---
title: 'Extended Vertical LCP (EVLCP): Theory & Methods'
url: https://www.emergentmind.com/topics/extended-vertical-linear-complementarity-problem-evlcp
type: topic
---

# Extended Vertical LCP (EVLCP): Theory & Methods

The extended vertical linear complementarity problem (EVLCP) is the problem of finding \(x\in\mathbb R^n\) such that
\[
\min\{A_0x+q_0,\;A_1x+q_1,\;\dots,\;A_kx+q_k\}=0,
\]
with the minimum taken componentwise, for given matrices \(A_0,\dots,A_k\in\mathbb R^{n\times n}\) and vectors \(q_0,\dots,q_k\in\mathbb R^n\). In alternative notation, the block tuple may be written as \(\mathbf M=(M_0,\dots,M_k)\) and \(\mathbf q=(q_0,\dots,q_k)\). EVLCP contains the standard linear complementarity problem and the vertical LCP as special cases, and has been studied through solvability theory, global error bounds, perturbation analysis, and continuous-time dynamical solvers [2202.13036] [2210.01268] [2507.20652].

## 1. Formulation and scope

For positive integers \(n\) and \(k\), EVLCP\((\mathbf M,\mathbf q)\) asks for \(x\in\mathbb R^n\) satisfying
\[
\min\{\,M_0x+q_0,\;M_1x+q_1,\;\dots,\;M_kx+q_k\}=0.
\]
The componentwise minimum makes the problem piecewise linear, but not globally linear. The standard LCP is recovered when \(k=1\), \(M_0=I\), \(M_1=M\), and \(q_0=0\). The vertical LCP is recovered when \(M_0=I\) and \(q_0=0\); in the literature summarized here, the vertical LCP is associated with Cottle–Dantzig.

The formulation unifies several model classes. Reported applications include network theory, control, game theory, and Hamilton–Jacobi–Bellman equations, as well as piecewise-linear electrical networks, singular control problems, real option valuation, generalized Leontief input-output models, and generalized bimatrix games. The breadth of this application list reflects the fact that the componentwise minimum naturally represents switching, selection, and regime-dependent linear structure rather than a single linear law.

A common misconception is that EVLCP is only a notational extension of the LCP. The literature treats it instead as a genuine umbrella model: the number of blocks is arbitrary, the minimum is taken across all blocks, and the algebraic conditions governing uniqueness and stability are correspondingly more intricate than in the single-matrix LCP setting.

## 2. Row \(\mathcal W\)-property and unique solvability

The central structural condition is the row \(\mathcal W\)-property. A block matrix \(\mathbf M=(M_0,\dots,M_k)\) has the row \(\mathcal W\)-property if, for every \(x\in\mathbb R^n\),
\[
\min(M_0x,\dots,M_kx)\le 0\le \max(M_0x,\dots,M_kx)
\quad\Longrightarrow\quad x=0,
\]
where the inequalities are componentwise. This condition is known to be equivalent to unique solvability of EVLCP for every right-hand side: EVLCP\((\mathbf M,\mathbf q)\) has a unique solution \(x^*\) for every \(\mathbf q\) if and only if \(\mathbf M\) has the row \(\mathcal W\)-property [2202.13036].

A determinant-type characterization, attributed to Gowda–Sznajder in the summarized material, states that the row \(\mathcal W\)-property holds if and only if for every choice of nonnegative diagonal matrices \(X_0,\dots,X_k\) with \(X_0+\cdots+X_k\) positive on the diagonal,
\[
\det(X_0A_0+\cdots+X_kA_k)\neq 0.
\]
A directly operational equivalent form uses diagonal selectors:
\[
S_M=\sum_{i=0}^k D_i M_i,\qquad
\mathcal D=\Bigl\{(D_0,\dots,D_k)\;\bigm|\;D_i=\operatorname{diag}(d_i),\ d_i\in[0,1]^n,\ \sum_{i=0}^k D_i=I\Bigr\},
\]
and \(\mathbf M\) has the row \(\mathcal W\)-property if and only if \(S_M\) is nonsingular for every \((D_0,\dots,D_k)\in\mathcal D\) [2210.01268].

Two additional necessary and sufficient conditions were highlighted in the 2022 error-bound study. First, the row \(\mathcal W\)-property is equivalent to nonsingularity of
\[
D_0A_0+D_1A_1+\cdots+D_kA_k
\]
for every diagonal choice \(D_j=\operatorname{diag}(d_j)\) with \(d_j\in[0,1]^n\) and \(\sum_{j=0}^k D_j=I\). Second, \((A_0,\dots,A_k)\) has the row \(\mathcal W\)-property if and only if, for every diagonal \(D\in[0,I]\), the two-block system
\[
(A_0,\;D_1A_1+\cdots+D_kA_k)
\]
is also row \(\mathcal W\), with \(D_1+\cdots+D_k=I\) [2202.13036]. These formulations matter because they replace combinatorial reasoning over row patterns by diagonal convex-combination tests.

## 3. Global error bounds

A major development in the recent EVLCP literature is a new derivation of global error bounds from a scalar identity for minima. For real numbers \(a_j,b_j\), there exist \(\lambda_j\in[0,1]\) with \(\sum_j\lambda_j=1\) such that
\[
\min_j a_j-\min_j b_j=\sum_j \lambda_j(a_j-b_j).
\]
Applied rowwise to EVLCP, with \(x^*\) denoting the unique solution and
\[
r(x):=\min\{A_0x+q_0,\dots,A_kx+q_k\},
\]
this yields diagonal matrices \(D_0,\dots,D_k\) satisfying \(\sum_{j=0}^k D_j=I\) and
\[
r(x)=\sum_{j=0}^k D_jA_j(x-x^*).
\]
Under the row \(\mathcal W\)-property, the matrix
\[
M:=D_0A_0+\cdots+D_kA_k
\]
is nonsingular, and therefore
\[
\|x-x^*\|
\le
\left(
\max_{D_0+\cdots+D_k=I,\;D_j\ \mathrm{diag.}\ge 0}
\|M^{-1}\|
\right)\,\|r(x)\|.
\]
The norm may be any vector-matrix norm. This bound is global and two-sided in the sense used in the source exposition, and its derivation avoids row permutations entirely [2202.13036].

The comparison with earlier bounds is explicit. In the standard LCP setting, Mathias–Pang obtained
\[
\|x-x^*\|
\le
\bigl(1+\|A_1\|\alpha(A_1)\bigr)\,\|r(x)\|,
\qquad
\alpha(A_1)=\min_{\|z\|_\infty=1}\max_i (A_1z)_i^+.
\]
Chen–Xiang showed, in any \(p\)-norm,
\[
\|x-x^*\|_p
\le
\max_{D\in[0,I]}\|(I-D+DA_1)^{-1}\|_p\,\|r(x)\|_p,
\]
which is equivalent to the EVLCP bound above when \(A_0=I\) and \(k=1\). Zhang–Chen–Xiu proved, under the row \(\mathcal W\)-property,
\[
\|x-x^*\|\le a(A)\,\|r(x)\|,
\]
with
\[
a(A)=\max_{A'\in \operatorname{RowRearr}(A),\;D\in[0,I]}
\|((I-D)A'_j+DA'_\ell)^{-1}\|,
\]
requiring exploration of row rearrangements and pair selections. The reported computational cost grows like \(((k+1)!)^n\), whereas the newer bound needs only a single worst-case inverse norm over diagonal convex combinations and thus eliminates the factorial blow-up [2202.13036].

An illustrative example with \(k=2\) and \(n=2\) used
\[
A_0=\begin{bmatrix}1&1\\[2pt]1&-2\end{bmatrix},\quad
A_1=\begin{bmatrix}1&0\\[2pt]0&1\end{bmatrix},\quad
A_2=\begin{bmatrix}2&1\\[2pt]1&0\end{bmatrix}.
\]
The row-rearrangement approach requires inspection of \(3!^2=36\) rearrangements and over a hundred inverse-norm evaluations, while the new theorem reduces the task to
\[
\max_{D_0+D_1+D_2=I}\|(D_0A_0+D_1A_1+D_2A_2)^{-1}\|_\infty=3,
\]
obtained through a small convex-programming problem over the simplex of selector variables. This example shows that the simplification is not merely formal; it directly reduces workload in bound estimation.

## 4. Perturbation bounds and sensitivity analysis

Perturbation theory for EVLCP analyzes how the solution changes when both the block matrices and the vectors are perturbed. With
\[
\tilde M_i=M_i+\Delta M_i,\qquad
\tilde q_i=q_i+\Delta q_i,
\]
assume that both \(\mathbf M\) and \(\tilde{\mathbf M}\) have the row \(\mathcal W\)-property, and let \(x^*\) and \(y^*\) be the corresponding unique solutions. A convex-combination identity analogous to the error-bound argument produces diagonal matrices \(\tilde D_i\) such that
\[
x^*-y^*
=
\sum_{i=0}^k
\left(\sum_{j=0}^k \tilde D_jM_j\right)^{-1}
\tilde D_i\bigl[(\tilde M_i-M_i)y^*+(\tilde q_i-q_i)\bigr].
\]
This leads to the amplification factors
\[
\alpha_i(M)
=
\max_{(D_j)\in\mathcal D}\|S_M^{-1}D_i\|,
\qquad
S_M=\sum_{j=0}^k D_jM_j,
\]
and the absolute perturbation bound
\[
\|x^*-y^*\|
\le
\sum_{i=0}^k
\alpha_i(M)\,\|(\tilde M_i-M_i)y^*+\tilde q_i-q_i\|.
\]
If only the vectors are perturbed, then
\[
\|x^*-y^*\|
\le
\sum_{i=0}^k \alpha_i(M)\,\|\tilde q_i-q_i\|.
\]
These formulas generalize sensitivity bounds known for the LCP and VLCP [2210.01268].

For simultaneous matrix and vector perturbations, the theory introduces
\[
\mathfrak M_\eta
=
\Bigl\{\mathbf A=(A_0,\dots,A_k)\ \bigm|\ \sum_{i=0}^k \alpha_i(M)\,\|A_i-M_i\|\le \eta<1\Bigr\},
\]
and proves that every \(\mathbf A\in\mathfrak M_\eta\) has the row \(\mathcal W\)-property together with
\[
\alpha_i(A)\le \frac{\alpha_i(M)}{1-\eta}=: \tau_i(M).
\]
Consequently, if \(\mathbf A,\mathbf B\in\mathfrak M_\eta\) and \(x^*,y^*\) solve the associated EVLCPs, then
\[
\|x^*-y^*\|
\le
\sum_{i=0}^k \tau_i(M)\,\|\bar q_i-\bar p_i\|
+
\Bigl(\sum_{i=0}^k \tau_i(M)\,\|A_i-B_i\|\Bigr)
\Bigl(\sum_{i=0}^k \tau_i(M)\,\|\bar p_i\|\Bigr).
\]

The same work also gives a relative perturbation estimate. If
\[
\|\Delta M_i\|\le \epsilon_i\|M_i\|,
\qquad
\|\Delta q_i\|\le \epsilon_i\,\|(-q_i)_+\|,
\]
and
\[
\eta=\sum_{i=0}^k \epsilon_i\,\alpha_i(M)\,\|M_i\|<1,
\]
then \(\tilde{\mathbf M}\) has the row \(\mathcal W\)-property and
\[
\frac{\|x^*-y^*\|}{\|x^*\|}
\le
\frac{2\eta}{1-\eta}.
\]
Computable bounds for \(\alpha_i(M)\) are available under additional matrix structure. If each \(M_i\) has positive diagonal and satisfies the comparison-matrix condition
\[
\rho\!\left(\max_i\{\Lambda_i^{-1}|C_i|\}\right)<1,
\]
then
\[
\alpha_i(M)\le \|\Gamma_i\|,
\qquad
\Gamma_i=(I-\max_j\{\Lambda_j^{-1}|C_j|\})^{-1}\Lambda_i^{-1}.
\]
If each \(M_i\) is strictly diagonally row-dominant, then
\[
\alpha_i(M)\le
\delta_i
:=
1\Big/\min_\ell\bigl\{(\langle M_i\rangle e)_\ell\bigr\}.
\]
These estimates convert the abstract perturbation constants into directly evaluable quantities [2210.01268].

The numerical examples in the perturbation study emphasize conservativeness rather than instability. In a two-by-two EVLCP example with \(\epsilon=10^{-3}\), the reported relative error was approximately \(1.55\times 10^{-2}\), while two bounds were approximately \(1.14\times 10^{-1}\) and \(7.36\times 10^{-2}\). In a three-by-three VLCP example with \(\epsilon=10^{-3}\), the relative error was approximately \(7.04\times 10^{-4}\) and the bound approximately \(2.84\times 10^{-2}\). For a Hamilton–Jacobi–Bellman discretization with \(n\in\{16,36,64,100\}\), the bounds tracked the actual relative error within a modest factor [2210.01268].

## 5. Reformulation as generalized absolute-value equations

A 2025 development recasts EVLCP as a special case of a new kind of generalized absolute value equations (NGAVE). The construction is recursive. One defines
\[
\phi_0(x)=B_0x+b_0,
\]
\[
\phi_1[B_1^1,b_1^1](x)=|B_1^1x+b_1^1|,
\]
\[
\phi_2[B_2^1,B_2^2,b_2^1,b_2^2](x)
=
\Bigl|B_2^1x+b_2^1+\phi_1[B_2^2,b_2^2](x)\Bigr|,
\]
and, in general,
\[
\phi_i[\dots](x)
=
\Bigl|
B_i^1x+b_i^1+\phi_1[\dots](x)+\cdots+\phi_{i-1}[\dots](x)
\Bigr|.
\]
The NGAVE of order \(s\) is
\[
\varphi_s(x)=\phi_0(x)+\sum_{i=1}^s \phi_i[\dots](x)=0.
\]
The paper shows that EVLCP can be embedded into this nested absolute-value framework [2507.20652].

For the case \(s=2\),
\[
\min\{A_0x+q_0,\;A_1x+q_1,\;A_2x+q_2\}=0
\quad\Longleftrightarrow\quad
\varphi_2(x)=0,
\]
with
\[
\begin{aligned}
B_0&=-(2\Omega A_0+A_1+A_2),\\
B_1^1&=A_1-A_2,\qquad
B_2^1=2\Omega A_0-A_1-A_2,\qquad
B_2^2=A_1-A_2,
\end{aligned}
\]
and
\[
\begin{aligned}
b_0&=-(2\Omega q_0+q_1+q_2),\\
b_1^1&=q_1-q_2,\qquad
b_2^1=2\Omega q_0-q_1-q_2,\qquad
b_2^2=q_1-q_2,
\end{aligned}
\]
where \(\Omega\) is any positive diagonal matrix. Similar nested-absolute-value representations are stated for \(s=3,4\), and higher orders.

Under the spectral condition
\[
\sigma_{\min}(B_0)>\sum_{i,j}\|B_i^j\|,
\]
the NGAVE \(\varphi_s(x)=0\) admits a unique solution for every choice of the vectors \(b_0,b_i^j\). The proof proceeds by constructing
\[
M(x)=-B_0^{-1}\Bigl(\sum_i \phi_i(x)+b_0\Bigr)
\]
and showing that \(M\) is a contraction, so Banach’s fixed-point theorem applies. Under the same assumption, the paper gives the error-bound inequality
\[
\frac{\|\varphi_s(x)\|}{\|B_0\|+\sum_{i,j}\|B_i^j\|}
\le
\|x-x_*\|
\le
\frac{\|\varphi_s(x)\|}{\sigma_{\min}(B_0)-\sum_{i,j}\|B_i^j\|}.
\]
This route supplies a distinct sufficient condition for unique solvability and an accompanying residual-based bound, but now expressed in the NGAVE variables rather than directly in the row \(\mathcal W\)-framework [2507.20652].

## 6. Fixed-time dynamical solution methods and computational interpretation

The NGAVE reformulation supports a fixed-time stable continuous-time solver. The model is
\[
\frac{dx}{dt}=-\rho(x)\,g(\gamma,x),
\qquad
g(\gamma,x)=\gamma\,B_0^\top\,\varphi_s(x),
\]
where \(\rho(x)\) is chosen so that the system is continuous and yields fixed-time convergence, with parameters \(\rho_1,\rho_2>0\), \(\lambda_1\in(0,1)\), \(\lambda_2>1\), and \(\gamma>0\). With the Lyapunov function
\[
V(x)=\tfrac12\|x-x_*\|^2,
\]
the dynamics satisfy
\[
\frac{dV}{dt}\le -c_1V^{\kappa_1}-c_2V^{\kappa_2},
\qquad
\kappa_1=\frac{\lambda_1+1}{2}\in(0.5,1),\quad
\kappa_2=\frac{\lambda_2+1}{2}>1.
\]
Using a fixed-time stability lemma of Polyakov (2011), every trajectory reaches \(x_*\) in at most
\[
T_{\max}
=
\frac{1}{c_1(1-\kappa_1)}
+
\frac{1}{c_2(\kappa_2-1)},
\]
independently of the initial condition [2507.20652].

The reported numerical study considered three test sets with \(s=2,3,4\) and problem sizes up to \(n=m\times m\) for \(m=20,40,60,80\). The EVLCP instances were constructed so that the exact solution \(x_*\) was known and the condition \(\sigma_{\min}(B_0)>\sum\|B_i^j\|\) was verifiable. The baseline was the smoothing-neural-network model of Hou–Zhang–Qiu (2022), and the metrics were the residual \(\|\min\{A_jx+q_j\}\|\le 10^{-5}\), CPU time, and the relative error \(\|x-x_*\|/\|x_*\|\). The fixed-time dynamical system was reported to converge in CPU time orders of magnitude smaller than the neural network, to reach the \(10^{-5}\) residual tolerance uniformly, and to yield relative errors on the order of \(10^{-7}\)–\(10^{-8}\), compared with \(10^{-5}\)–\(10^{-6}\) for the baseline. Trajectory plots of \(\|x(t)-x_*\|\) were reported to confirm convergence well before the theoretical \(T_{\max}\), indicating that the analytic bound is conservative but valid [2507.20652].

Taken together, the recent literature presents two complementary computational viewpoints. One viewpoint, based on the row \(\mathcal W\)-property and diagonal selectors, focuses on sharp residual-to-solution error bounds and perturbation constants without combinatorial row rearrangement. The other viewpoint, based on NGAVE reformulation, produces inverse-free fixed-time dynamics and a separate spectral sufficient condition for uniqueness. This suggests that EVLCP is best understood not as a single algorithmic object, but as a family of piecewise-linear systems admitting multiple analytic lenses: complementarity-theoretic, perturbational, and dynamical.

Source: https://www.emergentmind.com/topics/extended-vertical-linear-complementarity-problem-evlcp