1: %Ch4.tex
2: %
3: % These are the notes for Lecture 3 of my IHP course
4: %
5: % Frank Sottile
6: % 26 October 2005
7: % Muenchen
8: %
9: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
10: %
11: \chapter{Upper Bounds}\label{Ch:Khov}
12: %
13: %\Red{{Wednesday, 16 November, 10:45--12:15}}
14: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
15:
16: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
17: %
18: %\section*{Reprise: Descartes's Rule}
19: %
20: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
21:
22: Recall Descartes rule of signs~(\cite{D1637} or Section~\ref{S2:Descartes}), which gives a
23: bound for the number of positive solutions to a univariate polynomial.\medskip
24:
25:
26: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
27: \noindent{\bf Theorem~\ref{T2:Descartes} (Descartes's rule of signs)}
28: {\it
29: The number, $r$, of positive roots of a univariate polynomial
30: %
31: \begin{equation}\label{E4:Univariate}
32: f(x)\ =\ c_0 x^{a_0} + c_1 x^{a_1} + \dotsb + c_m x^{a_m}\,,
33: \end{equation}
34: %
35: counted with multiplicity, is bounded above by the number of variations in sign of the
36: coefficients of $f$,
37: \[
38: \#\{i\mid 1\leq i\leq m\mbox{ and } c_{i-1} c_i<0\}
39: \ \leq\ r\,,
40: \]
41: and the difference between the variation and $r$ is even.\medskip
42: }
43: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
44:
45: In~\eqref{E4:Univariate} we assume that $a_0<a_1<\dotsb<a_m$ and no
46: coefficient $c_i$ vanishes.
47:
48: Thus a univariate polynomial with $m+1$ monomials has at most
49: $m$ positive roots.
50: This bound is sharp, as the polynomial
51: %
52: \begin{equation}\label{E4:Descartes}
53: (x-1)(x-2)\dotsb (x-m)
54: \end{equation}
55: %
56: has $m+1$ distinct terms and $m$ positive roots.
57: Replacing $x$ by $x^2$ gives a polynomial with $m{+}1$ terms and $2m$ nonzero real roots.
58:
59: This chapter and the next will discuss
60: extensions of this Descartes bound to systems of multivariate polynomials.
61:
62: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
63: %
64: \section{Khovanskii's fewnomial bound}
65: %
66: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
67:
68: Descartes's rule of signs suggests that the number of real roots to a system of
69: polynomials depends not on its degree, but rather on the complexity of its
70: description.
71: D.~Bernstein and A.~Kushnirenko formulated the principle that the topological complexity
72: of a set in $\R^n$ defined by real polynomials is controlled by the complexity of the
73: description of the polynomials, rather than by their degree or Newton polytopes.
74: This is exactly what Khovanskii found in 1980 with his celebrated
75: \DeCo{{\sl fewnomial bound}}.
76:
77: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
78: \begin{thm}[Khovanskii~\cite{Kh80}]\label{T4:Fewnomial}
79: A system of $n$ real polynomials in $n$ variables involving $l{+}n{+}1$ distinct
80: monomials will have at most
81: %
82: \begin{equation}\label{E4:Fewnomial_Bound}
83: 2^{\binom{l+n}{2}}\cdot (n+1)^{l+n}
84: \end{equation}
85: %
86: nondegenerate positive solutions.
87: \end{thm}
88: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
89:
90: We remark that nondegenerate solutions are isolated, and there are finitely
91: many of them.
92: This bound, like other bounds in this part of the subject,
93: considers solutions in the \DeCo{{\sl positive orthant} $\R_>^n$}.
94: A consequence of Khovanskii's bound is that for each $l$ and $n$, there is a number
95: \DeCo{$X(l,n)$} which is equal to the maximum number of positive solutions to a system of
96: $n$ polynomials in $n$ variables having $l{+}n{+}1$ distinct monomials.
97: A central question in this area is to determine the \DeCo{{\sl Khovanskii number}}
98: $X(l,n)$ exactly, or give good bounds.
99: Khovanskii's Theorem shows that $X(l,n)$ is bounded above by the
100: quantity~\eqref{E4:Fewnomial_Bound}.
101:
102: A complete proof of Theorem~\ref{T4:Fewnomial} may be found in Khovanskii's
103: book~\cite{Kh91}, where much else is also developed.
104: Chapter 1 of that book contains an accessible sketch.
105: Benedetti and Risler~\cite[\S 4.1]{Benedetti_Risler} have a
106: careful and self-contained exposition of Khovanskii's fewnomial
107: bound.
108: We give a sketch of the main ideas in the exposition of
109: Benedetti and Risler, to which we refer for further details
110: (this is also faithful to Khovanskii's sketch).
111: We remark that Sturmfels has also sketched~(\cite[pp.~39--40]{SPE} and
112: in~\cite{St98}) a version of the proof.
113: This omits some contributions to the root count and is therefore regrettably incorrect.
114:
115: Khovanskii looks for solutions in the positive orthant $\R_>^n$, proving a far
116: more general result involving solutions in $\R^n$ of polynomial functions in logarithms of
117: the coordinates and monomials.
118: For this, he took logarithms of the coordinates.
119: Set
120: %
121: \begin{equation}\label{E4:substitution}
122: \DeCo{z_i}\ :=\ \log(x_i)
123: \qquad\mbox{and}\qquad
124: \DeCo{y_j}\ :=\ e^{z\cdot a_j}\ =\ x^{a_j}\,,
125: \end{equation}
126: %
127: where $i=1,\dotsc,n$, $j=1,\dotsc,k$, and $a_j\in\R^n$ can be real
128: exponents.
129: Consider a system of functions of the form
130: %
131: \begin{equation}\label{Eq4:Khovanski_poly}
132: F_i(z_1,\dotsc,z_n,\, y_1,\dotsc,y_k)\ =\ 0
133: \qquad i=1,\dotsc,n\,,
134: \end{equation}
135: %
136: where each $y_j=y_j(z)$ is an exponential function $e^{z\cdot a_j}$
137: and the $F_i$ are polynomials in $n{+}k$ indeterminates.
138:
139: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
140: \begin{thm}[Khovanskii's Theorem]\label{T4:Kh_real_thm}
141: The number of nondegenerate real solutions to the
142: system~$\eqref{Eq4:Khovanski_poly}$ is at most
143: %
144: \begin{equation}\label{Eq4:Real_fewnomial bound}
145: \Bigl(\prod_{i=1}^n \deg F_i\Bigr) \cdot
146: \Bigl( 1 + \sum_{i=1}^n \deg F_i \Bigr)^k\cdot
147: 2^{\binom{k}{2}} \,.
148: \end{equation}
149: %
150: \end{thm}
151: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
152:
153: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
154: \noindent{\it Proof of Theorem~$\ref{T4:Fewnomial}$.}
155: Given a system of $n$ real polynomials in $n$ variables involving
156: $l{+}n{+}1$ distinct monomials, we may assume that one of the monomials is 1.
157: Under the substitution~\eqref{E4:substitution}, this becomes a
158: system of the form~\eqref{Eq4:Khovanski_poly}, where each $F_i$ is
159: a degree 1 polynomial in $k=l+n$ variables.
160: Then $\deg F_i=1$ and the bound~\eqref{Eq4:Real_fewnomial bound} reduces
161: to~\eqref{E4:Fewnomial_Bound}.
162: \QED\medskip
163: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
164:
165:
166: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
167: \noindent{\it Sketch of proof of Theorem~$\ref{T4:Kh_real_thm}$.}
168: We proceed by induction on $k$, skipping some technicalities involving Sard's Theorem.
169: When $k=0$, there are no exponentials, and the system is just a
170: system of $n$ polynomials in $n$ variables, whose number of
171: nondegenerate isolated solutions is bounded above by the B\'ezout
172: number,
173: \[
174: \prod_{i=1}^n \deg F_i\,,
175: \]
176: which is the bound~\eqref{Eq4:Real_fewnomial bound} when $k=0$.
177:
178: Suppose that we have the bound~\eqref{Eq4:Real_fewnomial bound} for
179: systems of the form~\eqref{Eq4:Khovanski_poly} with $k$ exponentials, and
180: consider a system with $k{+}1$ exponentials and one added variable \DeCo{$t$}.
181: %
182: \begin{eqnarray}
183: \label{Eq4:subsystem}
184: G_i(z,t)\ :=\ F_i(z_1,\dotsc,z_n,\; y_1,\dotsc,y_k,\,t\cdot y_{k+1})&=& 0\
185: \qquad i=1,\dotsc,n\\
186: \nonumber
187: t&=& 1
188: \end{eqnarray}
189: %
190: The subsystem~\eqref{Eq4:subsystem} defines an analytic curve
191: $C$ in $\R^{n+1}$, which we assume is smooth and transverse to
192: the hyperplane at $t=1$.
193:
194: Write $z_{n+1}$ for $t$ and consider the vector field $\xi$ in
195: $\R^{n+1}$ whose $r$th component is
196: %
197: \begin{equation}\label{Eq4:zeta_r}
198: \xi_r\ :=\ (-1)^{n+1-r}\det\left(\frac{\partial G_i}{\partial
199: z_j}\right)^{i=1,\dotsc,n}_{j=1,\dotsc,\hat{r},\dotsc,n+1}\ .
200: \end{equation}
201: %
202: This vector field is tangent to the curve $C$, and we write
203: $\xi_t=\xi_{n+1}$ for its component in the $t$-direction.
204: An important ingredient in our proof of Theorem~\ref{T4:Kh_real_thm}
205: is a special case of the Khovanskii-Rolle Theorem~\cite[pp.~42--51]{Kh91}.
206:
207: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
208: \begin{thm}[Khovanskii-Rolle Theorem]\label{T4:Kh-R}
209: The number of points of\/ $C$ where $t=1$ is bounded
210: above by
211: \[
212: N\ +\ q\,,
213: \]
214: where $N$ is the number of points of\/ $C$ where $\xi_t=0$ and
215: $q$ is the number of unbounded components of\/ $C$.
216: \end{thm}
217: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
218:
219: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
220: \noindent{\it Proof.}
221: Note that $\xi_t$ varies continuously along $C$.
222: Suppose that $a$ and $b$ are consecutive points along an arc of
223: $C$ where $t=1$.
224: Since $C$ is transverse to the hyperplane $t=1$, we have
225: $\xi_t(a)\cdot\xi_t(b)<0$, and so there is a point $c$ of $C$
226: on the arc between $a$ and $b$ with $\xi_t(c)=0$.
227:
228: The hyperplane $t=1$ cuts a compact connected component of $C$
229: into the same number of arcs as points where $t=1$.
230: Since the endpoints of each arc lie on the hyperplane $t=1$,
231: there is at least one point $c$ on each arc with $\xi_t(c)=0$.
232: Similarly, the hyperplane $t=1$ cuts a noncompact component into
233: arcs, and each arc with two endpoints in the hyperplane $t=1$ gives a
234: point $c$ with $\xi_t(c)=0$.
235: However, there will be one more point with $t=1$ on this component
236: than such arcs.
237: \QED\medskip
238: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
239:
240: We illustrate the argument in the proof below.
241: \[
242: \begin{picture}(220,160)(0,-7)
243: \put(0,0){\includegraphics[height=150pt]{figures/4/KhRo.eps}}
244: \put(197,76){$t=1$}
245: \put(72,144){$\xi_t=0$}
246: \put(68,147){\vector(-1,0){20}} \put(85.8,140){\vector(0,-1){17}}
247: \put(62,-8){$\xi_t=0$}
248: \put(59,-2.5){\vector(-4,1){15}} \put(95,-3.5){\vector(4,1){15}}
249: \put(198,130){$C$}
250: \end{picture}
251: \]
252:
253: The key to the induction in the proof of Khovanskii's formula is to replace the last
254: exponential by a new variable.
255: This substitution is omitted in Sturmfels's argument, which also does not use
256: the Khovanskii-Rolle Theorem~\ref{T4:Kh-R}.
257: Since we have
258: %
259: \begin{eqnarray*}
260: \frac{\partial G_i}{\partial z_r}
261: &=&
262: \frac{\partial F_i}{\partial z_r}
263: (z_1,\dotsc,z_n,\; y_1,\dotsc,y_k,t y_{k+1})\\
264: &&\ +\ \sum_{j=1}^k \frac{\partial F_i}{\partial y_j}
265: (z_1,\dotsc,z_n,\;y_1,\dotsc,y_k,t\, y_{k+1})\cdot
266: a_{j,r}\, y_j\\
267: &&\ +\ \frac{\partial F_i}{\partial y_{k+1}}
268: (z_1,\dotsc,z_n,\;y_1,\dotsc,y_k,t\, y_{k+1})\cdot
269: a_{k+1,r}\ t\, y_{k+1}\,,
270: \end{eqnarray*}
271: %
272: if we set $\DeCo{u}:= t y_{k+1}$ and define \DeCo{$\phi(z,u)$} to be the expression for
273: $\xi_t=\xi_{n+1}$~\eqref{Eq4:zeta_r} written in terms of $z$ and $u$, then
274: the total degree (in $z_1,\dotsc,z_n,y_1,\dotsc,y_k,\DeCo{u}$) of
275: $\phi_t(z,u)$ is at most $\sum_{i=1}^n \deg F_i$.
276:
277: Thus number $N$ of Theorem~\ref{T4:Kh-R} is the number of
278: solutions to the system
279: %
280: \begin{equation}\label{Eq4:Sys_A}
281: \begin{array}{rcl}
282: F_i(z_1,\dotsc,z_n,\; y_1,\dotsc,y_k,\,\DeCo{u})&=& 0\
283: \qquad i=1,\dotsc,n\\
284: \phi_t(z,\DeCo{u})&=& 0\,.\rule{0pt}{15pt}
285: \end{array}
286: \end{equation}
287: %
288: This has the form~\eqref{Eq4:Khovanski_poly} with $k$ exponentials.
289: Given any solution to the system~\eqref{Eq4:Sys_A}, we use the substitution $u=t
290: y_{k+1}$ to solve for $t$ and get a corresponding point $c$ on
291: the curve $C$ with $\xi_t(c)=0$.
292: We apply our induction hypothesis to the system~\eqref{Eq4:Sys_A} (which has $k$ exponentials
293: and $n{+}1$ equations in $n{+}1$ variables) to obtain
294: \[
295: N\ \leq \
296: \prod_{i=1}^n \deg F_i \cdot
297: \Bigl( \sum_{i=1}^n \deg F_i \Bigr)\cdot
298: \Bigl( 1 + 2\sum_{i=1}^n \deg F_i \Bigr)^k\cdot
299: 2^{\binom{k}{2}} \,.
300: \]
301:
302:
303:
304: We similarly estimate the number $q$ of noncompact components of
305: $C$.
306: We claim that this is bounded above by the maximum number of points of
307: intersection of $C$ with a hyperplane.
308: Indeed, since each noncompact component has two infinite branches,
309: there are
310: $2q$ points (counted with multiplicity) on the sphere $S^{n}$
311: corresponding to directions of accumulation points of the branches of $C$
312: at infinity.
313: Any hyperplane through the origin not meeting these points will have
314: at least $q$ of these points in one of the hemispheres into which it
315: divides the sphere.
316: If we translate this hyperplane sufficiently far toward infinity, it
317: will meet the branches giving these accumulation points, and thus
318: will meet $C$ in at least $q$ points.
319:
320: Thus $q$ is bounded by the number of solutions to a system of the form
321: %
322: \begin{equation}\label{Eq4:Sys_B}
323: \begin{array}{rcl}
324: F_i(z_1,\dotsc,z_n,\; y_1,\dotsc,y_k,\,\DeCo{u})&=& 0\
325: \qquad i=1,\dotsc,n\\
326: l_0+ l_1z_1 + l_2z_2 +\dotsb+l_nz_n + l_u \DeCo{u} &=& 0 \ ,\rule{0pt}{15pt}
327: \end{array}
328: \end{equation}
329: %
330: where $l_0,\dotsc,l_n,\,l_u$ are some real numbers.
331: This again involves only $k$ exponentials, and the last equation has degree 1, so we have
332: \[
333: q\ \leq\
334: \prod_{i=1}^n \deg F_i \cdot
335: 1\cdot
336: \Bigl(1 + \sum_{i=1}^n \deg F_i \ +1\Bigr)^k\cdot
337: 2^{\binom{k}{2}} \,.
338: \]
339: Combining these estimates gives
340: \[
341: N+q\ \leq\
342: \prod_{i=1}^n \deg F_i \cdot 2^{\binom{k}{2}} \cdot
343: \Bigl[
344: \Bigl( \sum_{i=1}^n \deg F_i \Bigr)\cdot
345: \Bigl( 1 + 2\sum_{i=1}^n \deg F_i \Bigr)^k
346: \ +\
347: \Bigl( 2 + \sum_{i=1}^n \deg F_i \Bigr)^k
348: \Bigr] \,.
349: \]
350:
351:
352: We can obtain a simpler (but larger) estimate by bounding $N+q$ by the number of
353: solutions to the single system of equations,
354: %
355: \begin{eqnarray*}
356: F_i(z_1,\dotsc,z_n,\; y_1,\dotsc,y_k,\,\DeCo{u})&=& 0\
357: \qquad i=1,\dotsc,n\\
358: \DeCo{F_{n+1}}\ :=\ (l_0+ l_1z_1 + l_2z_2 +\dotsb+l_nz_n + l_u \DeCo{u})
359: \cdot\phi_t(z,\DeCo{u}) &=& 0 \ .
360: \end{eqnarray*}
361: %
362: By our induction hypothesis, we have
363: \[
364: N+q\ \leq\
365: \prod_{i=1}^{n+1}\deg F_i \cdot
366: % \bigl( 1 + \sum_{i=1}^{n+1} \deg F_i \bigr)\cdot
367: \Bigl( 1 + \sum_{i=1}^{n+1} \deg F_i \Bigr)^k\cdot
368: 2^{\binom{k}{2}} \,.
369: \]
370: But we saw that $F_{n+1}$ has degree at most $\sum_{i=1}^n \deg F_i\ +1$, and so
371: the number, $M$, of solutions to the system with $k+1$
372: exponentials is bounded by
373: %
374: \begin{eqnarray*}
375: M \ \leq \ N+q&\leq& \prod_{i=1}^n \deg F_i \cdot
376: \Bigl(1+ \sum_{i=1}^n \deg F_i \Bigr)\cdot
377: \Bigl(1 + \sum_{i=1}^n \deg F_i +1+\sum_{i=1}^n \deg F_i\Bigr)^k \cdot 2^{\binom{k}{2}} \\
378: &=& \prod_{i=1}^n \deg F_i \cdot
379: \Bigl(1 + \sum_{i=1}^n \deg F_i \Bigr)^{k+1} \cdot 2^{\binom{k+1}{2}} \,,
380: \end{eqnarray*}
381: %
382: which completes the proof of Theorem~\ref{T4:Kh_real_thm}.
383: \QED\medskip
384:
385:
386: We see that the result of Theorem~\ref{T4:Kh_real_thm} is much more
387: general than the statement of Theorem~\ref{T4:Fewnomial}.
388: Also, the bound is not sharp.
389: While no one believed that Khovanskii's bound~\eqref{E4:Fewnomial_Bound} was anywhere near
390: the actual upper bound $X(l,n)$, it was been extremely hard to improve it.
391: We discuss the first steps in this direction.
392:
393:
394: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
395: %
396: \section{Kushnirenko's conjecture}
397: %
398: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
399:
400: One of the first proposals of a more reasonable bound than Khovanskii's for the number of
401: positive solutions to a system of polynomials was due to
402: Kushnirenko, and for many years experts believed that this may
403: indeed be the truth.
404:
405: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
406: \begin{conj}[Kushnirenko]\label{C4:Kushnirenko}
407: A system $f_1=f_2=\dotsb=f_n=0$ of real polynomials where each $f_i$ has
408: $m_i{+}1$ terms has at most $m_1m_2\dotsb m_n$ nondegenerate positive
409: solutions.
410: \end{conj}
411: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
412:
413: This generalizes the bound given by Descartes's rule of signs.
414: It easy to use the example~\eqref{E4:Descartes} for the sharpness of Descartes's
415: rule to construct systems of the form
416: \[
417: f_1(x_1)\ =\ f_2(x_2)\ =\ \dotsb\ =\ f_n(x_n)\ =\ 0\,,
418: \]
419: which achieve the bound of Conjecture~\ref{C4:Kushnirenko}.
420:
421: Soon after Kushnirenko made this conjecture, K.~Sevostyanov found a counterexample which was
422: unfortunately lost.
423: Nevertheless, this conjecture passed into folklore until Haas~\cite{Ha02}
424: found a simple example of two trinomials ($3=2+1)$ in variables $x$ and $y$ with
425: $5\,(>4=2\cdot 2)$ isolated nondegenerate positive solutions.
426: %
427: \begin{equation}\label{E4:Haas}
428: 10 x^{106}\ +\ 11 y^{53}\ -\ 11y\ =\
429: 10 y^{106}\ +\ 11 x^{53}\ -\ 11x\ =\ 0\,.
430: \end{equation}
431:
432: There have been other attempts to find better bounds than the
433: Khovanskii bound.
434: Sturmfels~\cite{St94} used the toric degenerations introduced in Chapter~\ref{Ch:sparse}
435: to show how to construct systems with many real roots (the root count
436: depends upon a mixture the geometry of the Newton polytopes and some
437: combinatorics of signs associated to lattice points)\fauxfootnote{Actually, he
438: used the toric proof of Bernstein's Theorem.}.
439: This inspired Itenberg and Roy~\cite{IR96} to propose a multivariate
440: version of Descartes's rule of signs, which was later found to be too
441: optimistic~\cite{LW98}.
442: An interesting part of this story is told in the cheeky paper of Lagarias and
443: Richardson~\cite{LR97}.
444:
445: More recently, Li, Rojas, and Wang looked closely at Haas's
446: counterexample to Kushnirenko's conjecture, seeking to obtain
447: realistic bounds for the number of positive solutions which depended only on the number of
448: monomials in the different polynomials.
449: For example, they showed that Haas's counterexample was the best
450: possible.
451:
452: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
453: \begin{thm}[Li, Rojas, and Wang~\cite{LRW03}]
454: A system consisting of two trinomials in two variables has at
455: most $5$ nondegenerate positive solutions.
456: \end{thm}
457: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
458:
459: Dickenstein, Rojas, Rusek, and Shih~\cite{DRRS} used exact
460: formulas for $\calA$-dis\-criminants~\cite{DFS} to study systems of two trinomials in two
461: variables which achieve this bound of five positive solutions.
462: They gave the following example, which indicates how difficult it is to find systems with
463: many real solutions.
464:
465: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
466: \begin{ex}
467: Consider the family of systems of bivariate sextics,
468: %
469: \begin{equation}\label{Eq:DRRS_system}
470: x^6\ +\ a y^3\ -\ y\ =\ y^6\ +\ bx^3\ -\ x\ =\ 0\,,
471: \end{equation}
472: %
473: where $a,b$ are real numbers.
474: When $a=b=78/55$, this has five positive real solutions
475: \[
476: (0.8136, 0.6065)\,,\
477: (0.7888, 0.6727)\,,\
478: (0.7404, 0.7404)\,,\
479: (0.6727, 0.7888)\,,\
480: (0.6065, 0.8136)\,.
481: \]
482: We now investigate the set of parameters $(a,b)$ for which this achieves the trinomial
483: bound of five positive solutions.
484: This turns out to be a single connected component in the complement of the discriminant
485: for this family of systems.
486: This discriminant is a polynomial of degree 90 in $a,b$ with 58 terms whose leading and
487: trailing terms are
488: %
489: \begin{multline*}
490: 1816274895843482705708030487016037960921088a^{45}b^{45} \ +\
491: \dotsb\ \mbox{56 terms}\ \dotsb\ +\ \\
492: 1102507499354148695951786433413508348166942596435546875.
493: \end{multline*}
494: %
495: We display this discriminant in the square $[0,4]\times[0,4]$, as well as three successive
496: magnifications, each by a factor of 11.
497: The shaded region in the last picture is the set of pairs $(a,b)$ for
498: which~\eqref{Eq:DRRS_system} achieves the trinomial bound of five positive real
499: solutions.
500: \[
501: \includegraphics[height=110pt]{figures/4/DRRS.1.eps}
502: \includegraphics[height=110pt]{figures/4/DRRS.2.eps}
503: \includegraphics[height=110pt]{figures/4/DRRS.3.eps}
504: \includegraphics[height=110pt]{figures/4/DRRS.4.eps}
505: \]
506: \end{ex}
507: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
508:
509: To compare the trinomial bound in~\cite{LRW03} to the fewnomial
510: bound~\eqref{E4:Fewnomial_Bound}, note that we may multiply one of the polynomials by a
511: monomial so that the two trinomials share a monomial.
512: Then there are at most $5=2+2+1$ distinct monomials occurring in the two trinomials.
513: The fewnomial bound for $l=n=2$ is
514: \[
515: X(2,2)\ \leq \ 2^{\binom{4}{2}}\cdot (2+1)^4\ =\ 5184\,.
516: \]
517: We remark that a trinomial system is not quite a general
518: fewnomial system with $l=n=2$.
519: Still, the bound of five real solutions lent credence to the belief $X(2,2)$ is closer to
520: five than to 5184 and that Khovanskii's fewnomial bound~\eqref{E4:Fewnomial_Bound} could be
521: improved.
522:
523: In addition to providing the counterexample to Kushnirenko's conjecture, Sevostyanov
524: also established the first result of fewnomial-type.
525: He showed the existence of an absolute bound $\sigma(d,m)$ for the number of real
526: solutions to a system
527: \[
528: f(x,y)\ =\ g(x,y)\ =\ 0\,,
529: \]
530: where $f$ is a polynomial of degree $d$ and $g$ has $m$ terms.
531: The proof of this result, like his counterexample, has unfortunately been lost.
532: This result however, was the inspiration for Khovanskii to develop his theory of
533: fewnomials.
534:
535: Recently, Avenda\~{n}o~\cite{Av}, established a precise version of a special case of
536: Sevostyanov's theorem.
537:
538: \begin{thm}
539: Suppose that $f(x,y)$ is linear and $g(x,y)$ has $m$ terms.
540: Then the system
541: \[
542: f(x,y)\ =\ g(x,y) =\ 0\,,
543: \]
544: has at most $6m-4$ real solutions.
545: \end{thm}
546:
547:
548:
549: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
550: %
551: \section{Systems supported on a circuit}\label{S4:circuit}
552: %
553: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
554: Restricting the analysis of Section~\ref{S3:simplex} to real solutions shows that
555: $X(0,n)=1$.
556: Recently, it was shown that $X(1,n)=1{+}n$.
557: We discuss this here.
558:
559:
560: A collection $\calA$ of $n{+}2$ vectors in $\Z^n$ which affinely spans $\R^n$ is
561: called a \DeCo{{\sl circuit}}.
562: The circuit is \DeCo{{\sl primitive}} if its $\Z$-affine span is all of $\Z^n$.
563: When $0\in\calA$, this means that $\Z^n=\Z\calA$.
564:
565: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
566: \begin{thm}[Bertrand, Bihan, and Sottile~\cite{BBS}]\label{T4:BBS}
567: A polynomial system supported on a primitive circuit has at most $2n+1$
568: nondegenerate nonzero real solutions.
569: \end{thm}
570: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
571:
572: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
573: \begin{thm}[Bihan~\cite{Bihan}]\label{T4:Bihan}
574: A polynomial system supported on a circuit has at most $n{+}1$
575: nondegenerate positive solutions, and there exist systems supported on a circuit having
576: $n{+}1$ positive solutions.
577: \end{thm}
578: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
579:
580: This can be used to construct fewnomial systems with relatively many positive
581: solutions.
582:
583: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
584: \begin{cor}[Bihan, Rojas, Sottile~\cite{BRS}]\label{C4:BRS}
585: There exist systems of $n$ polynomials in $n$ variables having $l{+}n{+}1$ monomials and
586: at least
587: $\lceil\frac{n}{l}\rceil^l$ positive solutions.
588: \end{cor}
589: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
590:
591:
592: This gives a lower bound for $X(l,n)$ of $\lceil\frac{n}{l}\rceil^l$, and is the best
593: construction when $l$ is fixed and $n$ is large.
594: It remains an open problem to give constructions with more solutions, or
595: constructions with many solutions when $l$ is not fixed.
596:
597: The construction establishing Corollary~\ref{C4:BRS} is quite simple.
598: Suppose that $n=kl$ is a multiple of $l$, and let
599: \[
600: f_1(x_1,\dotsc,x_k)\ =\ f_2(x_1,\dotsc,x_k)\ =\ \dotsb\ =\
601: f_k(x_1,\dotsc,x_k)\ =\ 0\,,
602: \]
603: be a system with $k{+}2$ monomials and $k{+}1$ positive solutions.
604: Such systems exist, by Theorem~\ref{T4:Bihan}.
605: Write $F(x)=0$ for this system and assume that one of its monomials is a constant.
606: For each $i=1,\dotsc,l$, let $y^{(i)}=(y_1^{(i)},\dotsc,y^{(i)}_k)$ be a set of $k$ variables.
607: Then the system
608: \[
609: F(y^{(1)})\ =\ F(y^{(2)})\ =\ \dotsb\ =\ F(y^{(l)})\ =\ 0\,,
610: \]
611: has $(k+1)^l$ solutions, $kl$ variables, and $1+l+kl$ monomials.
612: When $n=kl+r$ with $r<;$, adding extra variables $y_i$ and equations
613: $y_i=1$ for $i=1,\dotsc,r$ gives a system with $(\lfloor\frac{n}{l}\rfloor +1)^l$
614: positive solutions and $1+l+n$ monomials.\medskip
615:
616:
617: When $l=1$ the fewnomial bound~\eqref{E4:Fewnomial_Bound} becomes
618: \[
619: 2^{\binom{1+n}{2}} \cdot (n+1)^{1+n}\,,
620: \]
621: which is considerably larger than Bihan's bound of $n{+}1$.
622: Replacing $l{+}n$ by $n$ in the fewnomial bound, it becomes equal to Bihan's bound when
623: $l=1$.
624: When $l=0$, this same substitution in~\eqref{E4:Fewnomial_Bound} yields $1$, which is the
625: sharp bound when $l=0$.
626: In Chapter~\ref{Ch:Gale}, we will outline generalizations of Theorems~\ref{T4:BBS}
627: and~\ref{T4:Bihan} to arbitrary $l$, giving the bound
628: %
629: \begin{equation}\label{Eq4:newfewnomial}
630: X(l,n)\ <\ \frac{e^2+3}{4}2^{\binom{l}{2}}n^l
631: \end{equation}
632: %
633: for positive solutions and, when $\calA$ is primitive, a bound for all real solutions,
634: \[
635: \frac{e^{\Red{4}}+3}{4}2^{\binom{l}{2}}n^l\,.
636: \]
637: This is only slightly larger---the difference is in the exponents $2$ and $4$ of $e$.
638: These are proven in~\cite{BaBiSo,BS07}.
639:
640: By Corollary~\ref{C4:BRS} and the bound~\eqref{Eq4:newfewnomial},
641: \[
642: l^{-l}n^l\ <\ \left\lceil\frac{n}{l}\right\rceil^l\
643: <\ X(n,l)\ <\ \frac{e^2+3}{4}2^{\binom{l}{2}}n^l\,.
644: \]
645: This reveals the correct asymptotic information for $X(n,l)$, when $l$ is fixed,
646: $X(n,l)=\Theta(n^l)$.
647:
648:
649: Theorems~\ref{T4:BBS} and~\ref{T4:Bihan} are related, and we will outline
650: their proofs, following the papers in which they appear, where more details may be
651: found.
652: To begin, let
653: %
654: \begin{equation}\label{E4:system}
655: f_1(x_1,x_2,\dotsc,x_n)\ =\
656: f_2(x_1,x_2,\dotsc,x_n)\ =\ \dotsb\ =\
657: f_n(x_1,x_2,\dotsc,x_n)\ =\ 0
658: \end{equation}
659: %
660: be a system with support a circuit $\calA$.
661: Suppose that $0\in\calA$ and list the elements of the circuit
662: $\calA=\{0,a_0,a_1,\dotsc,a_n\}$.
663: After a multiplicative change of coordinates (if necessary), we may assume that
664: $a_0=\ell \e_n$, where $\e_n$ is the $n$th standard basis vector.
665: Since the system~\eqref{E4:system} is generic, row operations on the equations
666: put it into diagonal form
667: %
668: \begin{equation}\label{E4:Diagonal_circuit}
669: x^{a_i}\ =\ w_i\ +\ v_i x_n^\ell
670: \qquad\mbox{for }i=1,\dotsc,n\,.
671: \end{equation}
672: %
673:
674: When $\calA$ was a simplex we used integer linear algebra to reduce the
675: equations to a very simple system in Section~\ref{S3:simplex}.
676: We use (different) integer linear algebra to simplify this system
677: supported on a circuit.
678:
679: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
680: %
681: \subsection{Some arithmetic for circuits}
682: %
683: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
684: Suppose that $\{0, a_0, a_1,\dotsc,a_n\}\subset\Z^n$ is a primitive circuit.
685: We assume here that it is nondegenerate---there is no affine dependency
686: involving a subset.
687: (The bounds in the degenerate case are lower, replacing $n$ by the size of this
688: smaller circuit.)
689: After possibly making a coordinate change, we may assume that
690: $a_0=\ell\cdot \e_n$, where $\e_n$ is the $n$th standard basis vector.
691:
692: For each $i$, we may write $a_i=b_i+k_i\cdot \e_n$, where $b_i\in\Z^{n-1}$.
693: Removing common factors from a nontrivial integer linear relation among
694: the $n$ vectors $\{b_1,\dotsc,b_n\}\subset\Z^{n-1}$ gives us the
695: \DeCo{{\sl primitive relation}} among them (which is well-defined up to
696: multiplication by $-1$),
697: \[
698: \sum_{i=1}^p \lambda_i b_i\ =\
699: \sum_{i=p+1}^n \lambda_i b_i\,.
700: \]
701: Here, each $\lambda_i>0$, and we assume that the vectors are ordered
702: so that the relation has this form.
703: We further assume that
704: \[
705: N\ :=\ \sum_{i=p+1}^n \lambda_i k_i\ -\
706: \sum_{i=1}^p \lambda_i k_i\ >\ 0\,.
707: \]
708: Then we have
709: %
710: \[
711: N \e_n\ +\ \sum_{i=1}^p \lambda_i a_i\ -\
712: \sum_{i=p+1}^n \lambda_i a_i\ =\ 0\,,
713: \]
714: %
715: and so
716: %
717: \begin{equation}\label{E4:primitiverelation}
718: x_n^N \cdot \prod_{i=1}^p (x^{a_i})^{\lambda_i} \ -\
719: \prod_{i=p+1}^n (x^{a_i})^{\lambda_i}\ =\ 0\,.
720: \end{equation}
721: %
722:
723: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
724: %
725: \subsection{Elimination for circuits}
726: %
727: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
728:
729: Using~\eqref{E4:Diagonal_circuit} to
730: substitute for $x^{a_i}$ in~\eqref{E4:primitiverelation}
731: gives the univariate consequence
732: of~\eqref{E4:Diagonal_circuit}
733: %
734: \begin{equation}\label{Eq4:eliminant}
735: \DeCo{f}(x_n)\ :=\ x_n^N \prod_{i=1}^p (w_i+v_i x_n^\ell)^{\lambda_i}
736: \ -\ \prod_{i=p+1}^n (w_i+v_i x_n^\ell)^{\lambda_i}\ .
737: \end{equation}
738: %
739: Some further arithmetic of circuits (which may be found in~\cite{BBS}) shows
740: that $f(x_n)$ has degree equal to $n!\vol(\Delta_\calA)$.
741: This is in fact the eliminant of the system.
742:
743: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
744: \begin{lemma}\label{L4:eliminant}
745: The association of a solution $x$ of~\eqref{E4:Diagonal_circuit} to its $n$th
746: coordinate $x_n$ gives a bijection between the solutions
747: of~\eqref{E4:Diagonal_circuit} and the roots of $f~\eqref{Eq4:eliminant}$ which restricts to a
748: bijection between their real solutions/roots.
749: \end{lemma}
750: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
751:
752: While $f$ is the eliminant of the system, we do not have a Gr\"obner basis or
753: even a triangular system to witness this fact, and the proof proceeds by
754: explicitly constructing a solution~\eqref{E4:Diagonal_circuit} from a root
755: $x_n$ of $f$.
756:
757: The upper bound is found by writing $f=F-G$ as in~\eqref{Eq4:eliminant} and then
758: perturbing $f$,
759: \[
760: f_t(y)\ =\ t\cdot F(y)\ -\ G(y)\,.
761: \]
762: We simply estimate the number of changes in the the real roots of $f_t$ as
763: $t$ passes from $-\infty$ to $1$, which can occur only at the singular roots of
764: $f_t$.
765: While similar to the proof of Khovanskii's theorem, this is not inductive, but
766: relies on the form of the Wronskian $F'G-G'F$ whose roots are the singular
767: roots of $f_t$.
768: This may also be seen as an application of Rolle's Theorem.
769: We note that this estimation also uses Viro's construction
770: for $t$ near $0$ and $\infty$.
771:
772: These estimates prove the bounds in Theorems~\ref{T4:BBS} and~\ref{T4:Bihan}.
773: Sharpness comes from construction.
774: In~\cite{BBS} Viro's method for univariate polynomials is used to
775: construct polynomials $f$ are constructed with $2n{+}1$ real solutions, for special primitive
776: circuits.
777: Bihan~\cite{Bihan} constructs a system with $n{+}1$ positive solutions using
778: Grothendieck's {\it dessins d'enfants}.
779:
780: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
781: %
782: \subsection{A family of systems with a sharp bound}\label{S4:exact}
783: %
784: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
785:
786: We give a family of systems that illustrate the result of Theorem~\ref{T4:BBS}
787: (actually of an extension of it) and which may be treated by hand.
788: These systems come from a family of polytopes $\Delta\subset\Z^n$ for which we
789: prove a nontrivial upper bound on the number of real solutions to polynomial
790: systems with primitive support $\calA:=\Delta\cap\Z^n$.
791: That is, the integer points $\calA$ in $\Delta$ affinely span $\Z^n$,
792: so that general systems supported on $\Delta$ have $n!\vol(\Delta)$ complex solutions,
793: but there are fewer than $n!\vol(\Delta)$ real solutions to polynomial systems with
794: support $\calA$.
795: This is intended to not only give a glimpse of the more general results
796: in~\cite{BBS}, but also possible extensions which are not treated in~\cite{BS07}.
797:
798: Let $l > k>0$ and $n \ge 3$ be integers and
799: $\DeCo{\epsilon} = (\epsilon_1,\dotsc,\epsilon_{n-1})\in\{0,1\}^{n-1}$ have at least one
800: nonzero coordinate.
801: The polytope $\DeCo{\Delta_{k,l}^\epsilon} \subset \R^n$ is the convex hull of the points
802: %
803: \[
804: (0,\dots,0),\ (1,0,\dots,0),\dots, (0,\dots,0,1,0),\ (0,\dots,0,k),\
805: (\epsilon_1,\dots, \epsilon_{n-1},l)\,.
806: \]
807: %
808: The configuration $\DeCo{\calA_{k,l}^\epsilon}=\Delta_{k,l}^\epsilon\cap\Z^n$
809: also includes the points along the last axis
810: \[
811: (0,\dotsc,0,1),\ (0,\dotsc,0,2),\ \dotsc,\ (0,\dotsc,0,k{-}1)\,.
812: \]
813: These points include the standard basis and the origin, so $\calA_{k,l}^\epsilon$
814: is primitive in that $\Z\calA=\Z^n$.
815:
816: Set $|\epsilon|:=\sum_i\epsilon_i$. Then the volume of
817: $\Delta_{k,l}^\epsilon$ is $(l+k|\epsilon|)/n!$. Indeed,
818: the configuration $\calA_{k,l}^\epsilon$ can be triangulated into
819: two simplices $\Delta_{k,l}^\epsilon \setminus \{(\epsilon_1,\dots, \epsilon_{n-1},l)\}$
820: and $\Delta_{k,l}^\epsilon \setminus \{0\}$ with volumes $k/n!$ and $(l-k+k|\epsilon|)/n!$,
821: respectively. One way to see this is to apply the affine transformation
822: \[
823: (x_1,\dotsc,x_n)\ \longmapsto\
824: (x_1,\dotsc,x_{n-1},x_n-k+k\sum_{i=1}^{n-1}x_i)\,.
825: \]
826:
827: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
828: \begin{thm}\label{T4:aBound}
829: The number, $r$, of real solutions to a generic system of $n$ real
830: polynomials with support $\calA_{k,l}^\epsilon$ lies in the interval
831: \[
832: 0 \ \leq\ r\ \leq \ k + k|\epsilon|+2\,,
833: \]
834: and every number in this interval with the same parity as $l+k|\epsilon|$
835: occurs.
836: \end{thm}
837: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
838:
839: This upper bound does not depend on $l$ and, since
840: $k < l$, it is smaller than or equal to the number $l+k|\epsilon|$
841: of complex solutions.
842: We use elimination to prove this result.
843:
844:
845: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
846: \begin{ex}\label{E:km-polytope}
847: Suppose that $n=k=3$, $l=5$, and $\epsilon = (1,1)$.
848: \[
849: \begin{array}{rclcl}
850: \makebox[210pt][l]{Then the system}&&&&
851: \multirow{9}*{
852: \begin{picture}(80,105)(0,15)
853: \put( 4,0){\includegraphics[height=4.5cm]{figures/4/35-polytope.eps}}
854: \put(-1,24){$1$} \put(16, 3){$x$} \put(39,20){$y$}
855: \put(-1,42){$z$} \put(-1,63){$z^2$} \put(-1,85){$z^3$}
856: \put(50,120){$xyz^5$}
857: \put(55,60){$\Delta_{3,5}^{(1,1)}$}
858: \end{picture}}
859: \\
860: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
861: % A:=linalg[matrix]([[1,1,1],[1,2,3],[2,2,1]]);
862: % B:=linalg[matrix]([[1,1,1,1],[5,7,11,13],[4,8,16,32]]);
863: % AI:=linalg[inverse](A);
864: % linalg[multiply](AI,linalg[concat](A,B));
865: % [1 0 0 -5 -11 -23 -41]
866: % [0 1 0 8 18 38 72]
867: % [0 0 1 -2 -6 -14 -30]
868: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
869: x + y + xyz^5 +1+z+z^2+z^3&=&0&\quad&\rule{0pt}{16pt}\\
870: x + 2y + 3xyz^5 +5+7z+11z^2+13z^3&=&0&\quad&\\
871: 2x + 2y + xyz^5 +4+8z+16z^2+32z^3&=&0&\quad&\\
872: \makebox[210pt][l]{is equivalent to }&&\rule{0pt}{16pt}\\
873: x -(5 +11z +23z^2 +41z^3)&=&0\rule{0pt}{16pt}\\
874: y +( 8 +18z +38z^2 +72z^3)&=&0\\
875: xyz^5 -(2 +6z +14 z^2 +30 z^3)&=&0\\
876: \mbox{\ }
877: \end{array}
878: \]
879: And thus its number of real roots equals the number of real roots of
880: \[
881: z^5(5 +11z +23z^2 +41z^3)(8 +18z +38z^2 +72z^3) -
882: (2 +6z +14 z^2 +30 z^3)\,,
883: \]
884: %
885: % A:=(5 +11*z +23*z2 +41*z3):
886: % B:=(8 +18*z +38*z2 +72*z3)*z5:
887: % C:=(2 +6*z +14*z2 +30*z3):
888: % realroot(A*B-C);
889: %
890: which, as we invite the reader to check, is 3. \QED
891: \end{ex}
892: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
893:
894:
895: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
896: \noindent{\it Proof of Theorem~$\ref{T4:aBound}$.}
897: A generic real polynomial system with support $\calA_{k,l}^\epsilon$ has the
898: form
899: %
900: \[
901: \sum_{j=1}^{n-1} c_{ij}x_j \ + c_{in}x^\epsilon x_n^l\ \
902: +\ f_i(x_n)\ =\ 0\ \quad
903: {\rm for\ }i=1,\dots,n\,,
904: \]
905: where each polynomial $f_i$ has degree $k$ and
906: $x^\epsilon$ is the monomial
907: $x_1^{\epsilon_1}\dotsb x_{n-1}^{\epsilon_{n-1}}$.
908:
909:
910: Since all solutions to our system are simple, we may perturb the
911: coefficient matrix $(c_{ij})_{i,j=1}^n$ if necessary and then use
912: Gaussian elimination to obtain an equivalent system
913: %
914: \begin{equation}\label{system}
915: x_1-g_1(x_n)\ =\ \dotsb \ = \ x_{n-1}-g_{n-1}(x_n) \ =\
916: x^\epsilon x_n^l -g_{n}(x_n)\ =\ 0\,,
917: \end{equation}
918: %
919: where each polynomial $g_i$ has degree $k$.
920: Using the first $n-1$ polynomials to eliminate the
921: variables $x_1, \dots , x_{n-1}$ gives the univariate polynomial
922: %
923: \begin{equation}\label{E:eliminantex}
924: x_n^l \cdot g_1(x_n)^{\epsilon_1} \dotsb
925: g_{n-1}(x_n)^{\epsilon_{n-1}}\ -\ g_n(x_n)\,,
926: \end{equation}
927: %
928: which has degree $l+k|\epsilon|=v(\Delta_{k,l}^\epsilon)$. Any zero
929: of this polynomial leads to a solution of the original
930: system~\eqref{system} by back substitution. This implies that the
931: number of real roots of the polynomial~\eqref{E:eliminantex} is equal
932: to the number of real solutions to our original
933: system~\eqref{system}.
934:
935: The eliminant~\eqref{E:eliminantex} has no terms of degree $m$ for $k<m<l$, and
936: so it has at most $k+k|\epsilon|+2$
937: nonzero real roots, by Descartes's rule of signs.
938: This proves the upper bound.
939:
940: We omit the construction which shows that this bound is sharp.
941: \QED
942: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
943: