betaFeedback

WorksMathematics

A reciprocal Schoenberg inequality and its extensions

Abstract

We study reciprocal moments of the critical points of a polynomial with a simple zero at the origin. A positive similarity transforms a reciprocal derivative companion into a complex symmetric matrix. This gives a short proof of the reciprocal Schoenberg inequality and a refinement for every elementary symmetric function. We determine the optimal constants in the corresponding inequalities for every positive exponent, including the sublinear range, and extend the results to arbitrary positive charges. We also characterize the normal powers of the companion and of its symmetric representative. The resulting criterion gives the complete equality cases in two quartic inequalities and explains the appearance of noncollinear equality configurations at higher orders. We also obtain a closed generating determinant and a quantitative collinearity estimate.

Introduction

The Gauss–Lucas theorem places the critical points of a polynomial in the convex hull of its zeros. Moment inequalities give a different kind of information: they compare the distribution of the critical points with that of the zeros. A distinguished example is Schoenberg’s inequality. If a polynomial of degree N≥2N \ge2 has zeros ζ1,…,ζN\zeta_1,\ldots,\zeta_N and critical points ω1,…,ωN−1\omega_1,\ldots,\omega_{N-1}, then

∑j=1N−1∣ωj∣2≤N−2N∑j=1N∣ζj∣2+1N2∣∑j=1Nζj∣2.(1)\sum_{j=1}^{N-1}|\omega_j|^2 \le\frac{N-2}{N}\sum_{j=1}^{N}|\zeta_j|^2+\frac{1}{N^2}\left|\sum_{j=1}^{N}\zeta_j\right|^2 . \tag*{(1)}

Equality holds precisely when the zeros are collinear. Schoenberg proposed the centered form of this inequality in 1986 [9]. Pereira [8] and Malamud [7] proved it through matrix representations of differentiation. These representations also connect polynomial geometry with the majorization relations of de Bruijn and Springer.

Cheung and Ng [4] developed derivative companion matrices and used them to prove higher-order inequalities, including a conjecture of de Bruin and Sharma. Kushel and Tyaglov [6] obtained a sharper quartic inequality through circulant matrices. Borcea [3] extended the majorization framework to equilibrium points of arbitrary positive charge distributions. Exterior powers already play an essential role in this literature: they replace individual eigenvalues by products of eigenvalues.

Negative moments arise naturally when one measures distances from a distinguished zero. We therefore write

P(z)=z∏j=1n(z−zj),zj≠0,(2)P(z)=z\prod_{j=1}^{n}(z-z_j),\qquad z_j\ne0, \tag*{(2)}

and denote the zeros of P′P' by w1,…,wnw_1,\ldots,w_n. Since P′(0)≠0P'(0)\ne0, all the wjw_j are nonzero. Tang and Zhang [10 Conjecture 6.3] proposed the reciprocal inequality

∑j=1n∣wj∣−2≤3∑j=1n∣zj∣−2+∣∑j=1nzj−1∣2.(3)\sum_{j=1}^{n}|w_j|^{-2}\le3\sum_{j=1}^{n}|z_j|^{-2}+\left|\sum_{j=1}^{n}z_j^{-1}\right|^2 . \tag*{(3)}

Their Conjecture 1.13 asks for the optimal power-sum bound

∑j=1n∣wj∣−p≤(n+1)p+n−1n∑j=1n∣zj∣−p,p≥1.(4)\sum_{j=1}^{n} |w_j|^{-p} \le\frac{(n+1)^p+n-1}{n}\sum_{j=1}^{n} |z_j|^{-p}, \qquad p \ge1. \tag*{(4)}

We prove both inequalities, give their equality cases, and place them in a larger family. At p=1p=1, the equality assertion in Conjecture 1.13 requires a correction: the nonzero zeros need only lie on one ray, rather than coincide.

The basic observation is simple. Put D=diag⁡(z1−1,…,zn−1)D=\operatorname{diag}(z_1^{-1},\ldots,z_n^{-1}) and let JJ be the all-ones matrix. The matrix Q=D(I+J)Q=D(I+J) has eigenvalues wj−1w_j^{-1}. Its similar representative

C=(I+J)1/2D(I+J)1/2C=(I+J)^{1/2}D(I+J)^{1/2}

is complex symmetric and has a smaller Frobenius norm. Schur’s inequality applied to CC proves (3); applying it to exterior powers of CC retains the same collinearity equality condition. A second factorization removes the phases of the zj−1z_j^{-1} and leads, through the Araki–Lieb–Thirring inequality, to sharp constants for all positive exponents.

For unit charges, the phase comparison itself is contained in Cheung–Ng’s Theorem 2.1 after inversion and the use of the product identity. Likewise, the endpoint p=1p=1 of our elementary symmetric bounds follows from Borcea’s product majorization. Our bounds for p>1p>1 improve the coefficients supplied by that majorization, as explained in Remark 4.4. The additional phase-sensitive refinements and the equality analysis use the same companion, but require its symmetric representative. In particular, a divided-difference formula characterizes all its normal powers. This gives the complete equality classification left open in [10 Remark 7.2] and shows why the refined quartic inequality has only collinear equality cases, whereas higher even orders need not.

Definitions and notation

Throughout, z1,…,znz_1,\ldots,z_n are nonzero complex numbers and r1,…,rnr_1,\ldots,r_n are positive real numbers. Repetitions among the zjz_j are allowed. Define

Fr(z)=∏j=1n(z−zj)+z∑j=1nrj∏ℓ≠j(z−zℓ).(5)F_r(z)=\prod_{j=1}^{n}(z-z_j)+z\sum_{j=1}^{n}r_j\prod_{\ell\ne j}(z-z_\ell). \tag*{(5)}

Let w1,…,wnw_1,\ldots,w_n be its zeros, counted with multiplicity, and put

bj=zj−1,λj=wj−1,σ=∑jrj,u=(r1,…,rn)T.b_j=z_j^{-1}, \qquad\lambda_j=w_j^{-1}, \qquad\sigma=\sum_j r_j, \qquad u=(\sqrt{r_1},\ldots,\sqrt{r_n})^T.

The polynomial FrF_r has degree nn, leading coefficient 1+σ1+\sigma, and a nonzero constant term. Thus all these reciprocals are defined.

For distinct positions, the wjw_j are the zeros of

1z+∑jrjz−zj.\frac{1}{z}+\sum_j\frac{r_j}{z-z_j}.

the force function associated with positive charges 1,r1,…,rn1,r_1,\ldots,r_n. For repeated positions we use the polynomial definition (2.1), so that roots lost through rational cancellation are retained. Arbitrary positive charges α0,…,αn\alpha_0,\ldots,\alpha_n are obtained by taking rj=αj/α0r_j=\alpha_j/\alpha_0. The choice rj=1r_j=1 gives Fr=P′F_r=P' for (1.2). More generally, rj=1/ℓr_j=1/\ell gives the nn nonzero critical points of zℓ∏j(z−zj)z^\ell\prod_j(z-z_j), where ℓ\ell is a positive integer.

Our matrices are

D=diag⁡(b1,…,bn),K=I+uuT,H=K1/2,Q=DK,C=HDH.(6)D=\operatorname{diag}(b_1,\ldots,b_n), \qquad K=I+uu^T, \qquad H=K^{1/2}, \qquad Q=DK, \qquad C=HDH. \tag*{(6)}

Here HH is the positive square root. The matrices KK, HH are real symmetric, CT=CC^{T}=C, and C=HQH−1C=HQH^{-1}. Lemma 4.1 identifies their spectrum with (λj)(\lambda_j).

For a vector xx, let ek(x)e_k(x) denote its kkth elementary symmetric function. We abbreviate

Ek,p(w)=ek(∣w1∣−p,…,∣wn∣−p),bI=∏j∈Ibj,rI=∑j∈Irj\mathcal{E}_{k,p}(w)=e_k\left(\lvert w_1\rvert^{-p},\ldots,\lvert w_n\rvert^{-p}\right),\qquad b_I=\prod_{j\in I}b_j,\qquad r_I=\sum_{j\in I}r_j

with b∅=1b_{\varnothing}=1. A sum over ∣I∣=k\lvert I\rvert=k runs over all kk-element subsets of {1,…,n}\{1,\ldots,n\}. The Frobenius and operator norms are denoted by ∥⋅∥F\lVert\cdot\rVert_F and ∥⋅∥op\lVert\cdot\rVert_{\mathrm{op}}. We use X∗X^* for the conjugate transpose, X‾\overline{X} for entrywise conjugation, and [X,Y]=XY−YX[X,Y]=XY-YX. Collinearity below always means collinearity on a line through the origin. It is equivalent for the configurations (zj)(z_j) and (bj)(b_j).

Main results

Sharp inequalities for all positive exponents

Theorem 3.1. For 1≤k≤n1\leq k\leq n and p≥1p\geq1,

Ek,p(w)≤∑∣I∣=k(1+(1+σ)p−1σrI)∣bI∣p.(7)\mathcal{E}_{k,p}(w)\leq\sum_{\lvert I\rvert=k}\left(1+\frac{(1+\sigma)^p-1}{\sigma}r_I\right)\lvert b_I\rvert^p. \tag*{(7)}

For 0<p<10<p<1,

Ek,p(w)≤∑∣I∣=k(1+rI)p∣bI∣p.(8)\mathcal{E}_{k,p}(w)\leq\sum_{\lvert I\rvert=k}(1+r_I)^p\lvert b_I\rvert^p. \tag*{(8)}

If k<nk<n, equality in (3.1) holds exactly when all bjb_j coincide for p>1p>1, and exactly when they have the same argument for p=1p=1. Inequality (3.2) is strict for k<nk<n, but each of its coefficients is individually optimal. For k=nk=n, both inequalities are identities for every p>0p>0.

Corollary 3.2. Suppose r1=⋯=rn=ρ>0r_1=\cdots=r_n=\rho>0. The best constant in

Ek,p(w)≤An,k,p(ρ)ek(∣z1∣−p,…,∣zn∣−p)\mathcal{E}_{k,p}(w)\leq A_{n,k,p}(\rho)e_k\left(\lvert z_1\rvert^{-p},\ldots,\lvert z_n\rvert^{-p}\right)

is

An,k,p(ρ)={(1+kρ)p,0<p≤1,1+kn((1+nρ)p−1),p≥1.(9)A_{n,k,p}(\rho)= \begin{cases} (1+k\rho)^p, & 0<p\leq1,\\ 1+\dfrac{k}{n}\left((1+n\rho)^p-1\right), & p\geq1. \end{cases} \tag*{(9)}

For k<nk<n, the equality cases are those in Theorem 3.1; in the sublinear range the optimal constant is not attained. For k=nk=n, equality always holds.

Taking k=1k=1 and ρ=1\rho=1 proves (1.4). For example, P(z)=z(z−1)(z−2)P(z)=z(z-1)(z-2) attains equality at p=1p=1, although its nonzero zeros are distinct. At the other end, k=nk=n records the product identity

∏jλj=(1+σ)∏jbj.(10)\prod_j\lambda_j=(1+\sigma)\prod_j b_j. \tag*{(10)}

A refinement retaining the phases

Theorem 3.3. For 1≤k≤n1\leq k\leq n, define

Bk(b;r)=∑∣I∣=k((1+rI)2−∑j∈Irj2)∣bI∣2+∑∣A∣=k−1∣bA∣2∣∑j∉Arjbj∣2.(11)B_k(b;r)=\sum_{\lvert I\rvert=k}\left((1+r_I)^2-\sum_{j\in I}r_j^2\right)\lvert b_I\rvert^2 +\sum_{\lvert A\rvert=k-1}\lvert b_A\rvert^2\left\lvert\sum_{j\notin A}r_jb_j\right\rvert^2. \tag*{(11)}

Then

Ek,2(w)≤Bk(b;r).(12)\mathcal{E}_{k,2}(w) \le B_k(b;r). \tag*{(12)}

For k<nk<n, equality holds if and only if the zjz_j are collinear. For k=nk=n, equality always holds. Moreover,

∑∣I∣=k(1+(σ+2)rI)∣bI∣2−Bk(b;r)=∑∣A∣=k−1∣bA∣2∑i<ji,j∉Arirj∣bi−bj∣2.(13)\sum_{|I|=k}\left(1+(\sigma+2)r_I\right)|b_I|^2-B_k(b;r) = \sum_{|A|=k-1}|b_A|^2\sum_{\substack{i<j\\i,j\notin A}}r_i r_j|b_i-b_j|^2. \tag*{(13)}

For unit charges this reads

Ek,2(w)≤(k2+k+1)ek(∣b∣2)+∑∣A∣=k−1∣bA∣2∣∑j∉Abj∣2.(14)\mathcal{E}_{k,2}(w)\le(k^2+k+1)e_k(|b|^2)+\sum_{|A|=k-1}|b_A|^2\left|\sum_{j\notin A}b_j\right|^2. \tag*{(14)}

At k=1k=1 it is precisely (3), including its complete equality characterization. Identity (13) measures the improvement over Theorem 3.1 at p=2p=2 by a sum of squared pairwise distances.

Normal powers and even moments

Define monic polynomials fmf_m of degree mm by

f1(t)=t,am=∑jrjfm(bj),fm+1(t)=t(fm(t)+am).(15)f_1(t)=t,\qquad a_m=\sum_j r_jf_m(b_j),\qquad f_{m+1}(t)=t\left(f_m(t)+a_m\right). \tag*{(15)}

Thus f2(t)=t2+stf_2(t)=t^2+st, where s=∑jrjbjs=\sum_jr_jb_j.

Theorem 3.4. For every positive integer mm,

∑j∣wj∣−2m≤∥Cm∥F2≤∥Qm∥F2,(16)\sum_j|w_j|^{-2m}\le\lVert C^m\rVert_F^2\le\lVert Q^m\rVert_F^2, \tag*{(16)}

and the second difference is the weighted variance

∥Qm∥F2−∥Cm∥F2=σ∑jrj∣fm(bj)∣2−∣am∣2.(17)\lVert Q^m\rVert_F^2-\lVert C^m\rVert_F^2 =\sigma\sum_jr_j|f_m(b_j)|^2-|a_m|^2. \tag*{(17)}

The following equality criteria hold:

(1) QmQ^m is normal if and only if fm(b1)=⋯=fm(bn)f_m(b_1)=\cdots=f_m(b_n).

(2) CmC^m is normal if and only if the bjb_j are collinear or fm(b1)=⋯=fm(bn)f_m(b_1)=\cdots=f_m(b_n).

Thus equality in the first inequality of (16) has exactly the configurations in (2), while equality between its first and last terms has exactly those in (1). Every noncollinear equality configuration has at most mm distinct nonzero zeros.

For the quartic case, introduce

S=∑jrj∣bj∣2,s=∑jrjbj,T=∑jrjbj2,S=\sum_jr_j|b_j|^2,\qquad s=\sum_jr_jb_j,\qquad T=\sum_jr_jb_j^2,
V0=∑j∣bj∣4,V1=∑jrj∣bj∣4,W=∑jrj∣bj2+sbj∣2.(18)V_0=\sum_j|b_j|^4,\qquad V_1=\sum_jr_j|b_j|^4,\qquad W=\sum_jr_j|b_j^2+sb_j|^2. \tag*{(18)}

Corollary 3.5. One has

∑j∣wj∣−4≤V0+2V1+S2+2W+∣T+s2∣2\sum_j|w_j|^{-4}\le V_0+2V_1+S^2+2W+|T+s^2|^2
≤V0+2V1+S2+(σ+2)W.(19)\le V_0+2V_1+S^2+(\sigma+2)W. \tag*{(19)}

Equality in the first inequality holds if and only if the zjz_j are collinear. Equality between the first and last expressions holds if and only if either all bjb_j coincide, or they take exactly two distinct values a,ca,c satisfying

(1+Ra)a+(1+Rc)c=0,Ra=∑bj=arj,Rc=∑bj=crj.(20)(1+R_a)a+(1+R_c)c=0,\qquad R_a=\sum_{b_j=a}r_j,\quad R_c=\sum_{b_j=c}r_j. \tag*{(20)}

For rj=1r_j=1, the last expression in (19) is the bound in [10 Theorem 7.1]. Its equality cases, apart from coincident roots, are therefore

P(z)=z(z−ζ)h(z−η)n−h,(n−h+1)ζ+(h+1)η=0,P(z)=z(z-\zeta)^h(z-\eta)^{n-h},\qquad(n-h+1)\zeta+(h+1)\eta=0,

where 1≤h<n1\le h<n and ζ,η\zeta,\eta are nonzero. The opposite pair for n=2n=2 was already noted in [10 Remark 7.2]; the displayed relation gives the complete classification.

The symmetric companion and the sharp inequalities

Representation and the quadratic inequality

Lemma 4.1. The matrices QQ and CC in (6) have eigenvalues λ1,…,λn\lambda_1,\ldots,\lambda_n. The matrix CC is normal if and only if the bjb_j are collinear, and

∥C∥F2=∑j(1+2rj)∣bj∣2+∣∑jrjbj∣2.(21)\lVert C\rVert_F^2=\sum_j(1+2r_j)|b_j|^2+\left|\sum_j r_jb_j\right|^2. \tag*{(21)}

Proof. The rank-one determinant formula gives the polynomial identity

det⁡(λI−DK)=∏j(λ−bj)−∑jrjbj∏ℓ≠j(λ−bℓ)=λnFr(1/λ)Fr(0).\begin{aligned} \det(\lambda I-DK)=\prod_j(\lambda-b_j)-\sum_jr_jb_j\prod_{\ell\ne j}(\lambda-b_\ell) \\ &=\frac{\lambda^nF_r(1/\lambda)}{F_r(0)}. \end{aligned}

The expression on the right is monic and has precisely the required roots. This also proves (10). Since C=HQH−1C=HQH^{-1}, it has the same spectrum.

Set v=Duv=Du. Direct multiplication gives

[C,C∗]=H(vv∗−vˉvT)H.(22)[C,C^*]=H(vv^*-\bar{v}v^T)H. \tag*{(22)}

The (i,j)(i,j) entry of the middle matrix is rirj(bibˉj−bˉibj)\sqrt{r_ir_j}(b_i\bar{b}_j-\bar{b}_ib_j). It vanishes for every pair exactly when the bjb_j lie on one line through the origin. Finally, ∥C∥F2=tr⁡(D∗KDK)\lVert C\rVert_F^2=\operatorname{tr}(D^*KDK); expanding K=I+uuTK=I+uu^T gives (21).

Schur triangularization yields, for any square matrix XX,

∑j∣λj(X)∣2≤∥X∥F2,(23)\sum_j|\lambda_j(X)|^2\le\lVert X\rVert_F^2, \tag*{(23)}

with equality exactly when XX is normal. Indeed, the difference is the sum of the squared moduli of the strictly upper triangular entries of a Schur form. Lemma 4.1 and (23) prove the weighted reciprocal Schoenberg inequality, and hence (3), with equality.

Exterior powers

Write X[k]=⋀kXX^{[k]}=\bigwedge^k X. In the standard basis indexed by kk-element subsets, its (I,J)(I,J) entry is det⁡X[I,J]\det X[I,J]. The identities

(XY)[k]=X[k]Y[k],(X∗)[k]=(X[k])∗(XY)^{[k]}=X^{[k]}Y^{[k]},\qquad(X^*)^{[k]}=(X^{[k]})^*

follow from Cauchy–Binet. The eigenvalues of X[k]X^{[k]} are the kk-fold products of the eigenvalues of XX; its singular values are the corresponding products of singular values. For positive definite XX, functional calculus gives (Xa)[k]=(X[k])a(X^a)^{[k]}=(X^{[k]})^a for real aa.

We will use the following elementary equality observation.

Lemma 4.2. If XX is invertible and 1≤k<n1\leq k<n, then X[k]X^{[k]} is normal if and only if XX is normal.

Proof. It suffices to show that P↦P[k]P\mapsto P^{[k]} is injective on positive definite matrices. If P[k]=R[k]P^{[k]}=R^{[k]}, then (R−1/2PR−1/2)[k]=I(R^{-1/2}PR^{-1/2})^{[k]}=I. Every kk-fold product of the positive eigenvalues of R−1/2PR−1/2R^{-1/2}PR^{-1/2} is therefore 11. Comparing products that differ in one index shows that all the eigenvalues are equal; hence each is 11 and P=RP=R. Apply this to P=XX∗P=XX^* and R=X∗XR=X^*X. □

Proof of Theorem 3.3. Apply (4.3) to C[k]C^{[k]}. Its squared Frobenius norm is

∥C[k]∥F2=tr⁡((D[k])∗K[k]D[k]K[k])=∑∣I∣=∣J∣=kbI‾bJdet⁡K[I,J]2.(24)\left\|C^{[k]}\right\|_F^2=\operatorname{tr}\left((D^{[k]})^*K^{[k]}D^{[k]}K^{[k]}\right)=\sum_{\lvert I\rvert=\lvert J\rvert=k}\overline{b_I}b_J\det K[I,J]^2. \tag*{(24)}

The minors of K=I+uuTK=I+uu^T have a particularly simple form:

det⁡K[I,I]=1+rI,det⁡K[A∪{i},A∪{j}]2=rirj(i≠j).\det K[I,I]=1+r_I,\qquad\det K[A\cup\{i\},A\cup\{j\}]^2=r_ir_j\quad(i\ne j).

Every other minor is zero, since two or more rows outside the common index set are proportional. Grouping the off-diagonal terms in (4.4) by A=I∩JA=I\cap J gives exactly (3.5). Equality follows from Lemmas 4.1 and 4.2; the top exterior power is scalar.

For the refinement identity, use the weighted variance formula

(∑j∈Jrj)(∑j∈Jrj∣bj∣2)−∣∑j∈Jrjbj∣2=∑i<ji,j∈Jrirj∣bi−bj∣2(25)\left(\sum_{j\in J}r_j\right)\left(\sum_{j\in J}r_j\lvert b_j\rvert^2\right)-\left\lvert\sum_{j\in J}r_jb_j\right\rvert^2=\sum_{\substack{i<j\\i,j\in J}}r_ir_j\lvert b_i-b_j\rvert^2 \tag*{(25)}

on each complement J=AcJ=A^c. After multiplication by ∣bA∣2\lvert b_A\rvert^2, the coefficient contributed to ∣bI∣2\lvert b_I\rvert^2 is ∑i∈Iri(σ−rI+ri)\sum_{i\in I}r_i(\sigma-r_I+r_i). Combining this with the first sum in (3.5) gives 1+(σ+2)rI1+(\sigma+2)r_I, which proves (3.7). □

Removing the phases

We recall two standard matrix inequalities in the precise forms needed here. If XX is invertible and sj(X)s_j(X) are its singular values, Weyl’s multiplicative inequality gives

∑j∣λj(X)∣p≤∑jsj(X)p,p>0.(26)\sum_j\lvert\lambda_j(X)\rvert^p\leq\sum_j s_j(X)^p,\qquad p>0. \tag*{(26)}

Equality holds if and only if XX is normal. To see the equality assertion, apply strict convexity of t↦eptt\mapsto e^{pt} to the majorization of logarithms of eigenvalue moduli by logarithms of singular values. Equality identifies the two multisets; (4.3) then implies normality. See [2] for multiplicative majorization.

For positive definite X,YX,Y, the Araki–Lieb–Thirring inequality is

tr⁡((X1/2YX1/2)p)≤tr⁡(XpYp),p≥1.(27)\operatorname{tr}\left((X^{1/2}YX^{1/2})^p\right)\leq\operatorname{tr}(X^pY^p),\qquad p\geq1. \tag*{(27)}

For p>1p>1, equality holds if and only if XY=YXXY=YX. We use Araki’s stronger log-majorization theorem [1]; a short justification of the equality condition is useful. For fixed p>1p>1, that theorem makes

g(t)=tr⁡((Xt/2YXt/2)p/t)g(t)=\operatorname{tr}\left((X^{t/2}YX^{t/2})^{p/t}\right)

nondecreasing on t>0t > 0. The function is real analytic there. Equality in (4.7) says g(1)=g(p)g(1) = g(p), so gg is constant first on [1,p][1,p] and then on (0,∞)(0,\infty). In particular,

0=g(p)−g(p/2)=12∥[Xp/2,Yp/2]∥F2.0 = g(p) - g(p/2) = \frac{1}{2}\lVert[X^{p/2},Y^{p/2}] \rVert_{F}^{2}.

Functional calculus now gives XY=YXXY = YX. This argument uses only the log-majorization theorem; see also Hiai [5] for a careful treatment of equality in related norm inequalities.

Lemma 4.3. Put R=diag⁡(∣b1∣,…,∣bn∣)R = \operatorname{diag}(\lvert b_1\rvert,\ldots,\lvert b_n\rvert) and A=R1/2KR1/2A = R^{1/2} K R^{1/2}. For all kk and p>0p > 0,

Ek,p(w)≤tr⁡((A[k])p).(28)\mathcal{E}_{k,p}(w) \le\operatorname{tr}\left((A^{[k]})^p\right). \tag*{(28)}

For k<nk < n, equality holds if and only if all bjb_j have the same argument.

Proof. Write D=URD = UR, with UU diagonal unitary, and choose a diagonal unitary VV with V2=UV^2 = U. The matrix M=D1/2KD1/2=VAVM = D^{1/2} K D^{1/2} = VAV is similar to QQ. Its singular values are the eigenvalues of AA. Apply (4.6) to M[k]M^{[k]}.

For k<nk < n, equality is equivalent to normality of MM, by Lemma 4.2. Since MM∗=VA2V∗MM^* = VA^2V^* and M∗M=V∗A2VM^*M = V^*A^2V, this is equivalent to UA2=A2UUA^2 = A^2U, and hence to UA=AUUA = AU. Every off-diagonal entry of AA is positive. Therefore UU is scalar, as required. □

Proof of Theorem 3.1 and Corollary 3.2. For p≥1p \ge1, apply (4.7) to R[k]R^{[k]} and K[k]K^{[k]} in (4.8). Since

Kp=I+(1+σ)p−1σuuT,K^p = I + \frac{(1+\sigma)^p-1}{\sigma}uu^T,

we obtain

Ek,p(w)≤tr⁡((Rp)[k](Kp)[k])=∑∣I∣=k∣bI∣p(1+(1+σ)p−1σrI).\mathcal{E}_{k,p}(w) \le\operatorname{tr}\left((R^p)^{[k]}(K^p)^{[k]}\right) = \sum_{\lvert I\rvert=k} \lvert b_I\rvert^p\left(1+\frac{(1+\sigma)^p-1}{\sigma}r_I\right).

For k<nk < n and p>1p > 1, equality also requires R[k]K[k]=K[k]R[k]R^{[k]}K^{[k]} = K^{[k]}R^{[k]}. Neighboring index sets A∪{i}A\cup\{i\} and A∪{j}A\cup\{j\} have a nonzero connecting entry in K[k]K^{[k]}, so ∣bi∣=∣bj∣\lvert b_i\rvert= \lvert b_j\rvert for every pair. Lemma 4.3 supplies the common argument. At p=1p = 1, only the latter condition remains.

For 0<p<10 < p < 1, the diagonal of a positive matrix is majorized by its eigenvalues. Concavity therefore gives

tr⁡((A[k])p)≤∑∣I∣=k(det⁡A[I,I])p=∑∣I∣=k(1+rI)p∣bI∣p.\operatorname{tr}\left((A^{[k]})^p\right) \le\sum_{\lvert I\rvert=k}\left(\det A[I,I]\right)^p = \sum_{\lvert I\rvert=k}(1+r_I)^p\lvert b_I\rvert^p.

For completeness, diagonalize A[k]A^{[k]} unitarily and apply scalar Jensen to each diagonal entry. Strict concavity shows that equality forces A[k]A^{[k]} to be diagonal. When k<nk < n, its entries joining neighboring index sets are nonzero, so the inequality is strict.

To prove that the coefficient of ∣bI∣p\lvert b_I\rvert^p cannot be reduced, take bj=1b_j = 1 for j∈Ij\in I and bj=ε>0b_j = \varepsilon> 0 otherwise, and let ε↓0\varepsilon\downarrow0. The nonzero eigenvalues in the limit are 1+rI1+r_I and k−1k-1 copies of 11. Thus the left-hand side tends to (1+rI)p(1+r_I)^p, while every other monomial on the right tends to zero.

For equal charges, the coefficients depend only on kk, giving (3.3). The preceding limiting configuration proves sharpness for p<1p < 1. For p≥1p \ge1, take all bjb_j equal: the eigenvalues are (1+nρ)b1(1+n\rho)b_1 and n−1n-1 copies of b1b_1, and

(n−1k)+(n−1k−1)(1+nρ)p(nk)=1+kn((1+nρ)p−1).\frac{\binom{n-1}{k}+\binom{n-1}{k-1}(1+n\rho)^p}{\binom{n}{k}} = 1+\frac{k}{n}\left((1+n\rho)^p-1\right).

The case k=nk=n is (3.4). □

Remark 4.4 (Comparison with product majorization). Borcea’s Theorem 3 [3], applied to products of n−kn-k equilibrium points and the convex function ∣⋅∣p\lvert\cdot\rvert^p, yields, after inversion using (3.4),

Ek,p(w)≤(1+σ)p−1∑∣I∣=k(1+rI)∣bI∣p,p≥1.(29)\mathcal{E}_{k,p}(w) \le(1+\sigma)^{p-1}\sum_{\lvert I\rvert=k}(1+r_I)\lvert b_I\rvert^p,\qquad p \ge1. \tag*{(29)}

Indeed, normalize the charges to total mass one. The coefficient of a product indexed by J⊆{1,…,n}J \subseteq\{1,\ldots,n\} in the forward inequality is 1−∑j∈Jrj/(1+σ)1-\sum_{j\in J}r_j/(1+\sigma); products containing the distinguished zero vanish. Set J=IcJ=I^c and divide by the total product of the equilibrium points. This gives (29).

For k=nk=n the product identity suffices.

At p=1p=1, (29) and (7) coincide. For p>1p>1 and k<nk<n, the coefficient in (29) exceeds ours by

((1+σ)p−1−1)(1−rIσ)>0.\left((1+\sigma)^{p-1}-1\right)\left(1-\frac{r_I}{\sigma}\right)>0.

Thus the endpoint is already a consequence of product majorization, while the new coefficients give a strict improvement away from that endpoint.

Normal powers and equality in higher moments

A variance identity

The next identity explains why the positive similarity improves every even-moment bound, not just the quadratic one.

Lemma 5.1. If LT=LL^T=L, then

∥HLH∥F2=∥L∥F2+2∥Lu∥2+∣uTLu∣2.(30)\lVert HLH\rVert_F^2=\lVert L\rVert_F^2+2\lVert Lu\rVert^2+\lvert u^TLu\rvert^2. \tag*{(30)}
∥LK∥F2=∥L∥F2+(σ+2)∥Lu∥2.(31)\lVert LK\rVert_F^2=\lVert L\rVert_F^2+(\sigma+2)\lVert Lu\rVert^2. \tag*{(31)}

Proof. Expand tr⁡(L∗KLK)\operatorname{tr}(L^*KLK) using K=I+uuTK=I+uu^T to obtain (30). The two single-rank terms have the same trace because LL is symmetric and uu is real. For (31), use K2=I+(σ+2)uuTK^2=I+(\sigma+2)uu^T in tr⁡(L∗LK2)\operatorname{tr}(L^*LK^2). □

Put

Lm=D(KD)m−1.L_m=D(KD)^{m-1}.

This matrix is symmetric, Cm=HLmHC^m=HL_mH, and Qm=LmKQ^m=L_mK. Moreover,

Lmu=fm(D)u.(32)L_mu=f_m(D)u. \tag*{(32)}

For m=1m=1 this is immediate. The induction step follows from Lm+1=DKLmL_{m+1}=DKL_m:

Lm+1u=D(fm(D)u+amu)=fm+1(D)u.L_{m+1}u=D\left(f_m(D)u+a_mu\right)=f_{m+1}(D)u.

Lemma 5.1 therefore proves (17). Its right-hand side is nonnegative by Cauchy–Schwarz, with equality precisely when all fm(bj)f_m(b_j) coincide. Together with Schur’s inequality, this proves the sandwich (16).

Suppose now that fm(bj)=af_m(b_j)=a for every jj. Then Lmu=auL_mu=au and, by symmetry, uTLm=auTu^TL_m=au^T, so LmL_m commutes with KK. Consequently

Qm=LmK=HLmH=CmQ^m=L_mK=HL_mH=C^m

commutes with both KK and QQ, hence with D=QK−1D=QK^{-1}. Since D∗D^* is a polynomial in DD (interpolate conjugation on its finite set of distinct diagonal entries), QmQ^m commutes with D∗D^* and therefore with Q∗=KD∗Q^*=KD^*. In particular, QmQ^m is normal. Conversely, normality of QmQ^m gives equality between the endpoints of (16), so (17) vanishes. This proves part (1) of Theorem 3.4.

Divided differences

For a polynomial ff, write

f[x,y]={f(x)−f(y)x−y,x≠y,f′(x),x=y.f[x,y]= \begin{cases} \dfrac{f(x)-f(y)}{x-y}, & x\ne y,\\ f'(x), & x=y. \end{cases}

Lemma 5.2. For m≥1m\ge1,

(Qm)ij=δijbim+biuiujfm[bi,bj].(33)(Q^m)_{ij}=\delta_{ij}b_i^m+b_i u_i u_j f_m[b_i,b_j]. \tag*{(33)}

Furthermore,

[Cm,C∗]=HΩmH,(Ωm)ij=uiuj(bi‾ bj‾−bi‾bj)fm[bi,bj].(34)[C^m,C^*]=H\Omega_mH,\qquad(\Omega_m)_{ij}=u_i u_j(\overline{b_i}\,\overline{b_j}-\overline{b_i}b_j)f_m[b_i,b_j]. \tag*{(34)}

Proof. All generating series below may be read formally. Set

Ψ(t)=1−∑jrjbjt1−bjt.\Psi(t)=1-\sum_j\frac{r_jb_jt}{1-b_jt}.

The recurrence (15) gives

∑m≥1fm(x)tm−1=x(1−xt)Ψ(t).(35)\sum_{m\ge1}f_m(x)t^{m-1}=\frac{x}{(1-xt)\Psi(t)}. \tag*{(35)}

Indeed, summing the recurrence first gives x(1+t∑m≥1amtm−1)/(1−xt)x(1+t\sum_{m\ge1}a_mt^{m-1})/(1-xt); summing this at x=bjx=b_j with weights rjr_j identifies the parenthesis as 1/Ψ(t)1/\Psi(t).

The rank-one inverse formula gives

((I−tQ)−1)ij=δij1−bit+tbiuiuj(1−bit)(1−bjt)Ψ(t).((I-tQ)^{-1})_{ij}=\frac{\delta_{ij}}{1-b_it}+\frac{tb_iu_iu_j}{(1-b_it)(1-b_jt)\Psi(t)}.

Taking the divided difference in xx of (35) and comparing coefficients proves (33), also when bi=bjb_i=b_j.

Finally,

H−1[Cm,C∗]H−1=LmKD∗−D∗KLm.H^{-1}[C^m,C^*]H^{-1}=L_mKD^*-D^*KL_m.

Use LmK=QmL_mK=Q^m, KLm=(Qm)TKL_m=(Q^m)^T, and (5.4). The diagonal terms cancel, leaving (5.5). □

Completion of Theorem 3.4. For any matrix XX, normality of XmX^m is equivalent to

[Xm,X∗]=0.(36)[X^m,X^*]=0. \tag*{(36)}

One implication follows by commuting with (X∗)m(X^*)^m. For the other, if XmX^m is normal, its adjoint is a polynomial in XmX^m, by the spectral theorem and interpolation on the finite spectrum. Since XX commutes with XmX^m, it commutes with (Xm)∗(X^m)^*; taking adjoints gives (5.7).

If the bjb_j are collinear, CC itself is normal. Otherwise, partition the indices according to the lines through the origin containing bjb_j. There are at least two classes. For indices in different classes, bˉibj−bˉjbi≠0\bar b_i b_j-\bar b_jb_i\ne0 and bi≠bjb_i\ne b_j. Thus (5.5) and (5.7) force fm(bi)=fm(bj)f_m(b_i)=f_m(b_j). Any two indices in the same class can be joined through an index in another class, so all fm(bj)f_m(b_j) coincide. The converse follows from part (1) and Cm=QmC^m=Q^m when the variance vanishes. Finally, fmf_m is monic of degree mm, so a level set of fmf_m contains at most mm distinct points. □

Remark 5.3 (Polynomial functions). The same argument applies to an arbitrary polynomial h(t)=∑m=0dcmtmh(t)=\sum_{m=0}^{d}c_mt^m. Put fh=∑m=1dcmfmf_h=\sum_{m=1}^{d}c_mf_m. Then

∑j∣h(λj)∣2≤∥h(C)∥F2≤∥h(Q)∥F2,\sum_j|h(\lambda_j)|^2\le\|h(C)\|_F^2\le\|h(Q)\|_F^2,
∥h(Q)∥F2−∥h(C)∥F2=σ∑jrj∣fh(bj)∣2−∣∑jrjfh(bj)∣2.\|h(Q)\|_F^2-\|h(C)\|_F^2=\sigma\sum_jr_j|f_h(b_j)|^2-\left|\sum_jr_jf_h(b_j)\right|^2.

The matrix h(Q)h(Q) is normal exactly when the fh(bj)f_h(b_j) coincide; h(C)h(C) is normal exactly when this holds or the bjb_j are collinear. To include the constant term in the norm calculation, use L=c0K−1+∑m≥1cmLmL=c_0K^{-1}+\sum_{m\ge1}c_mL_m in Lemma 5.1. The vector LuLu differs from fh(D)uf_h(D)u by a constant multiple of uu, which does not change the variance. The commutator formula follows by linearity, and the finite-spectrum argument above applies with h(X)h(X) in place of XmX^m.

The quartic bound and higher-order examples

Proof of Corollary 3.5. With v=Duv = Du, one has L2=D2+vvTL_2 = D^2 + vv^T. Directly,

∥L2∥F2=V0+2V1+S2,∥L2u∥2=W,uTL2u=T+s2.\lVert L_2\rVert_F^2 = V_0 + 2V_1 + S^2, \qquad\lVert L_2u\rVert^2 = W, \qquad u^T L_2u = T + s^2.

Lemma 5.1 gives both right-hand sides of (19).

It remains to identify equality. If f2(bj)f_2(b_j) is constant, the monic quadratic f2(t)=t2+stf_2(t) = t^2 + st takes that value at at most two distinct points. For two such points a,ca, c,

0=f2(a)−f2(c)a−c=a+c+s=(1+Ra)a+(1+Rc)c.0 = \frac{f_2(a)-f_2(c)}{a-c} = a + c + s = (1+R_a)a + (1+R_c)c.

Thus aa and cc lie on opposite rays. The alternative in part (2) of Theorem 3.4 therefore adds no noncollinear configurations when m=2m=2. Part (1) gives exactly (20) for equality in the unbalanced bound. □

At higher orders, noncollinear equality does occur.

Example 5.4 (A noncollinear sixth-moment equality). Take n=2n=2, r1=r2=1r_1=r_2=1, and

b1=1,b2=c=−5+i398.b_1 = 1, \qquad b_2 = c = \frac{-5+i\sqrt{39}}{8}.

Then 4c2+5c+4=04c^2+5c+4=0. For Q=D(I+J)Q=D(I+J), put τ=tr⁡Q=2(1+c)\tau=\operatorname{tr}Q=2(1+c). Its determinant is 3c=τ23c=\tau^2, so Cayley–Hamilton gives

Q2−τQ+τ2I=0,Q3=−τ3I.Q^2-\tau Q+\tau^2I=0, \qquad Q^3=-\tau^3I.

Consequently C3=Q3C^3=Q^3 is scalar, although CC is not normal. Hence equality holds throughout (16) for m=3m=3. In polynomial terms, this is the cubic P(z)=z(z−1)(z−1/c)P(z)=z(z-1)(z-1/c).

Example 5.5 (Regular polygons). Let rj=ρ>0r_j=\rho>0 and bj=βωj−1b_j=\beta\omega^{j-1}, where β≠0\beta\ne0 and ω=e2πi/n\omega=e^{2\pi i/n}. The determinant identity in Lemma 4.1 reduces to

det⁡(λI−Q)=λn−(1+nρ)βn.\det(\lambda I-Q)=\lambda^n-(1+n\rho)\beta^n.

Indeed, if g(λ)=λn−βng(\lambda)=\lambda^n-\beta^n, then ∑jbj∏ℓ≠j(λ−bℓ)=λg′(λ)−ng(λ)=nβn\sum_j b_j\prod_{\ell\ne j}(\lambda-b_\ell)=\lambda g'(\lambda)-ng(\lambda)=n\beta^n. Thus Qn=Cn=(1+nρ)βnIQ^n=C^n=(1+n\rho)\beta^nI. For n≥3n\ge3 this gives noncollinear equality at order 2n2n and at every positive multiple of that order.

Generating functions and stability

A closed generating determinant

All the quadratic exterior refinements can be collected in one rank-two determinant.

Proposition 6.1. For t≥0t\ge0, set

dj(t)=1+t∣bj∣2,L(t)=∑jrjt∣bj∣2dj(t),B(t)=∑jrjbjdj(t).d_j(t)=1+t|b_j|^2, \qquad L(t)=\sum_j\frac{r_jt|b_j|^2}{d_j(t)}, \qquad B(t)=\sum_j\frac{r_jb_j}{d_j(t)}.

Then

∏j(1+t∣λj∣2)≤det⁡(I+tC∗C)=∏jdj(t)((1+L(t))2+t∣B(t)∣2).(37)\prod_j(1+t|\lambda_j|^2)\le\det(I+tC^*C)=\prod_jd_j(t)\left((1+L(t))^2+t|B(t)|^2\right). \tag*{(37)}

The inequality holds coefficient by coefficient as a polynomial in tt. For any fixed t>0t>0, equality holds if and only if the bjb_j are collinear.

Proof. The coefficient of tkt^k in det⁡(I+tC∗C)\det(I+tC^\ast C) is ∥C[k]∥F2=Bk\lVert C^{[k]}\rVert_F^2=B_k. Theorem 3.3 proves the inequality and its equality statement (with the one-dimensional case automatic).

For the determinant formula, write E=diag⁡(dj(t))E=\operatorname{diag}(d_j(t)), v=Duv=Du, and s=∑jrjbjs=\sum_j r_jb_j. The matrix C∗CC^\ast C is similar to D∗KDKD^\ast KDK, and

D∗KDK=∣D∣2+∣D∣2uuT+vˉ(vT+suT).D^\ast KDK=\lvert D\rvert^2+\lvert D\rvert^2uu^T+\bar{v}(v^T+su^T).

The rank-two determinant identity therefore reduces the quotient by det⁡E\det E to the determinant of

(1+LtBˉs−B+sL1+L+tsBˉ).\begin{pmatrix} 1+L & t\bar{B}\\ s-B+sL & 1+L+ts\bar{B} \end{pmatrix}.

Here t∑jrj∣bj∣2/dj=s−Bt\sum_j r_j\lvert b_j\rvert^2/d_j=s-B. The displayed determinant is (1+L)2+t∣B∣2(1+L)^2+t\lvert B\rvert^2, as claimed. □

A quantitative collinearity estimate

Equality in the reciprocal Schoenberg inequality detects a line through the origin. Its deficit also controls the squared distance to the nearest such line. Keep S,s,TS,s,T from (18), and define

F=∥C∥F2,Δ=F−∑j∣λj∣2,dr2=min⁡θ∈R∑jrj(Im⁡(e−iθbj))2.F=\lVert C\rVert_F^2,\qquad\Delta=F-\sum_j\lvert\lambda_j\rvert^2,\qquad d_r^2=\min_{\theta\in\mathbb{R}}\sum_jr_j\bigl(\operatorname{Im}(e^{-i\theta}b_j)\bigr)^2.

Proposition 6.2. Let

G=S2−∣T∣2+2(S∣s∣2−Re⁡(sˉ2T)),η=max⁡jrj−1+2+σ.G=S^2-\lvert T\rvert^2+2\bigl(S\lvert s\rvert^2-\operatorname{Re}(\bar{s}^2T)\bigr),\qquad\eta=\max_jr_j^{-1}+2+\sigma.

Then

Δ≥G8∥C∥op2≥G8F,Δ≥dr24η.(38)\Delta\ge\frac{G}{8\lVert C\rVert_{\mathrm{op}}^2}\ge\frac{G}{8F},\qquad\Delta\ge\frac{d_r^2}{4\eta}. \tag*{(38)}

Also,

G=12∥[C,C∗]∥F2=tr⁡((C∗C)2)−∥C2∥F2.(39)G=\frac{1}{2}\lVert[C,C^\ast]\rVert_F^2=\operatorname{tr}\bigl((C^\ast C)^2\bigr)-\lVert C^2\rVert_F^2. \tag*{(39)}

No optimality is asserted for the constants in (38).

Proof. Put M=vv∗−vˉvTM=vv^\ast-\bar{v}v^T, so [C,C∗]=HMH[C,C^\ast]=HMH. Straightforward rank-one products give

∥M∥F2=2(S2−∣T∣2),uTMu=0,∥Mu∥2=2(S∣s∣2−Re⁡(sˉ2T)).\lVert M\rVert_F^2=2(S^2-\lvert T\rvert^2),\qquad u^TMu=0,\qquad\lVert Mu\rVert^2=2\bigl(S\lvert s\rvert^2-\operatorname{Re}(\bar{s}^2T)\bigr).

Expanding tr⁡(MKMK)\operatorname{tr}(MKMK) gives ∥[C,C∗]∥F2=2G\lVert[C,C^\ast]\rVert_F^2=2G. Expanding the square of the commutator proves the second identity in (39).

In a unitary Schur basis write C=Λ+NC=\Lambda+N, where Λ\Lambda is diagonal and NN is strictly upper triangular. Then ∥N∥F2=Δ\lVert N\rVert_F^2=\Delta, and

[C,C∗]=[C,N∗]+[N,Λ∗].[C,C^\ast]=[C,N^\ast]+[N,\Lambda^\ast].

Since ∥Λ∥op≤∥C∥op\lVert\Lambda\rVert_{\mathrm{op}}\le\lVert C\rVert_{\mathrm{op}}, it follows that ∥[C,C∗]∥F≤4∥C∥opΔ\lVert[C,C^\ast]\rVert_F\le4\lVert C\rVert_{\mathrm{op}}\sqrt{\Delta}. This proves the first two estimates.

Finally, minimizing the elementary trigonometric expression defining dr2d_r^2 gives dr2=(S−∣T∣)/2d_r^2=(S-\lvert T\rvert)/2. The second term in GG is nonnegative because ∣T∣≤S\lvert T\rvert\le S, and hence

G≥(S−∣T∣)(S+∣T∣)≥2Sdr2.G\ge(S-\lvert T\rvert)(S+\lvert T\rvert)\ge2Sd_r^2.

Cauchy–Schwarz gives ∣s∣2≤σS\lvert s\rvert^2\le\sigma S, so (21) implies F≤ηSF\le\eta S. Substitution proves the last estimate in (38). □

AI Declaration: Artificial intelligence (AI) tools were used to test the inequalities in this paper and to provide examples. An LLM also provided feedback on an earlier draft, pointed out several extensions to its original results, and assisted in re-writing of this paper. The author assumes full responsibility for the accuracy of all results presented here.

References

References

  1. [1]H. Araki, On an inequality of Lieb and Thirring, Lett. Math. Phys. 19 (1990), 167–170. doi:10.1007/BF01045887.DOI
  2. [2]R. Bhatia, Matrix Analysis, Graduate Texts in Mathematics, vol. 169, Springer, New York, 1997.
  3. [3]J. Borcea, Equilibrium points of logarithmic potentials induced by positive charge distributions. I. Generalized de Bruijn–Springer relations, Trans. Amer. Math. Soc. 359 (2007), 3209–3237. arXiv:math/0601519.arxiv.org/abs/math/0601519
  4. [4]W.-S. Cheung and T.-W. Ng, A companion matrix approach to the study of zeros and critical points of a polynomial, J. Math. Anal. Appl. 319 (2006), 690–707. doi:10.1016/j.jmaa.2006.06.071.
  5. [5]F. Hiai, Log-majorization and matrix norm inequalities with application to quantum information, Acta Sci. Math. (Szeged) 90 (2024), 527–549. arXiv:2402.16067.arxiv.org/abs/2402.16067
  6. [6]O. Kushel and M. Tyaglov, Circulants and critical points of polynomials, J. Math. Anal. Appl. 439 (2016), 634–650. doi:10.1016/j.jmaa.2016.03.005.DOI
  7. [7]S. M. Malamud, Inverse spectral problem for normal matrices and the Gauss–Lucas theorem, Trans. Amer. Math. Soc. 357 (2005), 4043–4064. arXiv:math/0304158.DOI
  8. [8]R. Pereira, Differentiators and the geometry of polynomials, J. Math. Anal. Appl. 285 (2003), 336–348. doi:10.1016/S0022-247X(03)00465-7.DOI
  9. [9]I. J. Schoenberg, A conjectured analogue of Rolle’s theorem for polynomials with real or complex coefficients, Amer. Math. Monthly 93 (1986), 8–13.DOI
  10. [10]Q. Tang and T. Zhang, Sharp Schoenberg type inequalities and the de Bruijn–Sharma problem, preprint, arXiv:2508.10341v3, 2025.arxiv.org/abs/2508.10341

Paper details

vlnc:W_x7ytsn3x
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 11

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents