Han–Sabag–Kostina–Hassibi: independent counterexample audit

What Lean checks

Lean proves the original matrix feasibility, rational witness, exact sparse infimum and all four individual assignment infima, plus the uniform gap throughout the stated open gain interval. Standalone strict source compilation, all-owned axiom audit and stock-kernel replay passed; the merged-package gate is pending.

Status: the counterexample and sparse lower bound pass independent mathematical and source-scope review. Prior-art review remains with the root agent; this note alone does not establish novelty.

Exact source

Barron Han, Oron Sabag, Victoria Kostina, and Babak Hassibi, Coded Kalman Filtering over MIMO Gaussian Channels with Feedback, arXiv:2406.17196v2 (5 August 2025), Conjecture 1, equations (23)–(24), printed page 6.

Primary full text: https://arxiv.org/pdf/2406.17196

The stated optimization problem fixes diagonal BB and HH, requires diagonal Π≽0\Pi\succeq0, and asks for J≻0J\succ0 satisfying

J=BJB−ΓT(I+HΠH)−1Γ+BΓTΓB.J=BJB-\Gamma^T(I+H\Pi H)^{-1}\Gamma+B\Gamma^T\Gamma B.

It minimizes trΠ\mathop{\mathrm{tr}}\Pi over these variables and conjectures that an optimum has exactly one nonzero entry in each column of Γ\Gamma. There is no further constraint relating Γ\Gamma to Π\Pi in the stated problem. The source discusses related constraints from the underlying communication model later, but those must not silently be inserted into equations (23)–(24).

The example below uses positive distinct diagonal entries of BB strictly below one, so there is no issue with the source's abbreviated condition bi<1b_i<1 or with uniqueness of the Lyapunov solution. It disproves equality of the dense and sparse infima, which is stronger than exploiting the source's possibly unattained-optimum wording.

Exact feasible point

Take

B=diag(45,35),H=diag(74,1),Π=diag(11681323,317),Γ=(1153408401961).B=\mathop{\mathrm{diag}}\left(\frac45,\frac35\right),\quad H=\mathop{\mathrm{diag}}\left(\frac74,1\right),\quad \Pi=\mathop{\mathrm{diag}}\left(\frac{1168}{1323},\frac3{17}\right), \quad \Gamma=\begin{pmatrix}1&1\\[2pt]\frac{53}{40}&\frac{840}{1961}\end{pmatrix}.

Then

D=(I+HΠH)−1=diag(27100,1720).D=(I+H\Pi H)^{-1} =\mathop{\mathrm{diag}}\left(\frac{27}{100},\frac{17}{20}\right).

Because BB is diagonal and |bibj|<1|b_i b_j|<1, the unique solution has entries

Jij=∑ℓ=12γℓiγℓj(bibj−dℓ)1−bibj.J_{ij}=\frac{\sum_{\ell=1}^2 \gamma_{\ell i}\gamma_{\ell j}(b_i b_j-d_\ell)}{1-b_i b_j}.

Direct rational arithmetic gives

J=diag(21157600,35289246113344)≻0,trΠ=2382522491.J=\mathop{\mathrm{diag}}\left(\frac{211}{57600}, \frac{35289}{246113344}\right)\succ0, \qquad \mathop{\mathrm{tr}}\Pi=\frac{23825}{22491}.

The off-diagonal cancellation is exact:

53408401961=2137=12/25−27/10017/20−12/25.\frac{53}{40}\frac{840}{1961}=\frac{21}{37} =\frac{12/25-27/100}{17/20-12/25}.

Both diagonal entries are strictly positive. Thus this is an exact feasible point, not a numerical optimizer output.

Complete sparse classification

Every column must be nonzero because a zero column forces the corresponding diagonal entry of JJ to vanish. With exactly one nonzero entry per column and two channels, there are precisely four assignment patterns.

If both columns use the same channel with diagonal value d=(1+h2π)−1d=(1+h^2\pi)^{-1}, nonzero column scaling is a diagonal congruence and does not change positive definiteness. The relevant matrix is

K(d)ij=bibj−d1−bibj.K(d)_{ij}=\frac{b_i b_j-d}{1-b_i b_j}.

For distinct 0<b1,b2<10<b_1,b_2<1,

detK(d)=(b1−b2)2(1−d)(b12b22−d)(1−b12)(1−b22)(1−b1b2)2.\det K(d)=\frac{(b_1-b_2)^2(1-d)(b_1^2b_2^2-d)} {(1-b_1^2)(1-b_2^2)(1-b_1b_2)^2}.

Since 0<d≤10<d\leq1, K(d)≻0K(d)\succ0 holds exactly when d<b12b22d<b_1^2b_2^2. Necessity follows from the determinant and the diagonal sign at d=1d=1; sufficiency follows from the positive diagonal entries and determinant. Here b12b22=144/625b_1^2b_2^2=144/625, so the power infimum for a common channel is (481/144)/h2(481/144)/h^2.

If the columns use different channels, JJ is diagonal. A mode bib_i assigned to channel hh needs d<bi2d<b_i^2, equivalently π>(bi−2−1)/h2\pi>(b_i^{-2}-1)/h^2. This condition is also sufficient. The two mode costs before division by the channel squared gain are 9/169/16 and 16/916/9.

For h1=7/4h_1=7/4, h2=1h_2=1, the four exact sparse power infima are therefore:

Assignment Infimum
Both modes to channel 1 481/441481/441
Both modes to channel 2 481/144481/144
Mode 4/54/5 to channel 1, mode 3/53/5 to channel 2 865/441865/441
Mode 4/54/5 to channel 2, mode 3/53/5 to channel 1 8065/70568065/7056

The minimum is 481/441481/441. Thus

infall feasible ΓtrΠ≤2382522491<481441=infone nonzero per columntrΠ,\inf_{\text{all feasible }\Gamma}\mathop{\mathrm{tr}}\Pi \leq\frac{23825}{22491} <\frac{481}{441} =\inf_{\text{one nonzero per column}}\mathop{\mathrm{tr}}\Pi,

with exact positive gap

481441−2382522491=70622491.\frac{481}{441}-\frac{23825}{22491}=\frac{706}{22491}.

The strict inequalities required by J≻0J\succ0 only mean that the sparse thresholds are infima. The dense feasible point is strictly below all of them, so allowing limits does not restore the conjecture.

A continuous family of counterexamples

Keep B,D,ΓB,D,\Gamma as above. For s=h2>0s=h^2>0 put

H=diag(h,1),Π=diag(7327s,317).H=\mathop{\mathrm{diag}}(h,1),\qquad \Pi=\mathop{\mathrm{diag}}\left(\frac{73}{27s},\frac3{17}\right).

This retains the same positive definite JJ and has dense feasible power 73/(27s)+3/1773/(27s)+3/17. The four sparse thresholds are

481144s,481144,916s+169,169s+916.\frac{481}{144s},\qquad \frac{481}{144},\qquad \frac9{16s}+\frac{16}{9},\qquad \frac{16}{9s}+\frac9{16}.

The dense point lies strictly below all four whenever

1360567<s<46751296.\frac{1360}{567}<s<\frac{4675}{1296}.

For the first and fourth comparisons these are exactly the upper and lower bounds. The two other comparisons give weaker lower bounds, as direct rational cross multiplication confirms. In particular the example lies in a nonempty open interval of channel gains; the contradiction is not tied to a floating-point special case.

Limits of this review

This review checks the conjecture as actually stated. It does not assert a separate counterexample to every engineering interpretation in the paper, nor does it need a reconstruction of the full coded Kalman filter. The source's conditional Theorem 2 cannot be invoked after its conjectural hypothesis fails; further consequences require their own derivation.

Independent review: algebra agent, 21 November 2026.