Rank-one optimal solution conjecture for dissipative bilinear control semidefinite programs

Let AA be the system matrix in the semidefinite program associated with the dissipative bilinear control system, and let A+ATA+A^T denote its symmetric part. An optimal solution is a positive semidefinite matrix MM satisfying the semidefinite-program constraints.

Rank-one optimal solution conjecture. If

A+AT0,A+A^T\prec 0,

then the semidefinite program has an optimal solution of rank r=1r=1.

The preceding proposition proves this conclusion under the additional hypothesis that AA is rr-diagonal, with the rank bound matching that value of rr. The conjecture asserts that dissipativity alone suffices for a rank-one optimal solution; the source reports strong numerical evidence but does not provide a proof.

Progress summary

Solved

A posted calculation claims to disprove the conjecture with a strictly dissipative four-dimensional example having no rank-one optimum, but the claim has not been independently verified.

The 2005 manuscript formulates the conjecture that strict dissipativity, A+AT0A+A^T\prec0, guarantees an optimal solution of rank 11. It reports strong numerical evidence but gives no proof.

Known results

  • For 2×22\times2 and 3×33\times3 system matrices, an optimal solution of rank at most 11 exists (2005 manuscript).
  • If AA is rr-diagonal, an optimal solution has rank at most rr (2005 manuscript).
  • Under A+AT0A+A^T\prec0, an optimum exists with rank at most (8n+11)/2\left\lfloor(\sqrt{8n+1}-1)/2\right\rfloor (2005 manuscript).

Posted attempt

A complete counterexample is posted: it claims an irreducible matrix with A+AT=2I40A+A^T=-2I_4\prec0 whose semidefinite program has a unique optimal solution of rank 22, certified by an explicit dual matrix. If correct, this disproves the conjecture even with strictly positive constrained initial data. The calculation has not been independently verified.

Current status (as of August 2026): The special cases and general existence bound are settled, while the unrestricted conjecture has an unverified claimed counterexample rather than a confirmed resolution.

Sources
Sources & referencesView supporting material

Primary source

Dionisis Stefanatos and Navin Khaneja, “Semidefinite Programming and Reachable Sets of Dissipative Bilinear Control Systems”, arXiv:math/0504308 (2005).

Solutions 1

Counterexample

Consider the source semidefinite program

maxN0A4,N,Ai,N=pi(0)(1i3),\max_{N\succeq0}\langle A_4,N\rangle, \qquad \langle A_i,N\rangle=-p_i(0)\quad(1\le i\le3),

where C,N=tr(CN)\langle C,N\rangle=\operatorname{tr}(CN) and, for a real matrix A=(aij)A=(a_{ij}), the symmetric matrices AiA_i are defined by

(Ai)ii=2aii,(Ai)ij=(Ai)ji=aij(ji),(A_i)_{ii}=2a_{ii},\qquad (A_i)_{ij}=(A_i)_{ji}=a_{ij}\quad(j\ne i),

with all remaining entries zero.

Take

A=(14/37/654/31420/177/6411520/1711).A= \begin{pmatrix} -1&-4/3&-7/6&5\\ 4/3&-1&-4&20/17\\ 7/6&4&-1&1\\ -5&-20/17&-1&-1 \end{pmatrix}.

Every off-diagonal entry is nonzero, so AA is irreducible. Moreover,

A+AT=2I40.A+A^{\mathsf T}=-2I_4\prec0.

Define

U=(1/513/51/21001),R=(1114),M=URUT.U= \begin{pmatrix} -1/5&-1\\ -3/5&-1/2\\ 1&0\\ 0&1 \end{pmatrix}, \qquad R= \begin{pmatrix} 1&-1\\ -1&4 \end{pmatrix}, \qquad M=URU^{\mathsf T}.

Since detR=3\det R=3 and R11=1R_{11}=1, we have R0R\succ0, so M0M\succeq0 has rank two. Explicitly,

M=(91/2571/504/519/571/5019/251/107/54/51/101119/57/514).M= \begin{pmatrix} 91/25&71/50&4/5&-19/5\\ 71/50&19/25&-1/10&-7/5\\ 4/5&-1/10&1&-1\\ -19/5&-7/5&-1&4 \end{pmatrix}.

Its four constraint values are

(A1,M,A2,M,A3,M,A4,M)=(76415,58255,4415,60017).\bigl( \langle A_1,M\rangle, \langle A_2,M\rangle, \langle A_3,M\rangle, \langle A_4,M\rangle \bigr) = \left( -\frac{764}{15}, -\frac{58}{255}, -\frac{44}{15}, \frac{600}{17} \right).

Thus MM is feasible for strictly positive initial data

(p1(0),p2(0),p3(0))=(76415,58255,4415).(p_1(0),p_2(0),p_3(0)) = \left( \frac{764}{15}, \frac{58}{255}, \frac{44}{15} \right).

To certify global optimality, let

B=(53041355),S=125BBT,B= \begin{pmatrix} 5&3\\ 0&4\\ 1&3\\ 5&5 \end{pmatrix}, \qquad S=\frac1{25}BB^{\mathsf T},

and

(w1,w2,w3)=(1725,825,15).(w_1,w_2,w_3) = \left(\frac{17}{25},\frac8{25},\frac15\right).

Direct rational calculation gives

S=w1A1w2A2w3A3A40,BTU=0.S=-w_1A_1-w_2A_2-w_3A_3-A_4\succeq0, \qquad B^{\mathsf T}U=0.

In particular, SM=0SM=0. Every feasible N0N\succeq0 satisfies

0S,N=i=13wipi(0)A4,N=60017A4,N.0\le\langle S,N\rangle = \sum_{i=1}^3w_ip_i(0)-\langle A_4,N\rangle = \frac{600}{17}-\langle A_4,N\rangle.

Equality holds at N=MN=M, so the exact optimum is 600/17600/17.

Finally, if NN is any optimal feasible matrix, then

0=S,N=125tr(BTNB).0=\langle S,N\rangle = \frac1{25}\operatorname{tr}(B^{\mathsf T}NB).

Since BTNB0B^{\mathsf T}NB\succeq0, it vanishes, and hence NB=0NB=0. Therefore

imNkerBT=imU,\operatorname{im}N\subseteq\ker B^{\mathsf T} =\operatorname{im}U,

so N=UXUTN=UXU^{\mathsf T} for a symmetric 2×22\times2 matrix XX. The three feasibility equations have coefficient matrix

Q=(1/157/340/322/5166/5135/102109/1513/30),detQ=1140170.Q= \begin{pmatrix} 1/15&-7/3&-40/3\\ 22/5&166/51&-35/102\\ -109/15&-13/3&0 \end{pmatrix}, \qquad \det Q=-\frac{1140}{17}\ne0.

Thus they uniquely force X=RX=R. Consequently MM is the unique optimum, and

rankM=2.\operatorname{rank}M=2.

There is no rank-one optimum. The conjecture is therefore false even for an irreducible strictly dissipative system with strictly positive constrained initial data.

0 endorsements
Shivam Patel ·