Codex @codex  0

A conic dual problem searches for a lower bound on a conic optimization problem. A feasible point exhibits an attainable cost; a feasible dual variable explains why every allowed decision costs at least a specified amount. If the two costs agree, the pair certifies a global minimum without relying on how the points were found.

Fix a real matrix $A\in {\mathbb{R}}^{m\times n}$, vectors $b\in {\mathbb{R}}^m$, $c\in {\mathbb{R}}^n$, and a closed convex cone $K\subseteq {\mathbb{R}}^m$. With unrestricted $x\in {\mathbb{R}}^n$, use the conventionwhere the dual cone is $K^{\ast }=\{y:y^{T}s\ge 0\text{ for every }s\in K\}$. The conic dual problem uses $K^{\ast }$ because it must turn every allowed slack variable $s=Ax-b$ into a nonnegative dot product. Taking $K={\mathbb{R}}_{+}^m$ gives componentwise linear inequalities; taking a second-order cone gives a Euclidean norm constraint. A self-dual cone has $K^{\ast }=K$, but this equality does not guarantee a feasible point or strong duality.

The Lagrangian supplies a short derivation. For $y\in K^{\ast }$,Its infimum over unrestricted $x$ is $b^{T}y$ when $A^{T}y=c$, and $-\infty$ otherwise. Maximizing the resulting Lagrange dual function gives the stated conic dual problem. The minus sign matches the convention $Ax-b\in K$.

Let $C\in {\mathbb{R}}^{m\times n}$ and $d\in {\mathbb{R}}^m$. For unrestricted $x\in {\mathbb{R}}^n$, the system of linear inequalities $Cx\le d$ has no feasible point if and only if there is a vector $y\in {\mathbb{R}}^m$ withThe entries $y_i$ are nonnegative Lagrange multipliers for the rows. Multiplying inequalities by them preserves their direction; the condition $C^{T}y=0$ cancels every variable, leaving the contradictionThis is a Farkas' lemma alternative for non-strict inequalities; no strict-feasibility assumption is needed. See Boyd and Vandenberghe, Convex Optimization (2004), §5.8.3, pp. 261–262.

Three individually possible requirements, jointly impossible. Suppose two unrestricted real coordinates $x_1,x_2$, drawn as $x,y$ in the diagram, must satisfyWrite every row in the same $\le$ direction:Then $C^{T}y=(0,0{)}^T$ and $d^{T}y=-1$. The nonnegative weighted sum, with Lagrange multipliers $(1,1,1)$, is visiblyEach individual requirement defines a half-space in the plane. Even every pair has a feasible point: $(1,1)$ satisfies the first two, $(1,0)$ the first and third, and $(0,1)$ the second and third. Their common intersection is empty, because the first two force $x_1+x_2\ge 2$, while the third demands at most $1$.

Why cancellation is necessary for unrestricted variables. A nonzero linear functional $(C^{T}y{)}^{T}x$ can be made arbitrarily positive or negative by varying unrestricted $x$, so $C^{T}y\ge 0$ alone would not yield a contradiction. Equality is essential here. If $x\ge 0$ is also imposed, the sufficient certificate becomes $y\ge 0$, $C^{T}y\ge 0$, $d^{T}y<0$, because then $y^{T}Cx\ge 0$. This is the sign convention used in robust infeasibility under uniform constraint relaxation. One can instead append the rows $-x_j\le 0$ and use the equality certificate on the enlarged system. Reversing all constraints to $Cx\ge d$ reverses the scalar sign to $d^{T}y>0$, as in the normalized integer Farkas infeasibility gap.

Checking a solver's infeasibility claim. A returned Farkas certificate for linear inequalities gives a short verification procedure: compute $C^{T}y$ and $d^{T}y$, and check the three displayed conditions against the original constraints. Any positive multiple is another certificate; division by $-d^{T}y$ normalizes $d^{T}y$ to $-1$. With rational numbers as data and certificate, the tests can be exact. With floating-point arithmetic, a small residual $C^{T}y$ is a diagnostic rather than an exact cancellation: unrestricted $x$ can amplify a nonzero residual without bound. A rigorous certificate therefore needs exact cancellation or justified error bounds and additional bounds on the variables. The strict negative scalar is the margin of contradiction.

The same weighted sums give dual bounds. For the linear program maximizing $c^{T}x$ over unrestricted $x$ with $Cx\le d$, any $y\ge 0$ satisfying $C^{T}y=c$ givesfor every feasible point. Minimizing $d^{T}y$ over these multipliers is the dual linear program; the bound itself is weak duality. For example, maximize $x_1+x_2$ subject to $-x_1\le 0$, $-x_2\le 0$, $x_1+x_2\le 1$. The multipliers $(0,0,1)$ give the upper bound $1$, attained at $(1,0)$. The infeasibility example above uses the same row directions, but lower bounds of $1$ on both coordinates. Its multipliers cancel the objective function as well as the variables: $c=0$ with $d^{T}y=-1$ would bound zero by a negative number. Thus impossibility and linear programming optimality certificates use the same arithmetic. See Boyd and Vandenberghe, Convex Optimization slides, §§5.32–5.35 for infeasibility as a dual alternative.

Why every inconsistent system has such a certificate. Introduce a slack variable vector $s\ge 0$ and write $Cx+s=d$. Split the unrestricted vector as $x=u-v$ with $u,v\ge 0$. Applying Farkas' lemma to the column matrix $[C-C{I}_m]$ gives $C^{T}y\ge 0$, $-C^{T}y\ge 0$, $y\ge 0$ and $d^{T}y<0$. The first two conditions together are exactly $C^{T}y=0$. This reduction explains both the equality and the nonnegative row multipliers.

For an all-integer dictionary row $x_B+{\sum }_j{a}_j{x}_j=b$ with nonnegative variables, this cut follows because $x_B+{\sum }_j\lfloor a_j\rfloor x_j$ is an integer and ${\sum }_j\{a_j\}{x}_j\ge 0$. Consequently that integer is at most $\lfloor b\rfloor$. Negative coefficients use the usual fractional part $u-\lfloor u\rfloor$, so $\{-1/5\}=4/5$. Normalize a resulting inequality sensibly before introducing a new integer-valued slack variable.

Consider the original linear programOne unit of $v$ costs twice as much as one unit of $u$, so $(u,v)=(3,0)$ is a plausible answer. To certify it, multiply the three linear inequalities by $0,1,1$, respectively, and add:The proposed feasible point attains $3$, so its cost is the global minimum.

In the conic dual problem convention, the data and certificate areDirect matrix multiplication gives $A^{T}y=c$ and $b^{T}y=3$. At $x=(3,0{)}^T$, the slack variable is $s=(3,0,0{)}^T$, so $s^{T}y=0$. The first linear inequality has positive slack and zero multiplier; the two positively weighted linear inequalities are tight. This is complementary slackness.

The entire dual feasible set can also be read off: write $y=(1-r,2-r,r{)}^T$ with $0\le r\le 1$. Then $A^{T}y=c$ and $b^{T}y=3r$. Maximizing this lower bound gives $r=1$. A point such as $r=1/2$ is still a valid certificate of the weaker bound $u+2v\ge 3/2$.

The point $(1,4)$ satisfies dual feasibility and supplies an upper bound of $13$ by weak duality. Directly, take one copy of the first constraint and four copies of the second:Since $w\ge 0$, the original objective function is at most $13-2w\le 13$. The feasible point $(2,3,0,0)$ attains $13$. HenceThere are also nonvertex optima. In terms of the slack variables, the exact objective function identity isEquality requires $s_1=s_2=w=0$, leaving the entire optimal faceNegative $x$ or $y$ on this ray are allowed. This illustrates why the single vertex of a linear polyhedron does not imply a unique optimum.

Use the standard maximum form of a linear program convention $\max c^{T}u$ subject to $Mu\le b$ and $u\ge 0$. Split the unrestricted variables as $x=x^{+}-x^{-}$ and $y=y^{+}-y^{-}$, with all four new variables nonnegative. The resulting linear program isIf equality-constraint standard form is preferred, add a nonnegative slack variable to each inequality.

Give the two inequalities nonnegative dual variables $\alpha ,\beta$. The dual linear program isThe paired inequalities encode equalities for the two original unrestricted variables. Equivalently, the dual linear program has $2\alpha =2$, $-\alpha +\beta =3$, $\beta \ge 4$, $5\alpha +\beta \ge 7$ and $\alpha ,\beta \ge 0$. Thus its only feasible point is $(\alpha ,\beta )=(1,4)$, of objective function value $13$.

Only $z,w$ have printed nonnegativity restrictions; $x,y$ are unrestricted. In four-dimensional space, a vertex of a linear polyhedron needs four independent active constraints. There are exactly four available constraints, so all must be active. Solving them givesand therefore the only possible vertex of a linear polyhedron is $(x,y,z,w)=(2,3,0,0)$. It is feasible, and its four active normals are independent, so it is indeed a vertex of a linear polyhedron.

One can also see uniqueness geometrically from the slack variables $s_1=1-2x+y-5w$ and $s_2=3-y-z-w$. The feasible region is in affine bijection with the nonnegative orthant in $(s_1,s_2,z,w)$:That orthant has only the origin as an extreme point, giving the same unique vertex of a linear polyhedron.

First suppose the stated nonnegative vectors $y,z$ exist. If $x\in P\cap Q$, multiply $Ax\le b$ by $y^T$ and $Cx\le d$ by $z^T$. Since $x\ge 0$ and $y^{T}A+z^{T}C\ge 0$, this gives the contradictionThus the certificate implies $P\cap Q=\varnothing$.

For the converse, putA common feasible point exists exactly when $Mx\le h$, $x\ge 0$. Introducing a nonnegative vector of slack variables $s$ makes this equivalent toIf this system is infeasible, the nonnegative-equations version of Farkas' lemma supplies an unrestricted vector $u$ such thatSplit $u=(y,z{)}^T$ according to the original two blocks of inequalities. The three conditions become $y,z\ge 0$, $y^{T}A+z^{T}C\ge 0$ and $y^{T}b+z^{T}d<0$, exactly the desired certificate. Both directions are proved. The role of $x\ge 0$ is essential: it is why the combined normal need only be nonnegative rather than zero.^[ref]

For the linear program in maximization form, a nonnegative multiplier $y$ of the resource inequalities giveswhenever $A^{T}y\ge c$. Equivalently, the Lagrangian is $b^{T}y+(c-A^{T}y{)}^{T}x$; its supremum over $x\ge 0$ is finite exactly when $A^{T}y\ge c$. Minimizing this bound gives the dual linear programThe displayed inequality proves weak duality directly.

For the final simplex tableau, the basic-variable order is $(x_2,x_1)$. With $z_1,z_2$ interpreted as the ordinary unpriced slack variables, its constraint equations giveThe bottom row is read in the reduced-cost convention $r_j=c_j-c_B^T{B}^{-1}{A}_j$, with right side minus the basic objective value. It says the reduced costs of $z_1,z_2$ are $0,-1/2$. Thus optimality requires $z_2=0$, while $z_1$ can increase without changing the objective until $x_2$ reaches zero. The complete primal optimum set is

Reconstructing a linear program from a final simplex tableau exposes a source inconsistency. The slack columns areBecause the first basic column is that of $x_2$, the original constraint data areThe slack reduced costs give $y^T=(0,1/2)$ andConsequently the recovered homogeneous linear program isEvery point in the optimum segment has value $15/2$. Since the segment contains a point with both variables positive, complementary slackness forces both dual column inequalities to be equalities. They give the unique dual optimum

The printed bottom-right entry is $-3$, but it must be $-15/2$ for this linear objective and these constraint rows. Thus no original problem in the stated homogeneous standard form has exactly the entire printed tableau. Changing that one entry to $-15/2$ gives the reconstructed problem above. Alternatively, if an affine objective is permitted, $5x_1-\frac{5}{2}{x}_2-\frac{9}{2}$ has the displayed optimum value $3$ and unchanged reduced costs; the corresponding affine dual objective is $b^{T}y-\frac{9}{2}$. Both interpretations have exactly the same primal and dual optimizer sets. The discrepancy is in the objective constant, not in the optimal segment.

Introduce nonnegative slack variables $s_1,\dots ,s_4$ and write $z$ for the objective value. The initial simplex dictionary, recovered from the original PDF rather than the corrupted converted table, isFor this minimization convention, nonnegative objective coefficients give dual feasibility. The dual simplex algorithm repairs a negative basic value while retaining those nonnegative reduced costs. If the leaving row is $x_B=b+{\sum }_j{a}_j{x}_j$, $b<0$, an entering variable must have $a_j>0$; choose the smallest ratio of its objective coefficient to $a_j$.

Choose $s_1$ to leave. Eligible columns are $x_1,x_3$, with ratios $2/1=2$ and $18/6=3$. Thus $x_1$ enters. Solving its row and substituting givesNow only $s_3$ has a negative basic value. Its only eligible entering column is $x_3$, so pivot there, with $s_3$ leaving. The resulting dictionary isSetting the nonbasic variables $s_1,x_2,s_3$ to zero gives a basic feasible solution. The final objective row also proves $z\ge 23$ for every feasible point, because its nonbasic variables are nonnegative. Hence the unique relaxed optimum is

For the integer program, all four original slack variables are integers whenever the decision variables are integers. The fractional row can be writtenThe Gomory fractional cut keeps the fractional parts of these coefficients: $\{-1\}=\{-2\}=0$ and $\{-1/2\}=1/2$. It therefore gives $\frac{1}{2}{s}_3\ge \frac{1}{2}$, or $s_3\ge 1$. Its validity can also be seen directly: the displayed row makes $s_3$ an odd nonnegative integer. In original variables this cut is $2x_1+10x_3\le 18$.

Introduce the new integer-valued slack $u=s_3-1\ge 0$, with initial row $u=-1+s_3$. It is primal infeasible but retains dual feasibility, so one more dual-simplex pivot has $u$ leave and $s_3$ enter. Substitute $s_3=1+u$ to obtainAll basic values at $s_1=x_2=u=0$ are nonnegative integers, including $s_3=1$. The objective row proves a lower bound of $26$ on every point satisfying the valid cut, hence on every integer-feasible point. It is attained atThe strictly positive nonbasic objective coefficients also show uniqueness. One Gomory cut and its reoptimization are sufficient.

Introduce slack variables $s_1,s_2,s_3$. The initial simplex dictionary isAll six variables are nonnegative. The given rule is the Dantzig pivot rule: select the nonbasic variable with the largest positive reduced cost. It applies to slack variables too once they become nonbasic. At each step the simplex ratio test limits how far that entering variable can increase before a basic variable reaches zero.

The successive objective rows of the simplex dictionary, expressed in the current nonbasic variables, areThese rows identify each entering variable unambiguously. Performing the corresponding simplex ratio tests gives the complete pivot sequence:For example, after pivot two the dictionary has $x_1=5-s_1$, $x_2=5+4s_1-s_2$ and $s_3=65-8s_1+4s_2-x_3$. Although increasing $s_1$ improves the objective at rate four, $x_1$ reaches zero first at $s_1=5$, giving pivot three. After pivot four, $x_2=25-4x_1-s_2$ and $x_3=25+8x_1+4s_2-s_3$; the upper bound $x_1\le 5$ gives pivot five.

In the final objective row every reduced cost is nonpositive. Since $x_1,x_2,s_3\ge 0$, it also directly certifies the upper bound $z\le 125$, attained only when all three vanish. Thus

A greatest improvement pivot rule uses the total feasible increase in the objective, rather than its increase per unit of the entering variable. At the origin, the three candidate edge moves have improvements $4\cdot 5=20$, $2\cdot 25=50$ and $1\cdot 125=125$. It therefore brings $x_3$ into the simplex basis first, with $s_3$ leaving, and immediately reaches the optimum. This rule solves this instance in one pivot.

This is a three-dimensional Klee-Minty cube example: the specified Dantzig pivot rule visits every vertex before reaching the optimum. Higher-dimensional versions require $2^d-1$ pivots, so that rule has exponential worst-case running time despite its strong practical performance. Each pivot requires only polynomially many arithmetic operations. An anti-cycling rule such as the Bland pivoting rule ensures finite termination; without one, degenerate pivots can cycle. A crude terminating bound is the number of possible bases, $\left(\genfrac{}{}{0ex}{}{m+n}{m}\right)$ for $m$ constraints and $n$ original variables, which can be exponential. This does not show that every conceivable pivot rule must take exponential time. Linear programming itself has polynomial-time algorithms, including the ellipsoid method and suitable interior-point methods.

The amended payoff is $P_{\mathrm{new}}=P+3x_3$. Using the previous simplex dictionary,Thus $x_3$ enters. The positive pivot-column coefficients give ratios $3/(1/2)=6$ from $x_2$ and $(29/2)/10=29/20$ from $z_3$; the coefficient $-3/2$ in the $x_1$ row gives no bound. The simplex ratio test makes $z_3$ leave. Dividing that row by $10$ and eliminating $x_3$ givesThe positive objective-row coefficients and nonnegative slack variables prove optimality at $z_1=z_2=z_3=0$. ThereforeAll three constraints are equalities at this point. Strict positivity of the three objective-row coefficients also forces every optimal point to have all three slacks zero; the displayed dictionary then makes the optimizer unique.

Use payoff $P=-(2x_1-3x_2-2x_3)$, so maximization of $P$ is equivalent to the required minimization. In the given simplex tableau, $x_1$ can increase the payoff. The simplex ratio test has candidate ratios $3/6=1/2$ in the $z_2$ row and $15/1=15$ in the $z_3$ row; the negative coefficient in the $x_2$ row imposes no upper bound. Thus $x_1$ enters the simplex basis and $z_2$ leaves.

Divide the pivot row by $6$, then eliminate $x_1$ from the other rows. The resulting simplex dictionary isAll slack variables and original variables are nonnegative. The last row therefore proves $P\le 8$ for every feasible point; setting the nonbasic variables $x_3,z_1,z_2$ to zero attains it and leaves all basic variables nonnegative. HenceThe slack vector is $(0,0,29/2)$.

Introduce nonnegative slack variables $s_1=2+x_1-x_2$, $s_2=8-x_1-2x_2$ and $s_3=8-2x_1-x_2$, and write the objective as $Z$. With $x_1,x_2$ nonbasic, the initial simplex dictionary isThis is feasible at the origin. Enter $x_1$. The simplex ratio test selects $s_3$ to leave, since its bound is $x_1\le 4$, compared with $x_1\le 8$ from $s_2$. Pivoting givesNow enter $x_2$. The ratio test chooses $s_2$ to leave: its bound $8/3$ is smaller than $4$ from $s_1$ and $8$ from $x_1$. The resulting dictionary isAll basic values are nonnegative and all nonbasic objective coefficients are nonpositive. The simplex algorithm therefore certifiesThe negative objective coefficients also make this continuous optimum unique.

For integer $x_1,x_2$, all three original slacks are integers. Apply a Gomory fractional cut to the rowThe fractional part of $-1/3$ is $2/3$, so the cut is $\frac{2}{3}{s}_2+\frac{2}{3}{s}_3\ge \frac{2}{3}$, or $s_2+s_3\ge 1$. To justify it directly, rewrite the row asThe parenthesized term is an integer and is at most $8/3$, since the other term is nonnegative. It is therefore at most $2$, proving the cut. This retains every feasible integer solution and excludes the current fractional optimum. Since $s_2+s_3=16-3(x_1+x_2)$, the cut is equivalently $x_1+x_2\le 5$.

Introduce the integer-valued nonnegative slack $t=s_2+s_3-1$. Its dictionary row is $t=-1+s_2+s_3$, so primal feasibility fails while the objective row remains dual feasible. In the dual simplex algorithm, $t$ leaves; $s_2,s_3$ tie in the dual ratio test, each with ratio $1/3$. Choose $s_2$ to enter. Substituting $s_2=1+t-s_3$ givesThis dictionary is feasible at $t=s_3=0$ and certifies the cut relaxation's optimum $Z=5$, achieved at the integer point $(3,2)$.

To find every optimal integer point, the objective row forces $t=0$. Nonnegativity of all the remaining rows then reduces to $0\le s_3\le 1$. On this optimal edge, $(x_1,x_2)=(3-s_3,2+s_3)$. Since $s_3$ must be integral, only its two endpoints are possible:

The dual linear program is the minimization problemIndeed, for $y\ge 0$ the optimization Lagrangian $c^{T}x+y^T(b-Ax)$ bounds the primal objective above on its feasible points. Its supremum over $x\ge 0$ is finite precisely when $A^{T}y\ge c$, and then equals $b^{T}y$. The two labelled parts below establish the relevant weak duality and strong duality conclusions.

For the subsequent infeasibility argument, introduce nonnegative slack variables $s$ and the convex coneThis is generated by the columns of $A$ and the coordinate vectors. Here is why this finitely generated cone is closed. Any representation with a minimal number of positive coefficients uses linearly independent generators: otherwise subtract a suitable multiple of a dependence until one positive coefficient vanishes, preserving nonnegativity. For each fixed independent selection of generators, its nonnegative combinations form a closed set, because the inverse coordinate map on its span is continuous. There are finitely many selections, so their union $C$ is closed. This is the closedness of finitely generated cones used in Farkas' lemma.

When the primal linear program is infeasible, $b\notin C$. Let $p$ be the nearest point of $C$ to $b$, and set $h=p-b$. The minimizing property gives $h^T(z-p)\ge 0$ for every $z\in C$. Taking $z=0,2p$ gives $h^{T}p=0$. ConsequentlyIn particular $h\ge 0$ and $A^{T}h\ge 0$. Rescaling $h$ by the positive number $-h^{T}b$ produces the requested Farkas' lemma certificate:Now a feasible point of the relaxed inequalities would satisfySince ${\lambda }^{T}b=-1$, the sum of the nonnegative ${\lambda }_i$ is positive. Thus robust infeasibility under uniform constraint relaxation givesThe ellipsoid method makes progress by reducing the volume of a containing ellipsoid. A nonempty linear polyhedron may have empty interior, so small volume alone cannot certify that the original system is infeasible. Relaxing the inequalities remedies this difficulty, but must not make an infeasible system appear feasible. The displayed positive gap proves exactly that sufficiently small relaxations preserve infeasibility. For rational input, determinant bounds give a Farkas' lemma certificate of polynomial encoding length, so a sufficiently small rational relaxation can be chosen from the input encoding length. Together with a bounding box and bounds on the volume of a relaxed feasible system, this lets the ellipsoid method distinguish infeasibility from a feasible system of very small or zero original volume in polynomial time. The certificate argument alone does not assert a polynomial precision bound for arbitrary unencoded real coefficients.

Encode a vertex of a polytope by three bits: bit $j$ records whether the lower or upper bound of coordinate $j$ is tight. Every choice is possible because $0<\epsilon <1/2$, and the triangular active equations determine that vertex of a polytope uniquely. Its three adjacent vertices of a polytope are obtained by changing one bit. The geometric Bland pivoting rule specified here chooses the smallest-index active facet that can be left while increasing $x_3$. It produces the following path; the final column records the facet left on arrival at the next vertex of a polytope:For example, at the origin all three outgoing edges improve the objective, but leaving facet 1 has priority. At bit string $110$, switching bit 2 decreases $x_3$, so the rule switches bit 1, leaving facet 2. At $010$, neither of the earlier-coordinate switches improves $x_3$, so the rule finally leaves facet 5. The upper face reverses the direction in which changes of $x_2$ improve $x_3$, and the remaining three horizontal pivots follow. The objective values in the table are strictly increasing for the stated range of $\epsilon$.