\section{A five-state example: a feasible reduced fit that overestimates}\label{sec:family} Consider a true chain with two visible states and only three hidden states. Its observed data admit exact fits with two hidden states, so the reduced class is feasible, yet every such fit has entropy production far above that of the true chain, and the relative overestimate diverges as a parameter $\varepsilon\downarrow0$. The theorems of the later sections explain this behaviour in general: the unrestricted floor is reached with four hidden states, while for unequal conditional cross-excursion means two hidden states cannot reach it. \subsection{The true family and exact reduced fits} In the order $(1,2,h_1,h_2,h_3)$ set \begin{equation} \pi=(11/40,11/40,1/10,3/10,1/20),\qquad N_\varepsilon=\begin{pmatrix} 0&1&3&1&3\varepsilon/4\\ 1+\varepsilon/2&0&1&3&\varepsilon/4\\ 3&1&0&1&0\\1&3&1&0&0\\ \varepsilon/4&3\varepsilon/4&0&0&0 \end{pmatrix},\quad0<\varepsilon\le10^{-3}. \label{eq:true-family} \end{equation} Rows and columns of the flow table have the same sums. Define off-diagonal rates by $q_{xy}=N_{xy}/\pi_x$ and choose the negative row sum on the diagonal. The graph is connected and reciprocal, so the model is admissible. Direct edge calculation gives \begin{equation}\sigma(Q_\varepsilon)=\varepsilon\log3+ \frac{\varepsilon}{2}\log(1+\varepsilon/2) \le\frac{8001}{4000000}<\frac1{400}.\label{eq:true-entropy}\end{equation} The small direct-edge term has coefficient $\varepsilon/2$. Put \begin{equation} R_0=\frac13\begin{pmatrix}7&5\\5&7\end{pmatrix},\quad S=\frac1{16}\begin{pmatrix}3&9\\1&3\end{pmatrix},\quad L=\frac1{2880}\begin{pmatrix}251&337\\265&443\end{pmatrix}. \end{equation} The observations are $R_\varepsilon=R_0+\varepsilon S$, this fixed $L$, visible masses $11/40,11/40$, and direct flows $1,1+\varepsilon/2$. To verify $R,L$, the fast hidden pair has entrance and exit flow matrix $B_f=\bigl(\begin{smallmatrix}3&1\\1&3\end{smallmatrix}\bigr)$, killed flow matrix $J_f=\bigl(\begin{smallmatrix}5&-1\\-1&5\end{smallmatrix}\bigr)$, and hidden masses $D_f=\diag(1/10,3/10)$. Its matrices are \[B_fJ_f^{-1}B_f=R_0,\qquad B_fJ_f^{-1}D_fJ_f^{-1}B_f=\frac1{90}\begin{pmatrix}7&8\\8&13\end{pmatrix}. \] The slow state's occurrence and first-moment contributions are $\varepsilon S,S/20$, because its mean residence time is $1/(20\varepsilon)$. An exact two-hidden-state fit exists throughout the interval. Define \begin{equation} X=\begin{pmatrix}2/3&1/3\\1/3&2/3\end{pmatrix},\quad Y=\begin{pmatrix}55/84&47/116\\29/84&69/116\end{pmatrix},\quad D_h=\diag(7/80,29/80), \end{equation} \begin{equation} U_\varepsilon=R_\varepsilon X^{-1},\quad V_\varepsilon=Y^{-1}R_\varepsilon, \quad J_\varepsilon=Y^{-1}R_\varepsilon X^{-1}. \label{eq:fit-construction} \end{equation} Use $U,V$ for entrance and exit flows, $-(J_\varepsilon)_{12},-(J_\varepsilon)_{21}$ for hidden internal flows, and hidden masses from $D_h$. Keep the visible masses and direct flows. Rows of $X$ and columns of $Y$ sum to one, giving $J_\varepsilon\one=V_\varepsilon\one$ and $\one^TJ_\varepsilon=\one^TU_\varepsilon$. These establish hidden balance and identify the killed flow matrix. Visible balance follows from $(R_\varepsilon)_{12}-(R_\varepsilon)_{21}=\varepsilon/2$. All these flows are affine in $\varepsilon$. At zero, \[U_0=\begin{pmatrix}3&1\\1&3\end{pmatrix},\quad V_0=\frac1{76}\begin{pmatrix}217&14\\87&290\end{pmatrix},\quad J_0=\frac1{76}\begin{pmatrix}420&-189\\-116&493\end{pmatrix}. \] At $\varepsilon=1/1000$ the complete four-state flow table is \begin{equation}\frac1{304000}\begin{pmatrix} 0&304000&911943&304285\\ 304152&0&303981&912095\\ 868105&56315&0&755475\\ 347971&1159913&463971&0 \end{pmatrix}. \end{equation} Positivity at both endpoints proves reciprocal complete support throughout. Finally \[U_\varepsilon J_\varepsilon^{-1}V_\varepsilon=R_\varepsilon, \qquad U_\varepsilon J_\varepsilon^{-1}D_hJ_\varepsilon^{-1}V_\varepsilon =YD_hX=L. \] Thus these are exact feasible fits, not approximate numerical reconstructions. \begin{figure}[tb]\centering\input{figures/network.tex} \caption{True hidden subnetwork in \eqref{eq:true-family}. The fast pair $h_1,h_2$ remains strongly connected to both visible states, while the rates through $h_3$ are proportional to $\varepsilon$. Its stationary mass stays $1/20$, so its long residence times contribute a finite amount to the first moments. All depicted links are reciprocal. The direct visible link is also present.}\label{fig:network}\end{figure} \subsection{A weighted excursion bound} For $s>0$ let $W_{ij}(s)=\int_0^\infty(1-e^{-st})f_{ij}(t)\,dt$. Pairing each microscopic path with its reverse gives a nonnegative integrand $(w-w^{\rm rev})\log(w/w^{\rm rev})$. Multiplication by the reversal-invariant weight $1-e^{-st}\in[0,1]$ decreases it. Discard returns and apply log-sum on the cross class. This proves \begin{equation}\sigma(Q)\ge(W_{12}-W_{21})\log(W_{12}/W_{21}) \ge\frac{2(W_{12}-W_{21})^2}{W_{12}+W_{21}}.\label{eq:weighted-bound}\end{equation} Stationary normalization is kept in this argument: $W$ consists of rates, not separately normalized conditional probabilities. For any two-hidden-state exact fit, \eqref{eq:H12} gives $W=R-R(R+L)^{-1}R$ at $s=1$ in the fixed inverse-time unit. At $\varepsilon=1/1000$, \[W=\frac1{996369600}\begin{pmatrix}82443151&109353773\\86314757&143879311\end{pmatrix},\] and hence every fit satisfies \begin{equation}\sigma(Q')\ge\frac{2764563845043}{507703580647625}>\frac1{200} >\sigma(Q_{1/1000})+\frac1{400}.\end{equation} As $\varepsilon\downarrow0$, the forced matrix converges to \[W_0=\frac1{62277}\begin{pmatrix}5153&6835\\5395&8993\end{pmatrix},\] whose bound in \eqref{eq:weighted-bound} is $138240/25388257>1/200$. Continuity therefore keeps the fitted infimum bounded below while \eqref{eq:true-entropy} tends to zero. The relative overestimate diverges. No true endpoint generator at $\varepsilon=0$ is used as an admissible model. Theorem~\ref{thm:family-overestimate} below replaces this endpoint-and-limit argument by a bound that holds uniformly on the whole parameter interval and is larger by an order of magnitude. \subsection{Every reduced fit overestimates the true entropy production} Write $d_\varepsilon$ for the observed data of $Q_\varepsilon$. The true chain has three hidden states, so $Q_\varepsilon\in\mathcal C_3(d_\varepsilon)$ and $I_3(d_\varepsilon)\le\sigma(Q_\varepsilon)$. \begin{theorem}[Two-hidden-state fits overestimate the true value]\label{thm:family-overestimate} For every $0<\varepsilon\le10^{-3}$ the class $\mathcal C_2(d_\varepsilon)$ is nonempty, its infimum is attained, and \begin{equation} I_\infty(d_\varepsilon)=I_4(d_\varepsilon)\le I_3(d_\varepsilon)\le\sigma(Q_\varepsilon)<\frac{11}{10^4}, \qquad I_2(d_\varepsilon)\ge I_\infty(d_\varepsilon)+\frac7{100}. \label{eq:family-overestimate} \end{equation} Consequently every exact fit $Q'$ with at most two hidden states satisfies \[ \sigma(Q')>\sigma(Q_\varepsilon)+\frac6{100},\qquad \sigma(Q')>60\,\sigma(Q_\varepsilon),\qquad \frac{I_2(d_\varepsilon)}{\sigma(Q_\varepsilon)}\ge \frac{7}{100\,\varepsilon(\log3+\varepsilon/4)}\xrightarrow[\varepsilon\downarrow0]{}\infty . \] \end{theorem} \begin{proof} The fits \eqref{eq:fit-construction} show $\mathcal C_2(d_\varepsilon)\ne\varnothing$. Since $R_\varepsilon=R_0+\varepsilon S$, \[ \det R_\varepsilon=\frac{16-\varepsilon}{6}>0,\qquad (R_\varepsilon)_{12}+(R_\varepsilon)_{21}=\frac{10}3+\frac{5\varepsilon}8, \] so Theorems~\ref{thm:H1}, \ref{thm:H2} and \ref{thm:H3} apply: $I_\infty=I_4$, the reduced infimum is attained, and \eqref{eq:H4} holds. The inequalities $I_4\le I_3\le\sigma(Q_\varepsilon)$ hold because $\mathcal C_3\subset\mathcal C_4$ and $Q_\varepsilon\in\mathcal C_3(d_\varepsilon)$. Direct substitution of $R_\varepsilon$ and the fixed $L$ into the quantities of Theorem~\ref{thm:H3} gives \[ \Delta m=\frac{32(15-16\varepsilon)}{5(80+3\varepsilon)(80+27\varepsilon)},\qquad a=\frac1{60}\Bigl(5-\frac3{16-\varepsilon}\Bigr),\qquad b=\frac{19}{1200(16-\varepsilon)} . \] On $[0,10^{-3}]$, $\Delta m$ and $a$ decrease and $b$ increases. Hence $\Delta m\ge\Delta m(10^{-3})=95897600/6402400081>149/10^4$, $a\le a(0)=77/960$ and $b\le b(10^{-3})=95/95994<10^{-3}$. Since $77/960+2\cdot10^{-3/2}<1435/10^4$, \eqref{eq:H4} gives \[ I_2(d_\varepsilon)-I_\infty(d_\varepsilon)\ge \frac{2\cdot\frac{10}{3}\,(149/10^4)^2}{(1435/10^4)^2} =\frac{88804}{1235535}>\frac7{100}. \] By \eqref{eq:true-entropy} and $\log(1+x)\le x$, $\sigma(Q_\varepsilon)\le\varepsilon\log3+\varepsilon^2/4$, which is below $11/10^4$ because $\log3<1.0987$. As $I_\infty\ge0$, $I_2(d_\varepsilon)\ge7/100$, which exceeds $\sigma(Q_\varepsilon)+6/100$ and $60\,\sigma(Q_\varepsilon)$, and the displayed ratio bound follows. Every $Q'\in\mathcal C_2(d_\varepsilon)$ has $\sigma(Q')\ge I_2(d_\varepsilon)$. \end{proof} Thus restricting the hidden dimension from three to two does more than lose the lower-bound property: at the observed data of this chain, the minimal reduced fit reports more than sixty times the entropy production of the chain that generated the data. Theorem~\ref{thm:family-overestimate} concerns this explicit family; it is not a claim that $I_2(d(Q))>\sigma(Q)$ for every true chain $Q$ with unequal cross means, since $\sigma(Q)$ may exceed $I_\infty(d(Q))$ by more than the reduction bias. The constant $7/100$ is a lower bound, not the value of $I_2(d_\varepsilon)$; the explicit fits \eqref{eq:fit-construction} have entropy production approximately $2.06$ at $\varepsilon=10^{-3}$, which bounds $I_2(d_\varepsilon)$ from above. Theorem~\ref{thm:family-sharp} below raises the lower bound to $7/5$, using a computer-assisted step in exact arithmetic. \subsection{A sharper bound by direct parametrization} Theorem~\ref{thm:H3} uses only one Laplace weight of the cross-excursion durations. A direct description of all two-hidden-state fits gives a much larger bound. \begin{theorem}[Sharper bound, computer-assisted]\label{thm:family-sharp} For every $0<\varepsilon\le10^{-3}$, every $Q'\in\mathcal C_2(d_\varepsilon)$ satisfies $\sigma(Q')\ge7/5$. Hence \[ I_2(d_\varepsilon)\ge\frac75>I_\infty(d_\varepsilon)+\frac{139}{100},\qquad \sigma(Q')>1200\,\sigma(Q_\varepsilon). \] \end{theorem} \begin{proof} \emph{Parametrization.} Let $Q'\in\mathcal C_2(d_\varepsilon)$. A single hidden state gives a rank-one $R$ by \eqref{eq:H11}, and $\det R_\varepsilon\ne0$, so $Q'$ has exactly two hidden states. Take $A,C,K$ as in \eqref{eq:H11}, let $D_h=\diag(d_1,d_2)$ hold the hidden masses, and put $X=K^{-1}C$. The entrance, exit and killed flow matrices of $Q'$ are $U=A$, $V=D_hC$ and $J=D_hK$. Zero row sums give $K\one=C\one$, so $X\one=\one$. The entry $X_{gj}$ is the probability that an excursion started in $g$ exits to $j$, so $X\ge0$, and $R_\varepsilon=AX$ makes $X$ invertible. Relabel the hidden states so that $X$ has rows $(x,1-x)$ and $(y,1-y)$ with $x>y$, and put $\delta=x-y$. By \eqref{eq:H11}, $U=R_\varepsilon X^{-1}$ and $AK^{-1}=LX^{-1}$. Stationarity of the hidden masses is $(d_1,d_2)K=\one^TA$, hence $(d_1,d_2)=\one^TLX^{-1}$. Also $K^{-1}=A^{-1}LX^{-1}=XR_\varepsilon^{-1}LX^{-1}$, so \[ V=D_hXL^{-1}R_\varepsilon,\qquad J=D_hXL^{-1}R_\varepsilon X^{-1}. \] With $Y=LX^{-1}D_h^{-1}$ these are the formulas \eqref{eq:fit-construction}; the fits displayed there have $x=2/3$, $y=1/3$. Substituting $R_\varepsilon$ and $L$ gives $d_1=(43-108y)/(240\delta)$, $d_2=(108x-43)/(240\delta)$, \[ U=\frac1{48\delta}\begin{pmatrix} 112-192y-\varepsilon(36y-9)&192x-112+\varepsilon(36x-9)\\ 80-192y-\varepsilon(12y-3)&192x-80+\varepsilon(12x-3) \end{pmatrix}, \] $V_{1j}=(43-108y)\varphi_j(x)/(912\delta)$ and $V_{2j}=(108x-43)\varphi_j(y)/(912\delta)$, where \[ \varphi_1(t)=336t-100+\varepsilon(48t-17),\qquad \varphi_2(t)=3\bigl(24-32t+\varepsilon(48t-17)\bigr). \] The hidden internal flows are $-J_{12}=(108y-43)\,\psi(x)/(912\delta^2)$ and $-J_{21}=(108x-43)\,\psi(y)/(912\delta^2)$, where $\psi(t)=240t^2-364t+100+\varepsilon(192t^2-116t+17)$. \emph{Location.} Positive hidden masses give $y<43/1080$, $V_{12}\ge0$ gives $x\le(24-17\varepsilon)/(32-48\varepsilon)$, which increases in $\varepsilon$, so $x\le752/1000$. Since $108x-43>0$, $V_{21}\ge0$ gives $y\ge(100+17\varepsilon)/(336+48\varepsilon)\ge25/84>2976/10^4$. Finally $-J_{21}\ge0$ gives $\psi(y)\ge0$. Now $\psi$ is convex in $t$ and affine in $\varepsilon$, and $\psi(361/1000)$ and $\psi(43/108)$ are negative for $\varepsilon=0$ and for $\varepsilon=10^{-3}$. Hence $\psi<0$ on $[361/1000,43/108]$, and $y<361/1000$. Every fit therefore has $(x,y)$ in the rectangle $\mathfrak B=[583/1000,752/1000]\times[2976/10^4,361/1000]$, on which $\delta>1/5$. \emph{Entropy bound.} Drop the nonnegative direct term. For a reciprocal pair with flows $a,b>0$, $\log r\ge2(r-1)/(r+1)$ for $r\ge1$ gives $(a-b)\log(a/b)\ge g(a,b):=2(a-b)^2/(a+b)$. An absent pair has $a=b=0$ by reciprocal support and contributes zero; set $g(0,0)=0$. Hence $\sigma(Q')\ge G$, where $G$ is the sum of $g$ over the five hidden-incident pairs $(U_{11},V_{11})$, $(U_{12},V_{21})$, $(U_{21},V_{12})$, $(U_{22},V_{22})$ and $(-J_{12},-J_{21})$. \emph{Finite check.} On $\{a\ge b\ge0\}$ the function $g$ increases in $a$ and decreases in $b$, and symmetrically on $\{b\ge a\ge0\}$. Hence, if $a\in[a_-,a_+]$ and $b\in[b_-,b_+]$ with $a_->b_+\ge0$, then $g(a,b)\ge g(a_-,b_+)$, and similarly with the roles exchanged. The rectangle $\mathfrak B$ was bisected adaptively into $2824$ subrectangles. On each subrectangle, times $\varepsilon\in[0,10^{-3}]$, every closed-form flow above was enclosed by interval arithmetic with exact rational endpoints. Negative lower endpoints were replaced by $0$, because all flows of a fit are nonnegative. This gave a lower bound at least $7/5$ for $G$ on every subrectangle. This step is computer-assisted. It uses only exact rational arithmetic and is performed by the script \texttt{anc/certify\_I2.py} supplied with the source files. The script also re-derives the closed forms symbolically. Hence $\sigma(Q')\ge7/5$. The remaining claims follow from $I_\infty(d_\varepsilon)\le\sigma(Q_\varepsilon)<11/10^4$ (Theorem~\ref{thm:family-overestimate}): $7/5-11/10^4>139/100$ and $1200\cdot11/10^4<7/5$. \end{proof} Numerical minimization over $\mathfrak B$ suggests $I_2(d_\varepsilon)\approx1.52$, attained on the boundary $U_{12}=V_{21}=0$. This value is not proved and is not used.