This paper uncovers universal linear mean relationships involving polynomial roots and their derivatives across all degrees, revealing new patterns and connections to classical polynomials and number sequences.
Contribution
It generalizes mean slope relationships for polynomials of any degree and provides a systematic procedure to determine these relationships, extending known cases to all dimensions.
Findings
01
Established explicit relationships for polynomials up to degree 40.
02
Discovered a single universal relationship in even dimensions and a two-parameter family in odd dimensions.
03
Connected these relationships to classical polynomials and Stirling numbers.
Abstract
In any cubic polynomial, the average of the slopes at the 3 roots is the negation of the slope at the average of the roots. In any quartic, the average of the slopes at the 4 roots is twice the negation of the slope at the average of the roots. We generalize such situations and present a procedure for determining all such relationships for polynomials of any degree. E.g., in any septic f, letting fn denote the mean f value over all zeroes of the derivative f(n), it holds that 37f1−150f3+200f4−135f5+48; and in any quartic it holds that 5f1−6f2+1f3=0. Having calculated such relationships in all dimensions up to 40, in all even dimensions there is a single relationship, in all odd dimensions there is a two-dimensional family of…
No public reviews on file for this paper yet. If you reviewed it on a platform where reviews are public (OpenReview, ICLR, NeurIPS, ICML), you can paste yours below so the community can read it here.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Universal Linear Mean Relationships in All Polynomials
Gregory Gerard Wojnar
Daniel Sz. Wojnar
Leon Q. Brin
Abstract
In any cubic polynomial, the average of the slopes at the 3 roots is the
negation of the slope at the average of the roots. In any quartic, the
average of the slopes at the 4 roots is twice the negation of the slope at
the average of the roots. We generalize such situations and present a
procedure for determining all such relationships for polynomials of any
degree. E.g., in any septic f, letting fn denote the mean
f value over all zeroes of the derivative f(n), it holds
that 37f1−150f3+200f4−135f5+48f6=0; and in any quartic it holds that 5f1−6f2+1f3=0. Having calculated such relationships in all dimensions up to 49, in all even dimensions there is a single relationship, in all odd dimensions there is a two-dimensional family of relationships. We see connections to Tchebyshev, Bernoulli, & Euler polynomials, and Stirling numbers.
This paper started with (a) the
observation that the quadratic formula essentially provides the roots to be
x=μ±σwhere μ is the mean of the two roots and where σ
is the standard deviation of the two roots considered equally likely, (b) the
observation that the slopes at the roots are
±Discriminant=±2aσ, and (c) the trivial
observation that the average of the slopes at the roots equals the slope at
the average of the roots. Curiosity took us next to cubics, to find that
the Cardano-Tartaglia cubic formula can be perceived as providing the roots
to be
rk=E+ωkT+⇀+ω−kT−⇀ where
ω=−21+i23 is a primitive cube root of
unity, k∈{0,1,2}, and where
T±⇀:=32W±(2W)2−(2V)3with V being the variance of the 3
roots, W being the 3rd central moment (sometimes
referred to as the unscaled skewness), and E≡μ being the
expectation of the roots, considered as being equally likely. Moreover, the
coordinates of the cubic’s inflection point are
(E,−aW), and the slope at the inflection
point is the negation of the mean slope at the 3 roots. Indeed, the
inflection point slope will be −23aV. Next we saw
that the mean slope at the 4 roots of a quartic is the negation of twice
the slope at the point (E,f(E))… and the hunt was on to find all such relationships! It was not
just mean slope relationships either– e.g., one can of course note that the
mean of a cubic’s function values at the roots of the first derivative equals
the ‘mean’ of the function’s (sole) value at the root of the second derivative, viz.
−aW. What other relationships lie waiting to be discovered? Have we only seen the tip of an iceberg?
Perhaps the reader is starting to suspect that such relationships are really
typical.
1 The General Question
We establish some notations. For any degree D polynomial f:C→C let
R≡Rf denote the family (≡ multiset) of D roots, and let
R(ρ)≡Rf(ρ)
denote the family of (D−ρ)roots of the ρth
derivative f(ρ) (sometimes using primemarks
[′] for lower orders); we also write ∣Rf∣=D (even if some root(s) have multiplicity greater than 1). Denote the average value of
the function on the ρth derivative roots family as
f(R(ρ)), and similarly
use
f(δ)(R(ρ)) for the average value of the δ th derivative of
f over the same roots family. Sometimes we will emphasize the degree of
f by explicitly writing fD, and we put φDδρ:=fD(δ)(R(ρ)) .
Our initial results mentioned above suggest that we seek to determine
φDδρ for all D,δ,& ρ, and that we seek
relationships among various φDδρ, of the form
∑ραρφDδρ=0. (Other linear
combinations of the φDδρ that sum over D or over
δ are unnatural because of dimensionality considerations.)
Proof Routes
In cases where the cardinality of the roots family
R(ρ) is 2 or 3, we can compute by brute
force since the values of the roots will be given by either the quadratic or
cubic formula. Perhaps this method could even be pushed to the case of a
family of 4 roots via the Ferarri quartic formula, but the level of
complexity is greatly increased. In any case, for root families of size 5
or greater, we need some alternative route.
Again we benefit from establishing notations. We shall use what we refer to
as the quasi-binomial representation of polynomials, expressing coefficients
in terms of averaged symmetric polynomials in the roots. For example, a
general cubic is represented as
f(x)=a(x3−3rx2+3rx−r) where it turns out that
r is the average of the (13) roots, r is
the average of the (23)products of pairs of roots, and
r is the ‘average’ of the (33)
product of triplets of roots. With these, the earlier-mentioned
Cardano-Tartaglia cubic formula can be computed easily via
E=r,V=2(r2−r),and
W=2r3−3rr+r. In general we write
fD(x)=i+j=D0≤i,j≤D∑(−1)i(ijD)rixj where
ri denotes that quasi-binomial
parameter with i-many bars, (ijD)=i!j!D!, and we write
R⇀ for the ordered family (r,r,r,…). Notice that
f′(x)=3⋅a(x2−2rx+r)is just 3
times the quasi-binomial representation of a generic quadratic. This is
typical: with quasi-binomial representations of polynomials,
fD′=D⋅fD−1, i.e. taking the derivative coincides with truncating the highest order term from the quasi-binomial family R⇀; in other words, for a given quasi-binomial parameter family R⇀, we obtain a finite Appell sequence of derived functions. Also, in terms of the
quasi-binomial parameters, the statistically-presented version of the
quadratic formula is simply
x=r±r2−r.
It remains instructive to establish that the slope of a cubic at its
inflection point is the negated mean of the slopes at the three roots, via the
route of brute force computation. We want to evaluate f′ at the
three roots
rk=E+ωkT+⇀+ω−kT−⇀(for k∈{0,1,2}): f′(rk)=3⋅a(rk2−2rrk+r). For this we will use
[TABLE]
Now averaging over k∈{0,1,2} we encounter convenient terms such
as (ω2⋅0+ω2⋅1+ω2⋅2)T+⇀2=(ω0+ω2+ω1)T+⇀2=0, and (ω0+ω1+ω2)T+⇀=0, and three copies of 1⋅T+⇀T−⇀. Hence we obtain
rk2=r2+2T+⇀T−⇀. From the definitions, we
further simplify 2T+⇀T−⇀=V. Of course rk=E=r. Together
we now have f′(rk)=3⋅a(E2+V−2rr+r)=3⋅a(−r2+2(r2−r)+r)=3⋅a(r2−r)=23aV. Whereas the inflection point occurs at x=r, we
also compute f′(r)=3⋅a(r2−2rr+r)=−23aV, thus establishing our goal.
Looking back, our efforts were aided here by the fact that
3ω0+ω1+ω2=0, something particular to the case of
cube roots– something that we won’t have available for general cases. What
we will have in the general case is computing means of powers of the roots,
and here is our key to higher degree proofs: we want to be able to express
powers of the roots in terms of the quasi-binomial parameters, i.e. in
terms of the mean elementary symmetric polynomials in the roots.
Exemplifying this in the case of 3 roots, consider:
[TABLE]
[TABLE]
For any (multi)set Z of numbers, the elementary symmetric polynomials on Z are defined as ei(Z)≡ei:=A⊆Z∣A∣=i∑z∈A∏z (for i=0…∣Z∣;e0=1);
e.g. if Z=R as above with ∣R∣=3, then e2 is the sum of all (23)
products of taking 2 roots at a time, and e1 is the sum of all
(13) ‘products’ of taking 1 root at a time. Also define the power sums on Z per pi(Z)≡pi:=z∈Z∑zi; p0:=∣Z∣. With these definitions, equation (1) above is equivalent to p2=∑rk2=e12−2e2. Such effort to
determine the sum of powers (or average of powers) is an established result,
the Girard (1629) and Waring (1762) formula
[2] which is very closely connected to Newton’s
identities (ca.
1687) [10]. Both are based upon the
fact that for a (multi)set Z with ∣Z∣=n it holds that
i+j=n0≤i,j≤n∑(−1)jpj(Z)ei(Z)=0. (Note that
ei(R)=(in)ri.) This relationship can either be solved for
the ei or for the pj. Solving for an ek one obtains Newton’s recursive
identities ek=k1i+j=k0<j≤k∑(−1)j+1pjei. Solving instead for pj, one obtains
the Girard-Waring formula. To state the Girard-Waring formula compactly,
Put n:={1,2,…,n}
and
consider the n-tuples of natural numbers,
κ⇀≡(ki)i∈n∈Nn (0∈N), and define
\mathcal{K}_{{}_{\boldsymbol{\mathfrak{n}}j}}:=\left\{\overset{\rightharpoonup}{\kappa}\in\mathbb{N}^{n}\,\Big{|}\,\,\left\|\overset{\rightharpoonup}{\kappa}\right\|=j\right\} with
κ⇀:=i∈n∑iki, with n being the number of data (or root)
values. Note that for n≥j,Knj is isomorphic to Kjj=:Kj which is the set of integer
partitions of j; the only difference (which is insignificant) is that elements of Knj may have trailing zeroes in the partition. E.g., (3,0,0,1,0,0,0)∈K7 denotes the partition 1+1+1+4 of 7. The Girard-Waring formula is that pj=κ⇀∈Kj∑γκ⇀i∈n∏eiki, where γκ⇀:=j(−1)j∣κ⇀∣(−1)∣κ⇀∣(κ⇀∣κ⇀∣), where
κ⇀:=i∈n∑ki and where
(κ⇀∣κ⇀∣)=i∈n∏ki!∣κ⇀∣! is the multinomial coefficient over the family of subindices ki
in κ⇀. (See Gould
[2]). (The γκ⇀ are given as sequence A210258
in Sloane’s OEIS [12].) From our perspectives it
will become more natural to represent this relationship by replacing
pj and ei by their normalized forms pj=pj/n and
\overline{\,e_{{}_{i}}}\,=\,\overline{\,\,\,r}\!\!\!\!\!^{{}^{i}}\,\,\,\,=\,e_{{}_{i}}\big{/}\binom{n}{i}, admittedly superficial changes to
Girard-Waring. We thus obtain pj=κ⇀∈Kj∑cκ⇀i∈n∏riki where
[TABLE]
We emphasize that for any partition of the power j, i.e. ∀(ki)i∈n∈Kj, the coefficient of the corresponding term i∈n∏riki is given by this formula for cκ⇀. By holding fixed all but one of the many parameters in equation (2), one obtains many sequences of coefficients. E.g., the coefficients of tℓt2 for a family of n=3 gives sequence (1,7,36,162,…) (OEIS A080420 [12]) with formula 18j(j−5)3j−5 (degree j=ℓ+6). E.g., the coefficients of uℓu for a family of n=4 gives sequence (3,18,96,480,…) (not in OEIS) with formula 6j4j−3 (degree j=ℓ+2).
Tabulated results are in Tables GW.n=2 through
GW.n=7 and in Tables GW.deg=2 through
GW.deg=7, at the end of the paper. We particularly note that the coefficients in Table GW.n=2 are exactly the coefficients of the Tchebyshev polynomials of the first kind. Thus our Tables GW.n are generalizations of these Tchebyshev polynomials.
2 Example
An illustrative example is to consider the long-term goal of determining
relationships among φ4,0,ρ≡f4(R(ρ)) for
ρ∈{1,2,3}. A better example would address
φ6,0,ρ≡f6(R(ρ)) for
ρ∈{1,2,3,4,5}, since with f′ being a quintic we have no
hope of algebraically knowing those roots, but the larger example demands much
lengthier efforts. So begin with
f4(x)=a(x4−4rx3+6rx2−4rx+r). For ρ=3, we have the
cardinality 1 family R′′′=(r), thus
quickly we obtain
[TABLE]
For ρ=2, let the family of the roots of the 2nd derivative
be R′′=(s1,s2), and toward determining φ4,0,2≡f4(R(2)) let us consider
f4(si)=\linebreaka(si4−4rsi3+6rsi2−4rsi+r). From our mean
versions of the Girard-Waring formula (presented as Table GW.n=2 at the end of
the paper) we have:
s2=2s2−1s, s3=4s3−3ss, and s4=8s4−8s2s+1s2.
These give us
[TABLE]
We would be stuck here were it not for the fact that the si are roots
of a derivative of f, and thus we are blessed with the facts that
s=rands=r. This is the key issue enabling general degree success.
Thus we substantively extend Girard-Waring by considering
pj(Rf(m)) and
ei(Rf) where
Rf(m) is the roots family of the derivative
f(m). Simplifying our current degree 4 expression, we
now have
[TABLE]
For ρ=1, let the family of the roots of the 1st derivative
be R′=(t1,t2,t3), andtoward
determining φ4,0,1≡f4(R(1)) let us consider f4(ti)=\linebreaka(ti4−4rti3+6rti2−4rti+r). From our mean
versions of the Girard-Waring formula (Table GW.n=3) we have t2=3t2−2t, t3=9t3−9tt+t, and
t4=27t4−36t2t+4tt+6t2. These give us
[TABLE]
We would again be stuck here were it not for the fact that the ti are
roots of a derivative of f, and thus we are blessed with the facts that
t=r,t=r,and
t=r. Simplifying, we now have
[TABLE]
Henceforward we shall assume that the leading coefficient is a=1. We have
summarized our
(φD,0,ρ)ρ∈{1,2,…,D−1}
results for other degrees in Tables φ.2 through φ.7 at the end of the paper.
We summarize the above procedure. For a given order ρ of derivative,
consider the (D−ρ)roots of the derivative, and express the
normalized power sum means in terms of the derivative family’s quasi-binomial
parameters, making use of our normalized variant of the Girard-Waring
formulas. Then to evaluate the mean value of the original degree D
function over the derivative roots family, appropriately substitute these
mean power sum expressions in where the argument of the function occurs; this
takes advantage of the fact that averaging is a linear process, to wit, the
mean value of the polynomial function is the sum of the mean values of its
monomial components. Next, taking advantage of the fact that the
quasi-binomial parameters of a family of derivative roots is the same (albeit
truncated) as the quasi-binomial parameters of the degree D function, we
are able to simplify the expression of the φD,δ,ρ to be
entirely in terms of the degree D function’s quasi-binomial (i.e.
normalized symmetric function) parameters.
With such procedure in hand, we are enabled to determine
general degree D results ad libitum. Tabulated results are in Tables
φ.D for D=2 through D=7, at the end of the paper.
Let us look a bit more closely at the results thus far obtained in the above
example. We have:
[TABLE]
[TABLE]
[TABLE]
Observe that all of these have common terms
−4rr+1r, so there is more structure here than
what we have put our finger on. After sufficiently inspired inspection we
might realize the following unexpected relationship:
[TABLE]
Remarks**.**
(1)* We desire a more systematic way, with less inspiration
required, to obtain such relationships. In effect we are striving to solve
\alpha_{{}_{1}}\,\varphi_{{}_{4,0,1}}$$+{\alpha_{{}_{2}}\,\varphi_{{}_{4,0,2}}}+\alpha_{{}_{3}}φ4,0,3=0, where the φ4,0,ρ are
“vectors” of linear combinations of the five “basis elements” (r4,r2r,rr,r2,r). Thus the matter of finding all triplet
(α1,α2,α3)solutions is a simple
linear algebra issue involving row reduction.
(2)At first glance, trying to solve
\alpha_{{}_{1}}\,\varphi_{{}_{4,0,1}}$$+{\alpha_{{}_{2}}\,\varphi_{{}_{4,0,2}}}+\alpha_{{}_{3}}φ4,0,3=0 for the αρ seems like a
hopeless venture– for this degree 4 case, with 5 basis elements we are
essentially trying to solve 5 equations with only 3 degrees of freedom
in our αs. But the fact that all three of the
φ4,0,ρ have common terms
−4rr+1r saves us: if we can happen to
satisfy the other three basis elements with αs such
that ρ∈{1,2,3}∑αρ=0, then
the rr and r constraints will
automatically be satisfied. Besides that, there is further structure within
our φ4,0,ρ results– observe that the coefficients of both
r4 and rr2
are in the proportion φ4,0,1:φ4,0,2:φ4,0,3::9:8:3. This
again increases our hope of finding at least one
(α1,α2,α3)solution triplet.
(3) Observe that: (a) the coefficients within the
φD,δ,ρ are somewhat larger than the coefficients in the
eventual relationship
5\,\varphi_{{}_{4,0,1}}$$-{6\,\varphi_{{}_{4,0,2}}}+{1\text{
}\varphi_{{}_{4,0,3}}}=0; and (b) the number of basis elements inside each
φD,δ,ρ is greater than the number of means
φD,δ,ρ in our eventual relationship. These are
typical situations. Indeed, for relationships involving degrees up to 7,
we sometimes see coefficients within some φD,δ,ρ in the
tens of thousands, with as many as 15 basis elements, yet the eventual
relationships remain small, with coefficients near 100, and with 2 to 6φD,δ,ρ means involved. The number of basis elements
for different degree polynomials is the number of integer partitions of
the degree (see OEIS A000041 [12]):*
[TABLE]
*This all strongly suggests that our proof procedure is unnecessarily
convoluted, and that some more natural and simple proof path is yet to be
found. Streamlined proofs still elude us.
(4) Having noted in the chart above how quickly the number of
basis elements increases with increasing degree, we should be doubtful about
the prospects of finding
(α1,α2,…,αD−1) solution
tuples for higher degree D cases. Our hope rests in there being “special
circumstances” similar to those noted in Remark (2)above.
We shall find that such circumstances do hold.
Before we give a resum\a’e of our complete results, we feel compelled to
present the striking degree 6 result: There is a unique solution
to ρ∈{1,…,5}∑αρφ6,0,ρ=0, namely
[TABLE]
Perhaps equally surprising is that in degree 7, even though constraints
from 15 basis elements must be satisfied, we enjoy a 2-dimensional
family of solution relationships.
The graphical example had us consider f′(R) which is the average of the slopes at the roots, and this quantity is equal to the negation of twice the slope at the average of the roots. Note that this latter quantity is invariant w.r.t. vertical translations, hence the average of the slopes at the roots does not depend upon the polynomial’s constant term. This is typical:
Proposition 1**.**
If a horizontal line cuts a polynomial graph at as many points as the degree of the polynomial, the average of the slopes at the points of intersection is invariant w.r.t. modest vertical translations of the line.
Proof.
(Sketch) Given a vertical translation Δh, one can compute changes in the various slopes, ignoring terms that are higher order in Δh. Straightforward algebra shows that the sum of all such slope changes is [math] for degrees 2 through 7, and we are confident that the same method works in all degrees.
∎
The interpretation of ”modest” in the proposition is that the vertical translation should not cross an extremum. In fact when one accounts for multiplicities and computes, if necessary, complex-valued derivatives for complex-value roots, it is seen that the modesty condition is not a requirement.
Our attempts at an inductive proof of the general case above led to the following confident conjecture (supported by strong numerical evidence). The general proof of proposition 1 would follow as a corollary of the k=2 case of the following. Note that the conjecture gives statements in (reciprocal root units)k−1:
Conjecture 1**.**
*In any polynomial of degree ≥2 with roots set R having no repeated roots, it holds that
[TABLE]
Note that the above can be reinterpreted as: ∀ℓ≥1, the sum of relative rates is s∈Sc∑g(s)g(ℓ)(s)=0 where Sc is the roots family of any antiderivative (∫0xg)+c provided that Sc has no repeated roots and that deg(g)≥1.
We now present some of our results. More completely see Tables α.4 through
α.6, as well as Tables φ.2 through φ.7, at the end of the
paper.
Proposition 2**.**
The following is an exhaustive list of fundamental linear
relationships for degrees up to 8, among means of a polynomial’s values
when evaluated at roots of derivatives.
Notes:
The alphabetical labels at the right margins in the preceding & following
propositions indicate “derivative inheritance” relationships across
different degrees.
In all the cases in the above list, the sum of all positive
coefficients equals the sum of all negative coefficients. We have computed such relations up to degree 49, observing that (1) in all even degrees beyond 2 there is a single such fundamental linear relationship, and (2) in all odd degrees beyond 3 there is a 2-dimensional family of fundamental linear relationships. Moreover, the odd degrees enjoy the following clean relationship, which we regard as a main computational result with full confidence.
Conjecture 2**.**
For all odd degree polynomials of degree D, it holds that
[TABLE]
There are two ways to augment the above fundamental relationships with more
relationships: (a) instead of considering values of the function f we can
consider values of its derivatives f(n) or its (repeated)
antiderivatives, which we denote as f(−n); or (b) we can
expand the list of root families available by considering the root familes of
antiderivatives f(−n). When using antiderivatives, it
turns out that there is no dependence on the constant of integration; this is a consequence of proposition 1.
Proposition 3**.**
Expressing polynomials as their quasi-binomial
representations, ∫fD=D+11fD+1 where
the constant term of fD+1is the constant of integration c. In
other words,
R(−1)⇀=R⇀⊎(c). Restated yet again, if S⇀ is the
quasi-binomial parameter vector of any antiderivative of fR,
then
R⇀≤S⇀, i.e.
R⇀ is a (strict) subvector of
S⇀ via extension by the
constant(s) of integration. Further, if the domain variable of the
polynomial carries dimensional units, then the constant c is
(D+1)-dimensional.
Dually, if S⇀ is the
quasi-binomial parameter vector of any derivative of fR, then
S⇀≤R⇀, i.e.
S⇀ is a (strict) subvector of
R⇀ via truncation.
Example: Consider any polynomial fD,
e.g.f3(x)=x3−3x2r+3x1r−r with roots family R, and consider averaging
the values of the function over some set Z, x∈Z. We obtain
f3(Z)=x3−3x2r+3x1r−r. The average values
of the powers xm are obtained by our mean-value modification of
the Girard-Waring formulas. The case of Z being the roots
family of a derivative of f is straightforward, as noted in the dual statement in the
proposition, with Z⇀ being a
subvector of R⇀ via truncation.
The case of Z being the roots family of an antiderivative of f
requires greater attention: In the example here, the highest power term
present in f3(x) is order 3 (& dimensionality 3 if x bears units), but the
constant of integration in ∫f is of dimensionality 4; hence the
constant of integration cannot enter into the computation of
x3, etc., with similar behavior in the general situation.
Indeed, the values of x3, etc., only depend upon the entries in
Z⇀ that were already present in
R⇀. In the detail of the
present example we have, following our notation for a family of 4,
Z=F4≡U, u=r, u=r and u=r; thus
f3(Z)=1u3−3u2r+3u1r−1r
=1{16u3−18uu+3u}−3{4u2−3u}r+3{u1}r−r
[TABLE]
Note, in particular, that the average value
φ3,0,−1≡f3(R(−1))does not at all involve the constant of integration that affects
the family R(−1). This behavior is typical. Let
us emphasize: in the computation of mean function values
fD(R(−m)), individual
mean powers such as uk only involve the quasi-binomial
parameters of the original R⇀ and
never involve the constant(s) of integration of
R(−m). Thus we have:
Proposition 4**.**
The mean quantities φD,0,−m≡fD(R(−m)) and φD,δ,−m≡fD(δ)(R(−m)) (with δ,m>0) do not depend on
constant(s) of integration.
Proposition 5**.**
φD,1,0=D⋅φ(D−1),0,−1**
[TABLE]
Of course, the above proposition can be read in reverse as
givingφD,0,−1=(D+1)1⋅φ(D+1),1,0and
φD,0,−ρ=(D+ρ)!D!⋅φ(D+ρ),ρ,0(with ρ>0). These
facts are substantiated in our tables of φD,0,ρ expressions
by the fact that whenever ρ<0, the highest order quasi-binomial
parameter ri is always only order
D, never an order in excess of D.
For the following five propositions, see also Tables α.4 through
α.6.
Proposition 6**.**
The following is a partial list of linear relationships among
means of a polynomial’s derivative values when evaluated at roots of
derivatives.
[TABLE]
[TABLE]
*C. Degree(f)=4
1φ4,1,0+2φ4,1,2=0*
1φ4,1,0+2φ4,1,3=0**
1φ4,1,2−1φ4,1,3=0*
*D. Degree(f)=5
5φ5,1,2−6φ5,1,3+1φ5,1,4=0
E. Degree(f)=6
1φ6,1,2−3φ6,1,4+2φ6,1,5=0*
2φ6,1,3−5φ6,1,4+3φ6,1,5=0**
3φ6,1,2−4φ6,1,3+1φ6,1,4=0**
5φ6,1,2−6φ6,1,3+1φ6,1,5=0*
1φ6,1,2−2φ6,1,3+2φ6,1,4−1φ6,1,5=0*
(any 3 of the above are linearly dependent; any 2 are independent)*
Proposition 7**.**
The following is a partial list of linear relationships among
means of a polynomial’s 2nd derivative values when evaluated at
roots of derivatives.
[TABLE]
*C. Degree(f)=4
1φ4,2,0−2φ4,2,1=0
1φ4,2,1+1φ4,2,3=0
1φ4,2,0+2φ4,2,3=0
1φ4,2,0−1φ4,2,1+1φ4,2,3=0
D. Degree(f)=5
[TABLE]
*E. Degree(f)=6
5φ6,2,3−6φ6,2,4+1φ6,2,5=0
Proposition 8**.**
The following is a partial list of linear relationships
among means of a polynomial’s 3rd derivative values when
evaluated at roots of derivatives.
[TABLE]
*D. Degree(f)=5
[TABLE]
E. Degree(f)=6*
*
1φ6,3,0+9φ6,3,4=0**
1φ6,3,1+5φ6,3,3=0**
1φ6,3,2+2φ6,3,4=0**
1φ6,3,4−1φ6,3,5=0**
Proposition 9**.**
The following is a partial list of linear relationships
among means of a polynomial’s 4th derivative values when
evaluated at roots of derivatives.
[TABLE]
E. Degree(f)=6
3φ6,4,0−4φ6,4,1=0**
1φ6,4,1−3φ6,4,3=0**
1φ6,4,2−2φ6,4,3=0**
1φ6,4,3+1φ6,4,5=0**
Proposition 10**.**
*The following is a partial list of linear relationships
among means of a cubic polynomial’s function, derivatives, and/or
antiderivatives values when evaluated at roots of derivatives and
antiderivatives.
Degree(f)=3:
2φ3,−2,0* −5φ3,−2,1+3φ3,−2,2=0*
1φ3,−2,−1−3φ3,−2,1+2φ3,−2,2=0**
5φ3,−1,0−6φ3,−1,1+1φ3,−1,2=0**
1φ3,0,−1+2φ3,0,1=0**
1φ3,0,−2+5φ3,0,1=0**
1φ3,0,−3+9φ3,0,1=0**
1φ3,1,−1−2φ3,1,0=0**
1φ3,1,−2−3φ3,1,0=0**
3φ3,1,−3−4φ3,1,−2=0**
The last six of the above deserve particular comment. In the middle three
above, in the first terms (with coefficient 1) the cubic function f is
being averaged over the root families R(−1),R(−2),R(−3), which are
the root families of ∫f,∬f,and∭f. As noted in an
earlier proposition, these results are independent of choice of constants of
integration.
The following essentially restate above results with a different perspective.
See Tables α.4 through α.6,
Note: Regarding the alphabetical labels to the right of some relationships, we have used double-letter hybrid
labels to indicate that the relationship is linearly dependent upon the
denoted pair of preceding relationships, in cases of worthy of attention since the relationship coefficients vector is highly symmetric,
Proposition 11**.**
In any quartic polynomial the following relations hold:
[TABLE]
Proposition 12**.**
In any quintic polynomial the following relations hold:
[TABLE]
Also note that item (0f) states that for any quintic f,
the average of the function values at the two roots of f′′′ equals the
2:1 weighted average of (i) the value of the function at the
sole root of f(4) and (ii) the average of the
value of the function at the four roots of f′. Item
(0g) states that for any quintic f, the average of the
function values at the two roots of f′′′ also equals the 3:2 weighted
average of (i) the value of the function at the sole root of
f(4) and (ii) the average of the value of the
function at the three roots of f′′. We leave the remaining
interpretations to the reader.
Proposition 13**.**
In any sextic polynomial the following relations hold:
[TABLE]
Proposition 14**.**
In any septic polynomial the following relations hold:
[TABLE]
We find these relationships
distinctively impressive. There are many more identities, not recorded above, that are present when we
open up the limitless world of roots R(−n) of
f(−n):=ncopies∬⋯∫f, i.e. when we consider the limitless world of mean values
φD,δ,ρ where ρ<0. See Tables α.4 through
α.6.
Ancillary Computational Results
Consider the following observations regarding the coefficients of rmax≡rD in φD,0,ρ. Let hD(n) be the polynomial in n predicting the coefficient of rD for the φ.D family (φD,0,ρ)ρ<D−1. This polynomial hD(n) is always of the form hD(n)=D!(−1)DDρnχgD(n) where gD(n) is a monic irreducible (over Z) polynomial of degree M:=D−(2+χ) with integer coefficients, where χ=1 when D is odd and χ=0 when D is even.
Let gD(n)=k=2…D∑tk(D)nD−(k+χ). Note that t2(D)≡1 (monic condition); also for odd k, it holds that tk(D)=0 & tk+1(D)=0 if D≤k. The tk(D) are of the form Qk1uk(D) where Qk is the least common denominator of the terms in tk(D), such that uk(D) is a (generally not monic) polynomial in D with integer coefficients, with leading coefficient denoted as Nk.
Curiously, denominators Qk are the OEIS [12] sequence A053657, described in OEIS as “Denominators of integer-valued polynomials on prime numbers (with degree n): 1/a(n) is a generator of the ideal formed by the leading coefficients of integer-valued polynomials on prime numbers with degree less than or equal to n”, or equivalently “Also the least common multiple of the orders of all finite subgroups of GLn(Q) [Minkowski]”. Strikingly, the leading coefficients Nk of the polynomials uk(D) are coefficients of Nørlund’s polynomials, see OEIS A260326 (which had listed only 7 values previous to our work), described there as “Common denominator of coefficients in Nørlund’s polynomial D_2n(x).” [The OEIS citation refers to Nørlund’s 1924 book [8] (which discusses the higher order Bernoulli & Euler polynomials), Table 6 (p. 460), which only lists 7 values, for even indices 0 through 12, but Table 5 (p. 459) includes both even & odd entries as the leading coefficients of the primary components of the numerator polynomials. Our work has produced 198 terms with values eventually exceeding 5.8×1040.]
This paper is a spin-off from, and is material to, our recent
Insights Via Representational Naturality: New Surprises Intertwining Statistics, Cardano’s Cubic Formula,
Triangle Geometry, Polynomial Graphs [13].
Here is a summary of our tabulated results.
Tables φ.D: Function value means φD,δ,ρ in
terms of the polynomial representation parameters ri. These tables
extend the Girard-Waring formula.
Tables α.D: Linear Relationship Coefficients
αD,δ,ρ such that ρ∑αD,δ,ρφD,δ,ρ=0. Some
(D,δ,ρ) triplets admit more than one
(αD,δ,ρ) family.
Tables GW.n=2 through GW.n=7:
Normalized Girard-Waring coefficients cκ⇀:=nj(−1)j∣κ⇀∣(−1)∣κ⇀∣(κ⇀∣κ⇀∣)i∈n∏(in)ki where j= degree
and where κ⇀≡⟨k1,k2,…,kn⟩ is an integer partition vector of j, i.e.
∑iki⋅i=j.
Tables GW.deg=2 through GW.deg=7:
Normalized Girard-Waring coefficients cκ⇀:=nj(−1)j∣κ⇀∣(−1)∣κ⇀∣(κ⇀∣κ⇀∣)i∈n∏(in)ki as above.
Supportive Results: Function Value Means φDδρ
in Terms of Polynomial Representation Parameters ri .
To present our results, let
S:={s1,s2}=Rf(D−2) where D is the degree of f;
letT:={t1,t2,t3}=Rf(D−3); let U:=Rf(D−4) etc. We have the
following results.
N.B.: The coefficients in Table GW.n=2 are exactly the coefficients of Tchebyshev polynomials of the first kind. Thus these tables generalize Tchebyshev polynomials.
Table GW.n=2
Sum of Positive
Coefficients
Data Family F2≡S={s1,s2}
1:s1=s1
2:s2=2s2−1s
4:s3=4s3−3ss
9:s4=8s4−8s2s+1s2
21:s5=16s5−20s3s+5ss2
50:s6=32s6−48s4s+18s2s2−s3
120:s7=64s7−112s5s+56s3s2−7ss3
289:s8=128s8−256s6s+160s4s2−32s2s3+1s4
Table GW.n=3
Sum of Positive
Coefficients
Data Family F3≡T={t1,t2,t3}
1:t1=t1
3:t2=3t2−2t
10:t3=9t3−9tt+t
37:t4=27t4−36t2t+4tt+6t2
141:t5=81t5−135t3t+15t2t+45tt2−5tt
541:
t6=243t6−486t4t+54t3t+243t2t2−
36ttt−18t3+1t2
2080:
t7=729t7−1701t5t+189t4t+1134t3t2−
189t2tt−189tt3+7tt2+21t2t
Table GW.n=4
Sum of Positive
Coefficients
Data Family F4≡U={u1,u2,u3,u4}
1:u1=u1
4:u2=4u2−3u
19:u3=16u3−18uu+3u
98:u4=64u4−96u2u+16uu+18u2−u
516:
u5=256u5−480u3u+80u2u+180uu2−
30uu−5uu
2725:
u6=1024u6−2304u4u+384u3u+1296u2u2−
288uuu−108u3−24u2u+9uu+12u2
14400:
u7=4096u7−10752u5u+1792u4u+
112⋅72u3u2−2016u2uu−1512uu3+
112uu2+252u2u−112u3u+84uuu−7uu
Table GW.n=5
Sum of Positive
Coefficients
Data Family F5≡V={v1,v2,v3,v4,v5}
1:v1=v1
5:v2=5v2−4v
31:v3=25v3−30vv+6v
205:v4=125v4−200v2v+40vv+40v2−4v
1376:
v5=625v5−1250v3v+250v2v+500vv2−
100vv−25vv+1v
9251:
v6=3125v6−7500v4v+1500v3v+
4500v2v2−1200vvv−400v3−
150v2v+60vv+60v2+6vv
62210:
v7=15625v7−43750v5v+8750v4v+35000v3v2−
10,500v2vv−7000vv3+700vv2+1400v2v−
875v3v+700vvv−70vv+35v2v−14vv
Table GW.n=6
Sum of Positive
Coefficients
Data Family F6≡W={w1,w2,w3,w4,w5,w6}
1:w1=w1
6:w2=6w2−5w
46:w3=36w3−45ww+10w
371:w4=216w4−360w2w+80ww+75w2−10w
3026:
w5=1296w5−2700w3w+600w2w+1125ww2−
250ww−75ww+5w
24707:
w6=7776w6−19440w4w+4320w3w+
12150w2w2−3600www−1125w3−
540w2w+225ww+200w2+36ww−1w
201748:
w7=46656w7−136080w5w+30240w4w+
113400w3w2−37,800w2ww−23625ww3+
2800ww2+5250w2w−3780w3w+
3150www−350ww+252w2w−105ww−7ww
Collated, instead, along powers, the same information is:
**Table GW.deg=2
**
∣Fn∣
Sum of Positive
Coefficients
Quadratic
2
2
s2=2s2−1s
3
3
t2=3t2−2t
4
4
u2=4u2−3u
5
5
v2=5v2−4v
66w2=6w2−5w
**Table GW.deg=3
**
∣Fn∣
Sum of (+)
Coefficients
Cubic
2
4s3=4s3−3ss
3
10t3=9t3−9tt+t
4
19u3=16u3−18uu+3u
5
31v3=25v3−30vv+6v
646w3=36w3−45ww+10w
**Table GW.deg=4
**
∣Fn∣
Sum of (+)
Coefficients
Quartic
2
9s4=8s4−8s2s+1s2
3
37t4=27t4−36t2t+4tt+6t2
4
98u4=64u4−96u2u+16uu+18u2−u
5
205v4=125v4−200v2v+40vv+40v2−4v
6371w4=216w4−360w2w+80ww+75w2−10w
**Table GW.deg=5
**
∣Fn∣
Sum of (+)
Coefficients
Quintic
2
21s5=16s5−20s3s+5ss2
3
141t5=81t5−135t3t+15t2t+45tt2−5tt
4
516
u5=256u5−480u3u+80u2u+
180uu2−30uu−5uu
5
1376
v5=625v5−1250v3v+250v2v+
500vv2−100vv−25vv+1v
63026
w5=1296w5−2700w3w+600w2w+
1125ww2−250ww−75ww+5w
**Table GW.deg=6
**
∣Fn∣
Sum of (+)
Coefficients
Sextic
2
50s6=32s6−48s4s+18s2s2−s3
3
541
t6=243t6−486t4t+54t3t−
243t2t2−36ttt−18t3
4
2725
u6=1024u6−2304u4u+384u3u+
1296u2u2−288uuu−
108u3−24u2u+9uu+12u2
5
9251
v6=3125v6−7500v4v+1500v3v+
4500v2v2−1200vvv−400v3−
150v2v+60vv+60v2+6vv
624707
w6=7776w6−19440w4w+4320w3w+
12150w2w2−3600www−1125w3−
540w2w+225ww+200w2+36ww−1w
**Table GW.deg=7
**
∣Fn∣
Sum of (+)
Coefficients
Septic
2
120s7=64s7−112s5s+56s3s2−7ss3
3
2080
t7=7294t7−1701t5t+189t4t+1134t3t2−
189t2tt−189tt3+7tt2+21t2t
4
14400
u7=4096u7−10752u5u+1792u4u+
112⋅72u3u2−2016u2uu−1512uu3+
112uu2+252u2u−112u3u+84uuu−7uu
5
62210
v7=15625v7−43750v5v+8750v4v+35000v3v2−
10,500v2vv−7000vv3+700vv2+1400v2v−
875v3v+700vvv−70vv+35v2v−14vv
6201748
w7=46656w7−136080w5w+30240w4w+
113400w3w2−37,800w2ww−23625ww3+
2800ww2+5250w2w−3780w3w+
3150www−350ww+252w2w−105ww−7ww
ACKNOWLEDGMENTS
thanks to all…
Bibliography13
The reference list from the paper itself. Each links out to its DOI / PubMed record.
11. E. Barbeau, Polynomials, Springer-Verlag, 1989.
22. H. Gould, The Girard-Waring Power Sum Formulas for Symmetric Functions and Fibonacci Sequences, Fibonacci Quaterly, 37(2) (1999), 135-140.
33. G. Keady, The Zeros and the Critical Points of a Polynomial: Second Moments and Least Squares Fits of Lines, 2010, accessed 02/17/2014 as https://www.researchgate.net/profile/Grant_Keady/publications/
44. G. Keady, Lines of Best Fit for the Zeros and for the Critical Points of a Polynomial, American Mathematical Monthly, 118 (March 2011), 262-264.
55. M. Marden, Geometry of Polynomials, Mathematical Surveys No. 3, American Mathematical Society, Providence, RI, 1966.
66. M. Marden, Geometry of the Zeros of a Polynomial in a Complex Variable, American Mathematical Society, New York, 1949.
77. D. Minda and S. Phelps, Triangles, Ellipses, and Cubic Polynomials, American Mathematical Monthly, 115 (2008), 679-689.
88. N. Nørlund, Vorlesungen über Differenzenrechnung, vol. 13 in Die Grundlehren der mathematischen Wissenschaften, Verlag von Julius Springer, Berlin, 1924.