .
Section 6,”OoC-Rule”: The primary condition is completely new. If it is not valid,
the measurement or the midpoint criterion is not used, but we show that it can be re-
placed with a very simple rule, which continues the algorithm and reduces the inva-
lidity of the next measurement. The OoC-rule does not measure the distance to the
conic, but it controls the digitization, such that it leaves the OoC state as soon as pos-
sible. The OoC-rule tries to correct the situation as good as possible. When we detect
OoC, for all the measurements, the train is going off the rails, and the OoC-rule must
put the train again on the rails or at least prevent that the train will go off the rails.
3. In stead of an advanced distance tool, we use a very simple tool, “the simple midpoint meas-
urement”. This is the core [2] of all midpoint algorithms (even for algorithm T of [1]): we
measure the difference of the squared distances of two candidate points, f.e. CP and BP to
the polar of their midpoint MP . Working out and simplifying this expression gives (24).
When the the measurement is valid, the primary OoC-condition is true, and this expression
reduces to M M
Lxy M 2
M
X Y
S F 2
G
+
− ∗ ∆ . Therefore, only the sign of the residue of the midpoint
must be checked in a given monotonic segment. .
4. Section 5,“Relative curve measurement” theorem” proves that the relative squared dis-
tances of two candidate points to the conic reduces to the simple midpoint measurement,
provided that both measurements are valid and not Out-of-Accuracy. But inaccuracy has no
effect on the digitization when the squared grid distance is smaller than half the squared
worst-case tolerated tolerance range. The trick used to prove this theorem is the “Construc-
tion of the pole EP ” of section 2. .
5. Section 7, “The OoA event in more details”, describes shortly, when Out-of-Accuracy
may occur. To prove the location of the OoA-segment, we must define the inner and outer of
a conic, but at the same time we can simplify the formal LEFTb -definition with a very users’
friendly definition, as LEFT CCW DETb b b= ⊕ . In stead of LEFTb the new independent variable
becomes CCWb (Fig. 2.). .

22
6. Supplements:
a) Appendix 2: Average %-speed gain.
b) Appendix 3: Examples of bad and good digitalizations.
c) Appendix 4: Simple example which shows that Algorithm T of D. Knuth [1] can be OoA.
d) Berserkless 8-connected midpoint algorithm in pseudo-code (2 + 1 info pages). The
modules T15 and T16 are important. The form is, intentionally, conform with algorithm T
of D. Knuth [1].
e) Complete 8-connected midpoint algorithm in Mathematica-cdf format.
A user who wants supplements d or e has to send me an email. .
Relative measurements of distances, OoC and OoA
Measuring the shortest distance to a conic F(x,y)=0, ultra fast, is still a challenging problem
(solution of a non-linear system) :
CFP and BFP are the footpoints of CP and BP on the conic, and their Euclidean distances from
point CP and point BP to the conic are respectively Cρ and Bρ (Fig.3).
The midpoint of the footpoints is the point C B C B B C
M M M
F F F F F F
F F F
W W
, W
2 2 2
+ + +
= ⇒ = =
P P G G
P G (25).
The footpoints must satisfy their non-linear systems,
C C CF C F C C FF 0, 0,= × = −ρ G ρ P P≅ , and B B BF B F B B FF 0, 0,= × = −ρ G ρ P P≅ .
With Mathematica, we can use “RegionDistance” to find the distance of a point to the conic
and the minimal distance of the points BP , CP and DP to the conic. But this measurement, as
the measurement with the midpoint method, can be invalid (OoC), when the candidate point
does not measure the distance to the actual monotonic conic segment. .
Possible relative measurements
The relative distance is the difference between the squared distances from two points to some
“reference”, and a relative measurement is the measurement of the relative distance. As “refer-
ence”, we will only consider a conic and the polar of a point with respect to a conic. Limiting
the relative measurement to conic curves, is not really a limitation in practice, because about
every shape can be approximated by “piecewise conics”, called quadratic Bézier splines
(conic splines) or squines [1, pp. 48, pp. 181]. .
Replacing the “reference” with a conic or a polar gives, for two given points CP and BP , called
the candidate points, the following relative measurements:
1. The “relative curve measurement” measures the difference between the squared distances
from two points to the conic, hence it measures 2 2
C B−ρ ρ .
2. The “relative polar measurement” measures the difference between the squared distances
from two points to the polar of a special constructed pole EP with respect to the conic, hence
it measures 2 2
CE BE−r r .
3. The “relative or simple midpoint measurement” measures the difference between the
squared distances from two points to the polar of the midpoint MP of these points with
respect to the conic, hence it measures
}
( )
(24)
2 2
CM BM M C B2
M
1
F F F
G
− = −r r . .

23
All midpoint methods for conics use this criterion, hence the midpoint method measures the
“relative midpoint distance”. .
Fig. 3. Case T2
Construction of the pole PE
The tangents, in the footpoints, intersect in the pole FP , hence the polar FPT of the pole FP
intersects the conic in the footpoints CFP and ..with midpoint MFP . The chord vector is by
definition ( )F C B FF L F F PS − ∈L P P T≅ and the gradient in FP is FG . Applying the incremental
equation to the footpoints gives F M M FL F F F p2S 0= =L G G Tγ γ ⇒ ( ) ( )C F B F
2 2
F p F p=G T G Tγ γ . In
general ( ) ( )F F
2 2
C p B p≠ρ T ρ Tγ γ , therefore we construct the pole EP with gradient EG , such that:
.
1. The pivoting point of the polar EpT is the point MFP .
Hence, M M MF E E E F FW 0 W 0+ = ⇔ + =P G P Gγ γ (26).
2. We turn the polar EpT around the pivoting point MFP such that ( ) ( )E E
2 2
C p B p=ρ T ρ Tγ γ (27).
The tangent of the pivoting angle Eφ is a function of
( )
( )
C C B B
E
C C B B
function sin , sin
tg
function cos , cos
ρ α ρ α
φ =
ρ α ρ α
,
with C B CC F F F F,α = − −P P P PΡ , B C BB F F F F,α = − −P P P PΡ (28).
It can be proved that Etg 1φ ≤ , when the measurement is valid.
3. The polar EpT cuts the conic in the points CEP and BEP , and the tangents in these points cut
in the pole EP of EpT . The chord vector is ( )E C BE L E ES −L P P≅ and the gradient in .. is per-
pendicular to the chord EL . The midpoints MEP and MFP of respectively the chords EL and
FL , belong to the polar of EP , therefore

24
ME E EW 0• + =P Gγ (29),
MF E EW 0• + =P Gγ (30),
( )M ME F E E E0 0• − = ⇔ =P P G L Gγ γ .
The poles FP and EP belong to the polar of MFP , and E F−P P is parallel to the chord
( )F C BF L F FS= −L P P or parallel to polar FPT of the pole FP .
We will use this construction in the “relative curve measurement” theorem of section 5.
The three possible measurements ( Fig.1., Fig. 3. )
The candidate points for every monotonic direction with optimal start point AP are the points
BP , CP , and DP for a 8-connected digitization, or the points BP , CP for a 4-connected digitiza-
tion. As we select pairwise, we need three measurements for a 8-connected digitization and one
for a 4-connected digitization. The conic is divided into separate segments, in each of which x
and y are both monotonic. The Booleans of the increments of x , and y are xb , yb ((1) ( 2)),
and the Boolean Lxyb defined as Lxy x y LEFTb b b b⊕ ⊕≅ (6), is fixed in each monotonic segment.
For a monotonic conic segment, the next simple measurements are possible:
1.The M-measurement M C BF (F F )− using points { BP , CP } and their midpoint
( )M C B / 2+P P P≅ with C BF -F Lxyb b== , selects point BP if MF Lxyb b 1⊕ == , else point CP (31).
The Boolean of MF is MFb , and the Boolean of ( )C BF F− is Lxyb . .
The sign of ( )M C BF F F− is the general midpoint criterion, for a 4-connected digitization.
When the “relative curve measurement” is valid and accurate, it corresponds with the Boole-
an expression 2 2
MC B
F Lxyb b bρ −ρ
= ⊕ (52) meaning: if 2 2
C B
bρ −ρ
is true, the midpoint method choos-
es point BP else it chooses point CP . .
2.The H-measurement H D BF (F F )− using points { DP , BP } and their midpoint ( )H D B / 2+P P P≅
with D BF -F Lxyb b== , selects point BP if HF Lxyb b 1⊕ == , else point DP (32).
3.The V-measurement V C DF (F F )− using points { CP , DP } and their midpoint ( )V C D / 2+P P P≅
with C DF -F Lxyb b== , selects point DP if VF Lxyb b 1⊕ == , else point CP (33).
Proof: The result of the “relative curve measurement” theorem of section 5.
In this paper we mostly consider the M-measurement, and we assume that the reader can apply
the same reasoning to the other measurements.
A 100 % stable hardware realization is possible
It is our purpose to pre-design a hardware algorithm, therefore we avoid exceptions, because
simplicity favors regularity, and therefore we will take care of all the possible “valid”
measurements.

25
The basic forms MF Lxyb b⊕ , HF Lxyb b⊕ , VF Lxyb b⊕ and M Lxybλ ⊕ (59) can be easily converted
to hardware and they select the shortest distance to the conic, ultra fast. But each measurement
can be invalid and an invalid measurement is not considered. The OoC-rule (59) is only applied
when there are no valid measurements, therefore only the M-OoC-rule applies. A valid
measurement has always the highest priority, except when the H-measurement selects point BP
and the V-measurement selects point CP , and the M-measurement is invalid. In that case the
OoC-rule selects one of these points. In all other cases, point DP gets the highest priority.
When all the measurements are invalid, we apply the OoC-Rule, which selects the most stable
point out of { BP , CP }.
3. PRIMARY OOC-CONDITION OF A MEASUREMENT
Algebraic OoC-condition
The midpoint measurement is OoC, if the sense of the direction of the digitization is not con-
form to the sense of the monotonic direction.
The monotonic vector iT measures the monotonic direction, and the sense of the directed polar
ZpT of the midpoint ZP , measures the sense of the direction of the digitization. Hence, the polar
ZpT of ZP must be monotonically equal to the monotonic vector iT , notated as Zp iT T (or
simply said conform with the monotonic vector iT ). With x x y yA B A B= ∗ + ∗A Bγ , and
0>A Bγ defined as x xA B 0∗ > and y yA B 0∗ > , the algebraic OoC-condition for a valid
measurement, associated with the midpoint ZP is
Z Zp i p i 0⇔ >T T T Tγ (34).
The notation between quotes,
B C D M V HP P P P P P i" , , , , , , the polars of the surrounding points of the midpoint of the footpoints,... "T T T T T T T
indicates, in detail, which polars are involved , but it always means that the measurement(s)
corresponding with the poles of the polars, must be valid.
Primary OoC-conditions in Boolean form
With (1), (2) and the Boolean .. defined as Lxy x y LEFTb b b b⊕ ⊕≅ , the necessary and sufficient
conditions that a midpoint measurement is valid are:
a. for the measurement using the midpoint MP : 1 2M M Mb b b 1= ∧ == ,
with 1 MM y Y Lxyb b b b= ⊕ ⊕ , 2 MM x X Lxyb b b b= ⊕ ⊕ (35);
b. for the measurement using the midpoint HP : 1 2H H Hb b b 1= ∧ == ,
with 1 HH y Y Lxyb b b b= ⊕ ⊕ , 2 HH x X Lxyb b b b= ⊕ ⊕ (36);
c. for the measurement using the midpoint VP : 1 2V V Vb b b 1= ∧ == ,
with 1 VV y Y Lxyb b b b= ⊕ ⊕ , 2 VV x X Lxyb b b b= ⊕ ⊕ (37).

26
We call these conditions the primary OoC-conditions. These conditions can be checked for
each measurement, and they must be valid for all other known points, but also for all unknown
poles, such as EP , FP , ZP , etcetera.
Proof: We will only prove the conditions for the pole EP , hence we have to prove
1 2E E Eb b b 1= = = , with 1 EE y Y Lxyb b b b= ⊕ ⊕ , 2 EE x X Lxyb b b b= ⊕ ⊕ , and 1 2E E Eb b & b= ; or in
S-form:
1 EE y Y LxyS S S S 1= − ∗ ∗ = , 2 EE x X LxyS S S S 1= ∗ ∗ = , hence 1 2E E ES S S 1= ∗ = .
Ep i LEFT E x yS ( ) (S S )= × + ∆T T k G i jγ γ = Lxy x y E x E yS S S ( ) S ( ) S × ∗ + × ∗ ∆ i k G j k Gγ γ
Lxy y E Lxy x ES S ( ) S S ( ) = − + ∆ j G i Gγ γ E ELxy y Y E Lxy x X ES S S Y S S S X= − ∆ + ∆
1 2E E E ES Y S X= ∆ + ∆ .
The polar of EP is monotonically equal to the monotonic vector iT , when Zp i 0>T Tγ , there-
fore 1ES , 2ES and ES must equal one.
Necessary but Insufficient secondary conditions in Boolean form
When the measurements are valid then
a. for the measurement using the midpoint MP : ( ) ( ) C BC B E C B M F -F Lxyb b b b− −
= = ==P P G P P Gγ γ
(38).
b. for the measurement using the midpoint HP : ( ) ( ) D BD B E D B H F -F Lxyb b b b− −
(39).
c. for the measurement using the midpoint VP : ( ) ( ) C DC D E C D V F -F Lxyb b b b− −
(40).
We call these conditions the secondary OoC-conditions. Each measurement will apply this
condition.
The secondary OoC-conditions are necessary but not sufficient.
We can replace EG with MG , HG , VG , CG , BG , ZG , etc.
Proof: We will only prove (38) with index E:
E EC B E y x E y Y E x X E( ) (S S ) (S S Y S S X )− = − ∆ = − ∆P P G j i Gγ γ . Applying the primary OoC-
conditions Ey Y LxyS S S∗ = − and Ex X LxyS S S∗ = gives, ( )C B E Lxy E E( ) S Y X− = − + ∆P P Gγ .
Proving with index M gives ( )C B M Lxy M M( ) S Y X− = − + ∆P P Gγ . Applying the incremental
equation (23) proves (38).
4. DTLTI-SYSTEM
Looking at (8, 9), the digitizing of 2D-curves can be seen as a Deterministic Discrete-Time
Linear Time-Invariant System [10]. The digitized point at time nt corresponds to the stage n.
For a digitized 2D-curve, the state difference equation is
x nxn 1 n
y nyn 1 n
S 0 ux x
0 S uy y
+
+
      
= +       
       
, and
the inputs are { }nx nyu ,u 0,∈ ∆ and nx nyu u+ = ∆ for a 4-connected 2D-curve, and

INTERNATIONAL JOURNAL OF COMPUTER GRAPHICS & ANIMATION (IJCGA) VOL. 5, NO. 1, JANUARY 2015
27
nx nyu u or 2+ = ∆ ∆ for a 8-connected 2D-curve. This system is time-invariant when
x
y
S 0
0 S
 
 
 
is independent of the time, and this is the case, as long as the stages belong to their monotonic
quadrant. Therefore, the monotonic part of a digitized curve can be modeled as a DTLTI-
system, the input can be considered as the set of feasible decisions
{ } ( ){ }n n x y x yU S , S , S S= = ∆ ∆ ∆ +u i j i j and the state equation (or system function) becomes,
n 1 n n+ = +P P u . The states equal the grid position vectors of the discrete 2D-curve, nP equals
AP , and n 1+P equals BP , CP or DP . The partial trajectory r
0T equals
r 1 r
0 1 2 r 0 0 1 r 1 0 0 0( , , , , ) ( , , , , ) ( , )−
−= = =P P P P P u u u P TΛ Λ PPP , so it depends only on the initial state 0P
and the policy r 1
0
−
PPP of the first r decisions 0 1 r 1( , , , )−u u uΛ . The cost per stage n n nc ( , )P u is the
criterion used to select nP from the candidate points BP , CP and DP given the point A n 1−=P P ,
and the possible moves n 1U − ( ){ }x x x yS ,S , S S= ∆ ∆ ∆ +i j i j , hence the cost per stage n n nc ( , )P u is
independent of the decision nu and depends only on the forward partial trajectory
n
0 0 1 2 n 0 0 1 n 1( , , , , ) ( , , , , )−= =T P P P P P u u uΛ Λ . We assume, that the Least Square Distance is the
criterion n n nc ( , )P u used for digitizing the conic. The set of feasible decisions is independent of
the stage, provided that the stages remain in their monotonic quadrant. Therefore, the system is
time-invariant and deterministic as long as the stages belong to their monotonic quadrant. The
objective is to find a complete trajectory Tthat minimizes the cost of the complete trajectory
( ) ( )
n N
N N
0 0 n n n
n 0
V V c ( , )
=
=
= = ∑T T P u .
For a time-invariant deterministic dynamic system, the recursive procedure can be based on a
forward induction process, where the first stage to be solved is the initial stage of the problem,
and problems are solved moving forward one stage at a time, until all stages are included
[11],[12]. .” Hence, the starting point AP of (8) must be optimal ! The basis of the forward re-
cursive optimization procedure is a dual to Bellman’s statement: ”An optimal policy has the
property that, whatever the ensuing state and decisions are, the preceding decisions must con-
stitute an optimal policy with respect to the state existing before the last decision.
If the dual of Bellman’s’ principle of optimality holds, then
r 1 r r r
0 0 0 r 1 r 1 r 1 0 r 1 r 1 r 1
local
J min[V ( )] min[c ( , )]] J min[c ( , )]]+
+ + + + + += + == +T P u P u
144424443
⇔ The global minimum equals the sum of the local minima.
The solution of the DP-problem does not say how we have to select the best point or how we
have to find the best decision vector. It just says, if Bellman’s principle or its dual holds, then
you have a local problem, and if you can find a solution for that local problem, that solution is
also valid for the global problem. It is clear that the principles hold for a 4- or 8-connected 2D-
curve, and a 26-connected 3D-curve, but not for a 6-connected 3D-curve! Therefore, the Tri-
pod 6-Connected 3D Line algorithm [13] is not global optimal.
5. “RELATIVE CURVE MEASUREMENT” THEOREM
The “relative curve measurement” theorem proves that
2 2
C B E E M C B E2
E
2
(1 ) F ( )
G
− = − ε τ −ρ ρ P P Gγ or 2 2
M C B E E EC B
F ( ) (1 )b b b b b− τ −ερ −ρ
= ⊕ ⊕ ⊕P P Gγ .
When a measurement is valid, the parameter Eε is always smaller than one, and when the meas-
urement is accurate, we have E 0τ > , hence MC B F Lxyb b bρ > ρ
= ⊕ .

28
Proof: In section 2, we constructed the pole EP such that ( ) ( )E E
2 2
C p B p=ρ T ρ Tγ γ .
Using the identity ( ) ( )
2 22 2
i pE i E i E= ∗ −ρ T ρ G ρ Gγ γ , from appendix 1, it is now easy to calculate
2 2
C B−ρ ρ because ( ) ( )
2 22 2 2 2
C E C E B E B E∗ − = ∗ −ρ G ρ G ρ G ρ Gγ γ .
Therefore ( )2 2 2
C B E− ∗ρ ρ G equals
C B F F C B F FC B C B
C B E C B E
( ) ( )
( ) ( )
+ − + − − −
   
   + ∗ − =
   
   P P P P P P P P
ρ ρ G ρ ρ Gγ γ
14243 14243 M C BM E F E C B E F F E2 ( ) ( )  − ∗ − − −   P G P G P P G P P Gγ γ γ γ
= ( ) ( ) ( )( )
( )
M C B
E M M
E C B E
F
E M
M E E F E E C B E F F E
W 0
2 W W
τ
+
ε −
         + − + ∗ − − −           
P G
P P G
P G P G P P G P P G
γ
1442443 γ
γ γ γ γ
1442443 1442443 1442443
.
Hence, the “relative curve measurement” is 2 2
C B E E M C B E2
E
2
(1 ) F ( )
G
− = − ε τ −ρ ρ P P Gγ (41).
These parameters have only sense when the measurement is valid, and they equal
F
FC FB E F E
E L
C B E C B E
( - )
S
( - ) ( )
ε
−
P P G L G
P P G P P G
γ γ
≅ ≅
γ γ
(42), and E M M
E
M
W
F
+
τ
P Gγ
≅ (43).
The directed distance of the point CFP to the polar of EP is C
C
F E E
F E E2
E
W
G
+
=
P G
r G
γ
. Therefore,
• C B
C B
F F E E
F E F E E2
E
( ) 2W
G
+ +
+ =
P P G
r r G
γ MF E E
E2
E
W
2 0
G
+
= =
P G
G
γ
(44),
because the pivoting point MFP belongs to the polar of EP and FP .
• C BF E F E E CE BE( )− = ε −r r r r (45) and you may expect that Eε will be very small, as E
E F
lim
→
ε equals
0, because F F 0==L Gγ . Algebraic, it is rather complex to prove, that E(1 )b 1−ε == . One has
to prove that FL 2< ∗∆ and that the pivoting angle =
E FE p p F E( , ) ( , ) 4φ = ≤ = ≤ πT T G GΡ Ρ . Therefore, we decided to cancel that proof in this
paper.
• C B E E M E E E M M
CE BE E E E2 2 2
E E E
( ) 2W W W
2 2
G G G
+ + + +
+ = = =
P P G P G P G
r r G G G
γ γ γ
(46).
Equation (46) is the result of applying the “switching” property for polars (18). We know
that the residue in the midpoint MP equals M M M MF W= +P Gγ , hence the natural choice is to
link E M MW+P Gγ to M M MW+P Gγ , using the parameter E M M
E
M
W
F
+
τ
P Gγ
≅ .
The parameter Eτ becomes one for Mτ .
Therefore (41) defines different measurements in function of the parameters Eε and Eτ :
• the “relative midpoint measurement” measures the difference between the squared distances
from the candidate points CP and BP to the polar of MP , and it equals
( )2 2
CM BM M M C B M M C B2 2
M M
2 1
F ( ) F F F
G G
− = τ − = −r r P P Gγ (47),
This measurement is accurate, and has no tolerance, but it can be invalid (OoC);

29
• the “relative polar measurement” 2 2
CE BE E M C B E2
E
2
F ( )
G
− = τ −r r P P Gγ (48)
measures the difference between the squared distances from the candidate points CP and BP
to the polar of EP with respect to the conic;
• the “relative curve measurement” 2 2
C B E E M C B E2
E
2
(1 ) F ( )
G
− = − ε τ −ρ ρ P P Gγ measures the dif-
ference between the squared distances from the candidate points CP and BP to the conic.
When the measurement is valid, we have E(1 )b 1−ε == and the secondary OoC-conditions
C B E( ) Lxyb b− =P P Gγ , etc. hold. The parameters Eτ and Fτ are responsible for Out-of-Accuracy.
From the definition of OoA, we have E
b 1τ = when the measurement is accurate, and E
b 0τ =
when the measurement is Out-of-Accuracy (OoA). .
When the measurement is valid, we have,
• C B E( ) Lxyb b− =P P Gγ , etc. (49),
• E(1 )b 1−ε = (50),
• 2 2
M EC B
F Lxyb b b bτρ −ρ
= ⊕ ⊕ (51).
If the measurement is valid and accurate then E
b 1τ = . Therefore, when the measurement is val-
id and accurate, we have MC B F Lxyb b bρ > ρ
= ⊕ (52).
So, the “relative curve measurement” reduces to the “relative simple midpoint measurement”
provided that the measurement is valid and accurate.
Tolerance of the measurement
When the measurement is inaccurate it can be within the tolerance. The worst-case tolerance
range is defined as the maximal distance between the candidate points ( )2∆ , and the toler-
ance equals half the tolerance range.
If the measurement is inaccurate ( E
b 0τ = ) then we have 2 2
MC B
F Lxyb b bρ −ρ
= ⊕ , the midpoint
method is wrong, and the measurement is “Out-of-Accuracy” (OoA). But the measurement is
within tolerance if
worst-case tolerance range which can be tolerated
2 worst-case tolerance
2
∆ ≤ = ∗ (53).
Avoiding Aliasing using the (Nyquist-Shannon) Sampling Theorem:
The radius of curvature of a conic equals
3
Cur
G
R
DET
= , and the sampling theorem states that
CurR
1≥
∆
(common minimum values are 5 à 10, say Sn ). Therefore the grid-distance must be
smaller then Minimum( 2 worst-case tolerance∗ , Minimum( Cur
S
R
n
) ).
A circle with a radius equal to the grid-distance digitizes as a square!

30
Two-point method or midpoint method
We show in section 7, that OoA happens once in a while, when the midpoint is, in practice,
inside and very, very near to the conic. In that case we have,
• M MM E EF F< = λ (54),
• C B−ρ ρ = Tolerance,
• C tolerance≤ρ and B tolerance≤ρ ,
•
}
( )M M M
(20)
E M a a a M MF F Fτ = τ = τ + λ .
Hence, E M a aM M
F Fb bτ τ= and that resolves the long existing conflict between the midpoint method
and the two-point method. In practice, if one detects that the sign of MF and MaF are different
then one may have OoA. Section 7 proves that this way of detecting OoA, has many false sig-
nals. Simply said the midpoint method is better than the two-point method (for conics).
6. OOC-RULE
Beforehand, one calculates M x yA B- 2S S DλΛ = + (55), M M
b IF( 0, 1, 0)λ λ= Λ ≥ (56), and
M M M MFb IF( 0, b , b )λ λ λ= Λ = (56).
The update equations from MP to M
ˆP are: M M xmove x ymove y
ˆX X (b S A b S D)= + + ∆ (57), and
M M xmove x ymove y
ˆY Y (b S D b S B)= + + ∆ (58).
The OoC-Rule xmove Lxyb b bλ= ⊕ , ymove Lxyb b bλ= ⊕ must be applied when M H Vb b b 0= = = ,
viz. when all the measurements are invalid. For the proof, we focus on the M-measurement.
Proof:
From the incremental equation (23) and (9), we have C B y M x MF F 2(S Y S X )− = − ∆ .
When the primary condition is true then C B Lxy M MF F 2S ( Y X )− = − + ∆ , and if LxyS equals +1
then y M x MS Y S X 0− < . But when y M x MS Y S X 0− ≥ and LxyS equals +1, then the primary
condition is not true and the best solution is to make y M x MS Y S X− as negative as possible
without applying the invalid M-measurement. Otherwise, when the primary condition is true
and if LxyS equals -1 then y M x MS Y S X 0− > . But when y M x MS Y S X 0− ≤ and LxyS equals -1,
then the primary condition is not true and the best solution is to make y M x MS Y S X− as positive
as possible without applying the invalid midpoint method. Therefore, we have the next cases:
> Case 1, Lxyb 0= and Mb 0= :
From position AP , we have to find a single step, which brings us to the new position A
ˆP , and
the new difference becomes y M x M
ˆ ˆ(S Y S X )− . The only thing we can do, is to choose the step,
which makes y M x M
ˆ ˆS Y S X− the most positive. Therefore, y M x M y M x M
ˆ ˆS Y S X S Y S X− < − or
y M M x M M
ˆ ˆ0 S (Y Y ) S (X X )< − − − . Using the update equations gives,
• for a x-step x y0 S S D A< − , because M M x
ˆY Y S D− = ∆ and M M x
ˆX X S A− = ∆ ;
• for a y-step x y0 B S S D< − , because M M y
ˆY Y S B− = ∆ and M M y
ˆX X S D− = ∆ .

31
Therefore, given Lxyb 0= , a x-step is better than a y-step if x y x yB S S D S S D A− < − or if
M
2
x y
4
A B 2S S D 0
λ
∆
+ − <
1442443
, or M Mxmove Lxy ymove Lxyb b b and b b bλ λ= ⊕ = ⊕ (59).
> Case 2, Lxyb 1= and Mb 0= :
From position AP , we have to find a single step, which brings us to the new position A
ˆP , and
the new difference becomes y M x M
ˆ ˆ(S Y S X )− . The only thing we can do, is to choose the step,
which makes y M x M
ˆ ˆS Y S X− the most negative. Therefore, y M x M y M x M
ˆ ˆS Y S X S Y S X− < − or
y M M x M M
ˆ ˆS (Y Y ) S (X X ) 0− − − < . Using the update equations gives,
• for a x-step x yS S D A 0− < , because M M x
ˆY Y S D− = ∆ and M M x
ˆX X S A− = ∆ ;
• for a y-step x yB S S D 0− < , because M M y
ˆY Y S B− = ∆ and M M y
ˆX X S D− = ∆ .
Therefore, given Lxyb 1= , a x-step is better than a y-step if x y x yS S D A B S S D− < − or if
M
2
x y
4
A B 2S S D 0
λ
∆
+ − >
1442443
, or M Mxmove Lxy ymove Lxyb b b and b b bλ λ= ⊕ = ⊕ (59).
The control factor Mλ can become zero for a parabola and a hyperbola. In that case, it does not
matter if we make a x- or a y-move, and we can just as well make a move that corresponds with
the measurement. Hence, Mxmove F Lxyb b b= ⊕ or M M M MFb IF( 0, b , b )λ λ λ= Λ = .
7. THE OOA EVENT IN MORE DETAILS
To explain the OoA event, and to find the location of the midpoints { }M H V, ,P P P , we need to
know the sign of the inside of a conic and the sign of the residue at a point to the left or the
right of a polar.
Inside and outside of a conic
Determining the inside of a curve can be very complex. Knuth [1, pp. 44] used the Jordan
curve theorem, we use the topological definition: a point is inside a conic if the conic is
concave, seen from that point. Therefore the center of an ellipse is inside, and the center of a
hyperbola (intersection of the asymptotes) is at the outside of the conic. From [9], the residue
CTF at the center of a conic equals
DET
DIS
, therefore the Boolean of the residue of the center of
an ellipse or hyperbola is CTF DIS DETb b b= ⊕ (60).
The center of a parabola is at infinity, but empirical it is, as the center of a hyperbola, at the
outside. So, we can say that the center of a every conic is inside, if DIS 0> and outside, if
DIS 0≤ .
Therefore the algebraic definition of inside is PPinside F DETb b b= ⊕ (61) because
• CTDIS F DETb b b= ⊕ ,
• CTP inside DISb b== , and
• P CTF Fb b== as long as the point and the center are inside or outside the conic.

32
The sign of the residue of a point, determines if the point is inside or outside the conic, but it
also depends on the sign of the determinant. We need the inside-definition to locate the OoA-
segment, but at the same time, it gives a user’s friendly approach of defining LEFTb : traveling
clockwise or counterclockwise is the person’s view from the inside. Using this approach, it is
easy to prove that LEFTb equals LEFT CCW DETb b b= ⊕ (62).
The polar divides the plane in a + and - half plane
If P and EP are at the same side of the polar line of EP then E E EM
( - ) Fb b=P P Gγ (63) else
E E EM
( - ) Fb b=P P Gγ and F E EM
( - ) Fb b=P P Gγ (64) else F E EM
( - ) Fb b=P P Gγ with EFb the Boolean of the
residue E EF F( )= P .
The location of the midpoint PM defines the sign of τE
From (43), the switching property (18), and (29), the parameters can be written as
ME M E M M M E E M E EF W W ( - )τ ∗ = + = + =P G P G P P Gγ γ γ (65),
M M M M M M MF W F 1τ ∗ = + = ⇒ τ =P Gγ (66).
If the midpoint MP is
1.outside the conic, at the same side of the chord C BE EP P as EP then MF DETb b= and
M E E EM
( - ) F DETb b b= =P P Gγ , E M M E EM
F ( - )b b bτ = ⊕ P P Gγ ⇒ E 0τ > (67),
2.inside the conic, at the same side of the chord C BE EP P as EP then MF DETb b= and
M E E EM
F ( - )b b bτ = ⊕ P P Gγ ⇒ E 0τ < (68),
3.inside the conic, at the other side of the chord C BE EP P as EP then MF DETb b= and
M E E EM
F ( - )b b bτ = ⊕ P P Gγ ⇒ E 0τ > (69).
So, we have OoA if the midpoint MP is inside the conic at the same side of the chord C BE EP P as
EP , which is always at the outside of the conic. We call the area bounded by the chord C BE EP P
and the conic, the OoA-segment, which has the next properties:
• the measurement is inaccurate when the midpoint is inside or on the OoA-segment,
• applying the arithmetic mean equation to the points CEP and BEP gives
( ) ( )C B C B
M M
E E E E
E EF
4
− −
λ = = −
P P G Gγ
(70),
• MM EF < λ if MP is on or inside the OoA-segment (71),
• M E MM
F Fb b+λ = if MP is inside the OoA-segment (72).
Hence, for M ME M E M EF (F )τ = τ + λ , we have for (72) M
M
E E
EE
τ τ
= −
ττ
. Therefore, if we use
MM EF + λ instead of MF , the inaccuracy disappears. Unluckily, the exact value of MEλ is
unknown. If we replace MEλ with Mλ (two-point method), we enlarge the OoA-segment . This
means that we will have many false OoA-events.

33
8. COMPARISON WITH PREVIOUS WORK
The midpoint methods use about the same measurements, and the general midpoint algorithms
[3], [16], [4], [5, pp. 947, 951-961], [6] also use measurements to detect quadrant change
control instead of using the monotonic approach. The monotonic approach is better:
• it allows the deterministic, time-invariant modeling;
• the frame can be determined beforehand;
• no quadrant change control problems;
• the general algorithms translate the curve mostly to (0,0), but the frame can only be
determined after several tryouts.
Therefore, we will only compare the Berserkless Midpoint Algorithm with other midpoint al-
gorithms with a monotonic approach.
1. Comparison with algorithm T of D. Knuth [1, pp. 44-48, Exercise 182 at pp.47, 66 and 179]
This is a 4-connected midpoint algorithm, not based on [7], [5]. It is ultra fast, but it is not
100% stable. If you delete the results of the OoC-Rule and if you set H Vb b 0= = , then you
obtain, about the same non-stable algorithm, after some simplifications. The proposed
corrections [1, pp. 183] are comparable with others: they consider the next successive
quadrant(s), and therefore the next monotonic direction(s). Sometimes it works, but essential it
is wrong, because the monotonic direction is given a priory, and looking around the corner,
neglects what is wrong in the actual monotonic segment.
2. Comparison with Van Aken & Novak [7]
They use separate algorithms for separate conics, and their separate conics are monotonic.
They state that the accuracy
tolerance range
2
≤ , and that from empirical measurements, the
midpoint method is better than the two-point-method; we have done the same experiments and
obtained the same result, for grosso modo all the examples of [3] including thin and sharp
conics.
Van Aken & Novak state, at page 166-168, the third OoC-condition: H
D B
4
1
F F
λ
<
−
, with
2
H D B D B4 ( ) ( ) Bλ − − = ∆P P G G≅ γ , D B Lxy HF F S Y− = − ∆ .
For the M-measurement, the third OoC-condition becomes M
C B
4
1
F F
λ
<
−
(73).
This can be proved very easily, if the primary condition, and therefore the 2nd
OoC-condition is
true.
Proof: C B
C M
2
−
= +
G G
G G , and C B
B M
2
−
= −
G G
G G . Dot multiplying with C B( - )P P γ gives
respectively C B C B
C B C C B M
( - ) ( )
( - ) ( - )
2
−
= +
P P G G
P P G P P G
γ
γ γ and
C B C B
C B B C B M
( - ) ( )
( - ) ( - )
2
−
= −
P P G G
P P G P P G
γ
γ γ .
From the incremental equation (23), and the arithmetic mean equation (21), we have,
M
C B C C B
C B
4
2( - ) (F F )*(1 )
F F
λ
= − +
−
P P Gγ , and M
C B B C B
C B
4
2( - ) (F F )*(1 )
F F
λ
= − −
−
P P Gγ .
From the primary OoC-conditions (38) C B C C B B C B( - ) ( - ) (F F )b b b −= =P P G P P Gγ γ , therefore M
C B
4
1
F F
λ
<
−
.

34
Van Aken & Novak do not state the primary and secondary OoC-conditions, and we do not use
directly the 3th OoC-condition. Their statement to avoid OoC, based on the 3th OoC-condition,
is as acknowledged by them, completely wrong. The third OoC-conditions and the secondary
OoC-conditions are the result of the primary conditions; they are not wrong, but these
conditions do not give a 100% stable midpoint algorithm. .
Van Aken & Novak [7, pp. 168] agree that their algorithms are not successful in a region in
which two edges of a curve meet or cross each other. D. Knuth [1, pp. 46] avoids this situation
by its a priori condition : “No edge of the integer grid contains two roots of Q”.
When we apply the OoC-Rule:
• conics which have two roots on the line segment, connecting the candidate points, are not
excluded;
• sharp, very sharp turning and thin conics are not excluded;
• unusual cases do not tend to drive the conics berserk.
3. comparison with all midpoint algorithms
All general, non-line midpoint algorithms are not 100% stable, therefore they cannot be trans-
formed to hardware. All non-line midpoint methods, except [8], cannot be linked, directly, with
the “relative curve distance” that measures 2 2
C B−ρ ρ .
All earlier non-line midpoint algorithms do not apply the primary conditions, nor the the OoC-
rule; and most of them even do not apply the 2nd
conditions.
9. CONCLUSION
Before, the digitization of conics, using the midpoint or two-point method, had a bad
reputation:
• “Algorithms for discrete geometry are notoriously delicate: unusual cases tend to drive them
berserk”, and “Reasonable conics don’t make such sharp turns” [1, pp. 180];
• “The generation of thin and sharp turning hyperbolas remains unsolved” [3, pp. 36];
• “Digitizing general conics is very hard, the octant-changing test is tougher, the difference
computations are tougher. Do it only if you have to” [6, pp. 42, course cs123].
The phrases “midpoint technique” and “midpoint method” must be replaced by “midpoint
measurement”. The midpoint measurement finds the shortest distance from two points to a
conic or QSIC, provided that the starting point is optimal, and the measurement is valid and not
OoA; but even in the latter case, the tolerance of the measurement is
2
2
≤ ∆ in 2D and
3
2
≤ ∆ in 3D. Every digitized curve has non-zero tolerance and the OoA’ s of the midpoint
measurements are unimportant when they do not influence the tolerated tolerance.
The midpoint algorithm is ultra fast using the valid measurements or the OoC-rule, it is robust
and 100% stable using the OoC-rule and it is appropriate to be converted in hardware.
We also solved the long existing enigma between the midpoint method [5, 6, 7] and the two-
point method [8], and even the mystery of the midpoint method.
The Berserkless Midpoint Algorithm can be extended to 3D-QSIC-curves (intersections of
quadrics).
CNC-machines need the grid points, and this poses a real problem for QSICS. Mathematica
cannot calculate the QSIC itself, but it only shows the QSIC. The most difficult task is the
calculation of the extreme rational tangent points.

35
10. APPENDICES
Appendix 1 : Proof of the identity of section 5
Lagrange’s identity for the vectors A and B is 2 2 2 2
( ) ( )∗ = + ×A B A B A Bγ .
Hence, with i=A ρ and E=B G , we have ( ) ( )
2 22 2
i E i E i E× = ∗ −ρ G ρ G ρ Gγ .
From (15), we have EE LEFT pS ( )= ×G T k , and applying the vector triple product
( ) ( ) ( )× × = ∗ − ∗A C D A D C A C Dγ γ with Ep=C T and =D k gives
E E E Ei p i p i p i p( ) ( ) ( ) ( )× × = ∗ − ∗ = − ∗ρ T k ρ k T ρ T k ρ T kγ γ γ because the vector k is perpendicular to
the plane containing the distance vector iρ , the tangent EpT , and the gradient EG .
Hence ( ) ( )E E
22 2
i E LEFT i p i pS ( ) ( )× = − ∗ =ρ G ρ T k ρ Tγ γ = ( )
22 2
i E i E∗ −ρ G ρ Gγ
Appendix 2 : Average %-speed gain versus Mathematica’s ContourPlot[]
Berserkless Midpoint Algo-
rithm
Mathematica version 10.00 Btime Mtime
Mtime
−
Btime Mtime
8.122465 11.85668 31.49 %
8.007458 10.15458 21.14 %
7.920453 10.77762 26.51 %
7.904452 10.90962 27.55 %
Average %-speed gain using the Berserkless Midpoint Algorithm 26.67 % ± 5%
Mathematica’s AbsoluteTime[] measures the execution time of both algorithms. The genera-
tion of the drawings is the only difference of the two algorithms.
Appendix 3: Examples of bad and good digitalizations
Ellipses (Algorithm T of D. Knuth digitizes B-4 and D-4) :
A B C D
F(x,y)=x² + 15² y² - 15²
From (0,-1) to (15,0)
F(x,y)=-160x²-921y²+767xy-104x+249y
8
OoC-rule / Finish
horizontally
4
Alg. T of D. Knuth
Finish horizontally
From (0,0) to (7,3)
OoC-rule digitizes
the red points
Non-monotonic
& Berserk
Monotonic
algorithm
Non-monotonic
& Berserk
Berserkless midpoint
algorithm D-8

36
Wide hyperbola 2 2
F[x, y] 24x 4y 2 10xy 2 17x 2 7y 8= + + ∗ + ∗ + ∗ + , but figures rotated 90° CCW
From Mathematica, it suggests a
hyperbola with two thin branches
Wide branches of hyperbola, generated by the berserkless midpoint algorithm.
The line y+0.5=-2.5(x+0.5) does not cut the hyperbola.
Appendix 4 Simple example which shows that Algorithm T of D. Knuth [1] can be OoA:
T-algorithm of D. Knuth using the notations and formulas of [1, pp. 46-47]:
F[x,y]=20x²+20y²-291, Q[x,y]=F[x-0.5,y-0.5]=20x²+20y²-20x-20y-281,
a=c=20,b=0,d=c=-20,f=-281.
Cases T2,T3,T4,T5 of D. Knuth correspond with the direction vectors T1,T4,T2,T3.
Initialization T1:
Q[4,0]=-41, Qx[4,0]=140, Qy[4,0]=-20, QM=Q=Q[x,y+1]=-41, RM=R=Qx[x,y+1]=120,
SM=S=Qy[x,y+1]=40 (see [1, pp 45 & 46 and (179))
Loop T4:
x y QM or Q RM or R SM or S This is loop T4
4 0 -41 120 40 Do T6 if QM < 0
4 1 -1 120 80 Point (4,1) is Out-of-Accuracy
4 2 79 120 120 Point (3,1) is better
3 2 -41 80 120
3 3 79 80 160 Do T8 if QM >=0
FR=ImplicitRegion[F[x,y]==0,{x,y}], d[xi_,yi_]=N[RegionDistance[FR,{xi,yi}]
d[4,2]=0.658 > d[3,1]=0.652 for R² = 291 / 20 and
d[4,2]=0.664 > d[3,1]=0.646 for R² = 290 / 20 (midpoint on the circle).
Note that D. Knuth’s QM function is identical with the midpoint function MF .
11. ACKNOWLEDGMENT
We wish to acknowledge the careful review and helpful comments of Dr. H. V. Ramakrishnan,
Dr MGR Educational & Research Institute University Maduravoyal, Chennai.
402002040
100
50
0
50
100
x
y
402002040
100
50
0
50
100
xScale1
yScale1
402002040
100
50
0
50
100
x
y

37
12. REFERENCES
1. Donald E. Knuth, “The Art of Computer Programming, Bitwise Tricks & Techniques,”
Volume 4, Fascicle 1, Addison-Wesley, ISBN 0-321-58050-8, 20 December 2008.
2. V. Huypens, “The Sign Corrected Midpoint Decision Variable selects the candidate point
with the minimum Euclidean distance to the conic”, Geometric Modeling and Imaging,”
3rd International Conference GMAI July 2008, IEEE Service Center Order Nr. P3270,
ISBN 978-0-7695-3270-7, pp. 15-20.
3. M. Golipour-Koujali, “General Rendering and Antialiasing Algorithms for Conic Sec-
tions,” London South Bank University, London, England, May 2005, ManosPhDthesis.pdf.
4. M.L.V. Pitteway, “Algorithm for Drawing Ellipses or Hyperbolae with a Digital Plotter,”
Computer Journal, 10, pp. 282-289, 1967.
“Welcome to...”, wscg.zcu.cs/wscg2005/papers_2005/piiteway.pdf.
5. J.D. Foley, A. van Dam, Feiner, Hughes, "Computer graphics: Principles and practice,"
Second edition, Addison Wesley, reprinted with corrections, July 1997.
6. A. van DAM, “cs123(2010): Introduction to Computer Graphics, Scan Conversion,
“http://www.cs.brown.edu/courses/cs123/ lectures/scan_conversion.pdf”.
7. J. Van Aken, M. Novak, "Curve-Drawing Algorithms for Raster Displays," ACM Transac-
tions on Graphics, Vol. 4, No. 2, April 1985, pp. 147-169.
8. J.E. Bresenham, “A Linear Algorithm for Incremental Digital Display of Circular Arcs,”
Communications of the ACM, Vol. 20, No. 2, pp. 10-106, Feb. 1977.
9. John Henry Tanner and Joseph Allen, “An elementary course in analytic geometry”, before
1923, New York: Cincinnati, Chicago, American Book Company.
10. H. Kwakernaak, R. Sivan “Linear Optimal Control Systems, Chapter 6: Discrete-Time Sys-
tems,” Wiley-Interscience, 1972.
11. R.E. Bellman, “Dynamic Programming,” Princeton University Press, Princeton, NJ., 1957,
Republished 2003: Dover, ISBN 0486428095.
12. M. Sniedovich, "Dynamic Programming Foundations and Principles.," Second edition,
CRC Press, with chapter on forward dynamic programming, ISBN 978-4200146-3-1, Sept.
2010.
13. Cohen-Or D. and Kaufman Arie,”3D Line Voxelization and Connectivity Control”,Vol.
17, November 1997, pp. 80-87.
14. Laurent Dupont, Sylvain Lazard, Daniel Lazard, Sylvain Petitjean,”Near-Optimal Parame-
terization of the Intersection of Quadrics: Theory and Implementation”, Proc. of SoCG
(ACM Symposium on Computational Geometry), San Diego, pp. 246-255, 2003.
15. Sylvain Lazard, Luis Mariano Peñaranda, Sylvain Petitjean,”Intersecting Quadrics An Ef-
ficient and Exact Implementation”, Proc. of the twentieth annual symposium on Computa-
tional Geometry, June 8-11, 2004, Brooklyn, New York, USA, pp. 4
16. E. Banissi, M.K. Golipour,”A New General Incremental Algorithm for Conic Section”,
11th
International Conference on CGIV, Aug. 2014, pp. 46-51

Relative squared distances to a conic

More Related Content

What's hot (19)

Viewers also liked (14)

Similar to Relative squared distances to a conic (20)

Recently uploaded (20)

Relative squared distances to a conic