Schur stability conjecture for image reconstruction operators

Let WW be a primitive stochastic matrix, let BB be a positive semidefinite matrix satisfying

Beϱ(B)eBe \leqslant \varrho(B)e

and

πBe>0,\pi^\top Be>0,

where π\pi is the left Perron eigenvector of WW. Define

P(t)=W(ItB),R(t)=IW+(I+tB)1(2WI).P(t)=W(I-tB),\qquad R(t)=I-W+(I+tB)^{-1}(2W-I).

Schur stability conjecture. The operators P(t)P(t) and R(t)R(t) are stable, meaning that

ϱ(P(t))<1andϱ(R(t))<1,\varrho(P(t))<1\quad\text{and}\quad\varrho(R(t))<1,

for all 0<t<2/ϱ(B)0<t<2/\varrho(B).

Progress summary

Solved

An unverified posted calculation claims that explicit two-dimensional examples disprove both stability assertions, while the published paper only proves stability for sufficiently small steps and selected cases.

Banerjee and Chaudhury posed the conjecture in September 2025: under the stated hypotheses, both reconstruction operators should remain stable throughout the prescribed step-size interval.

Known results

  • Banerjee–Chaudhury (2025): both operators are stable for each sufficiently small t>0t>0, although the proof gives no uniform estimate for the threshold.
  • Banerjee–Chaudhury (2025): explicit stability ranges are obtained for particular classes of WW and BB.
  • Banerjee–Chaudhury (2025): the bound Beϱ(B)eBe\leqslant\varrho(B)e is shown necessary for the conjectured formulation; numerical examples indicate that the bound for R(t)R(t) need not be sharp.

Posted attempt

An unverified calculation claims that a positive-definite 2×22\times2 matrix BB and primitive stochastic 2×22\times2 matrices satisfy all hypotheses while both operators have spectral radius greater than 11 at t=4/5t=4/5. It further claims a parameterized family with instability for arbitrarily small positive tt, which would refute both assertions rather than merely prove partial progress. The calculation has not been independently verified.

Current status (as of August 2026): the local and selected special-case results are established, while the general conjecture is now accompanied by an unverified claimed counterexample and is not mathematically settled.

Sources
Sources & referencesView supporting material

Primary source

Debraj Banerjee and Kunal Narayan Chaudhury, “On the Schur Stability of Some Image Reconstruction Operators”, arXiv:2509.06715 (2025).

Solutions 1

Counterexample

Counterexample to both stability assertions, with arbitrarily small instability thresholds.

Conjecture 1.1 of Banerjee and Chaudhury, arXiv:2509.06715, considers

P(t)=W(ItB),R(t)=IW+(I+tB)1(2WI),P(t)=W(I-tB),\qquad R(t)=I-W+(I+tB)^{-1}(2W-I),

where WW is primitive and row-stochastic, BB is positive semidefinite, e=(1,,1)e=(1,\ldots,1)^\top, and the normalized left Perron vector π\pi satisfies

Beρ(B)e,πBe>0.Be\le \rho(B)e,\qquad \pi^\top Be>0.

It asserts that both spectral radii are strictly less than one throughout 0<t<2/ρ(B)0<t<2/\rho(B). The conjecture does not require BB to be entrywise nonnegative: the preceding discussion explicitly says that nonnegativity is not essential, and Example 1.2 includes negative off-diagonal entries.

Take

W=(017/3023/30),B=111(9443),t=45.W=\begin{pmatrix}0&1\\7/30&23/30\end{pmatrix}, \qquad B=\frac1{11}\begin{pmatrix}9&-4\\-4&3\end{pmatrix}, \qquad t=\frac45.

Indeed,

W2=(7/3023/30161/900739/900)>0,W^2= \begin{pmatrix}7/30&23/30\\161/900&739/900\end{pmatrix}>0,

so WW is primitive. Furthermore,

σ(B)={1,111},ρ(B)=1,Be=111(51)(11).\sigma(B)=\left\{1,\frac1{11}\right\},\qquad \rho(B)=1,\qquad Be=\frac1{11}\binom5{-1}\le\binom11.

Thus BB is even positive definite. The stationary vector is

π=137(730),πBe=5407>0,\pi=\frac1{37}\binom7{30}, \qquad \pi^\top Be=\frac5{407}>0,

and 0<4/5<2/ρ(B)=20<4/5<2/\rho(B)=2.

Exact rational calculation gives

P(4/5)=(16/5543/55167/550367/550),R(4/5)=(700/1593545/1593253/1593015797/15930).P(4/5)= \begin{pmatrix} 16/55&43/55\\ 167/550&367/550 \end{pmatrix}, \qquad R(4/5)= \begin{pmatrix} 700/1593&545/1593\\ 253/15930&15797/15930 \end{pmatrix}.

Nevertheless,

det(IP(4/5))=21375<0,det(IR(4/5))=22655<0.\det(I-P(4/5))=-\frac2{1375}<0, \qquad \det(I-R(4/5))=-\frac2{2655}<0.

For each real two-by-two matrix MM, its monic characteristic polynomial therefore has χM(1)<0\chi_M(1)<0 and χM(x)+\chi_M(x)\to+\infty as x+x\to+\infty. Each matrix consequently has a real eigenvalue strictly greater than one:

ρ(P(4/5))>1,ρ(R(4/5))>1.\rho(P(4/5))>1,\qquad \rho(R(4/5))>1.

Hence both claimed stability assertions fail simultaneously.

In fact the obstruction occurs arbitrarily close to zero. Keep the same positive definite BB, choose any 0<s<20<s<2, and put

v=\frac1{5-s},\qquad W_s=\begin{pmatrix}0&1\v&1-v\end{pmatrix}, \qquad \pi_s=\frac1{1+v}\binom v1.

Then WsW_s is primitive and stochastic, BeeBe\le e, and

πsBe=s11(6s)>0.\pi_s^\top Be=\frac{s}{11(6-s)}>0.

Direct symbolic calculation yields

det(IPs(t))=t(ts)11(5s),det(IRs(t))=t(ts)(5s)(t+1)(t+11).\det(I-P_s(t)) =-\frac{t(t-s)}{11(5-s)}, \qquad \det(I-R_s(t)) =-\frac{t(t-s)}{(5-s)(t+1)(t+11)}.

At t=st=s, both operators already have eigenvalue one, and for every s<t<2s<t<2, both have a real eigenvalue greater than one. Since ss can be arbitrarily small while ρ(B)=1\rho(B)=1 remains fixed, no positive stability threshold depending only on ρ(B)\rho(B) and the dimension is valid under the conjectured assumptions. The paper's separate local per-instance stability theorem remains unaffected.

0 endorsements
Shivam Patel ·