8.2 Linear Programming: Graphical Solutions and Simplex Fundamentals

8.2 Linear Programming: Graphical Solutions and Simplex Fundamentals

Linear programming models can be solved geometrically in two dimensions or algebraically in $n$ dimensions using George Dantzig's Simplex algorithm. Understanding both perspectives is vital for the FE exam: graphical analysis builds intuitive geometric reasoning regarding feasible regions and alternate optima, while Simplex fundamentals explain how optimization software navigates high-dimensional solution spaces.

---

1. The Two-Variable Graphical Solution Method

When a linear program involves only two decision variables ($x_1, x_2$), it can be solved geometrically in the Cartesian plane using four systematic steps:

Step 1: Plot non-negativity axes (First Quadrant: x_1 >= 0, x_2 >= 0)
                     │
                     ▼
Step 2: Plot constraint boundary lines (set inequalities to equalities)
                     │
                     ▼
Step 3: Shade half-planes to identify the Feasible Region (Convex Polygon)
                     │
                     ▼
Step 4: Sweep the Objective Level Curve (Isoprofit / Isocost) to find the optimal vertex

1. Plotting Constraint Boundaries and Feasible Regions

Each linear constraint defines a half-space bounded by a straight line:

$$a_{i1} x_1 + a_{i2} x_2 = b_i$$

To graph the boundary line, find its coordinate intercepts:

• $x_1$-intercept: Set $x_2 = 0 \implies x_1 = \frac{b_i}{a_{i1}}$
• $x_2$-intercept: Set $x_1 = 0 \implies x_2 = \frac{b_i}{a_{i2}}$

Select a test point (such as the origin $(0,0)$ if $b_i \neq 0$) to determine which side of the line satisfies the inequality. The simultaneous intersection of all half-spaces and the non-negativity conditions ($x_1 \ge 0, x_2 \ge 0$) forms the Feasible Region.

2. Convexity of the Feasible Region

A fundamental mathematical property of any LP feasible region is that it forms a convex set. A set $S$ is convex if, for any two points $\mathbf{x}_1, \mathbf{x}_2 \in S$, the straight line segment connecting them lies entirely within $S$:

$$\lambda \mathbf{x}_1 + (1 - \lambda) \mathbf{x}_2 \in S \quad \forall \lambda \in [0, 1]$$

Geometrically, the feasible region of an LP is a convex polygon (in 2D) or a convex polyhedron (in $n$ dimensions). It contains no indentations, holes, or separated regions.

3. Level Curves (Isoprofit and Isocost Lines)

The objective function $Z = c_1 x_1 + c_2 x_2$ represents a family of parallel lines, called level curves (isoprofit lines for maximization, isocost lines for minimization). Rearranging into slope-intercept form:

$$x_2 = -\left(\frac{c_1}{c_2}\right) x_1 + \frac{Z}{c_2}$$

• The slope of every level curve is fixed at $m = -\frac{c_1}{c_2}$.
• Increasing $Z$ (for positive $c_1, c_2$) shifts the line outward away from the origin in the direction of the gradient vector $\nabla Z = [c_1, c_2]^T$.
• In a maximization problem, the optimal solution is the last point of contact between the shifting level curve and the feasible region before exiting.

---

2. The Extreme-Point (Corner-Point) Theorem

The Fundamental Theorem of Linear Programming states:

If a linear program possesses a finite optimal solution, at least one optimal solution must occur at an extreme point (corner point / vertex) of the feasible region.

Because the feasible region is convex and the objective function is strictly linear, the objective function cannot achieve an unconstrained peak in the interior of the region. This theorem reduces the task of searching an infinite continuum of feasible points to evaluating a finite set of extreme points formed by the intersection of binding constraint boundaries.

---

3. Taxonomy of Linear Programming Solution States

Every linear program falls into exactly one of four distinct solution categories:

                               LP SOLUTION STATES
                                       │
         ┌─────────────────────────────┴─────────────────────────────┐
         ▼                                                           ▼
┌─────────────────┐                                         ┌─────────────────┐
│FEASIBLE SOLUTION│                                         │   INFEASIBLE    │
│     EXISTS      │                                         │   Empty region  │
└────────┬────────┘                                         │  No valid point │
         │                                                  └─────────────────┘
         ├─────────────────────────────┐
         ▼                             ▼
┌─────────────────┐           ┌─────────────────┐
│ BOUNDED OPTIMUM │           │    UNBOUNDED    │
│ Finite Z* value │           │  Z -> infinity  │
└────────┬────────┘           │ Missing bounds  │
         │                    └─────────────────┘
         ├─────────────────────────────┐
         ▼                             ▼
┌─────────────────┐           ┌─────────────────┐
│ UNIQUE OPTIMUM  │           │ALTERNATE OPTIMA │
│  Single vertex  │           │ Parallel slopes │
│  c^T x matches  │           │ Entire segment  │
│  unique point   │           │ of equal value  │
└─────────────────┘           └─────────────────┘

---

4. Standard Form of Linear Programs

To solve an LP algebraically via the Simplex algorithm, it must first be converted into Standard Form. Standard form requires:

1. Maximization objective (if minimizing, convert using $\min Z = -\max (-Z)$).
2. All structural constraints expressed as strict equalities ($=$).
3. All right-hand side constants non-negative ($b_i \ge 0$). If any $b_i < 0$, multiply the entire equation by $-1$ and flip the inequality sign before introducing auxiliary variables.
4. All decision variables non-negative ($x_j \ge 0$).

Auxiliary Variables Transformation Table

Big-M Method and Two-Phase Method: Because artificial variables do not exist in the physical system, they must be forced to zero in the final optimal solution. In the Big-M Method, artificial variables are assigned a massive penalty coefficient in the objective function: $-M R_i$ for maximization, or $+M R_i$ for minimization (where $M \gg 0$). If any artificial variable remains positive in the final optimal tableau ($R_i > 0$), the original problem is infeasible.

---

5. Simplex Algorithm Fundamentals & Tableau Operations

A standard LP with $m$ equality constraints and $n$ variables ($n > m$) possesses an underdetermined system of linear equations with infinite solutions. The Simplex algorithm exploits the extreme-point theorem by examining only Basic Feasible Solutions (BFS).

Basic vs. Non-Basic Variables

• Non-Basic Variables (NBV): Set $n - m$ variables equal to zero. These correspond to active constraint boundaries.
• Basic Variables (BV): Solve the remaining $m$ variables from the $m$ linear equations. If all basic variables satisfy $x_{Bi} \ge 0$, the solution is a Basic Feasible Solution (BFS), which corresponds algebraically to an extreme point of the feasible region.

Step-by-Step Simplex Execution

┌────────────────────────────────────────────────────────┐
│ START: Establish Initial BFS (Slack/Artificial Basis)  │
└───────────────────────────┬────────────────────────────┘
                            │
                            ▼
┌────────────────────────────────────────────────────────┐
│ OPTIMALITY TEST: Check reduced costs (c_j - z_j)       │
│ Are all reduced costs <= 0 (in standard max tableau)?  │
└─────────────┬────────────────────────────┬─────────────┘
              │ YES                        │ NO
              ▼                            ▼
    ┌──────────────────┐         ┌─────────────────────────────────┐
    │ STOP: Optimal    │         │ ENTERING VARIABLE SELECTION:    │
    │ Solution Reached │         │ Most positive (or most negative)│
    └──────────────────┘         │ reduced cost enters basis       │
                                 └────────────────┬────────────────┘
                                                  │
                                                  ▼
                                 ┌─────────────────────────────────┐
                                 │ MINIMUM RATIO TEST:             │
                                 │ theta = min {b_i / a_ik}        │
                                 │ for strictly positive a_ik > 0  │
                                 │ Smallest ratio leaves basis     │
                                 └────────────────┬────────────────┘
                                                  │
                                                  ▼
                                 ┌─────────────────────────────────┐
                                 │ GAUSS-JORDAN PIVOT:             │
                                 │ Update tableau via elementary   │
                                 │ row ops to produce new basis    │
                                 └────────────────┬────────────────┘
                                                  │
                                                  └─────── Back to Optimality Test

1. Entering Variable Selection (Optimality Condition): In a maximization tableau where the bottom row represents reduced costs $\bar{c}_j = c_j - z_j$, the non-basic variable with the most positive reduced cost (representing the highest marginal improvement per unit increase) is selected as the entering variable (pivot column $k$). If all $\bar{c}_j \le 0$, the current tableau is optimal.

2. Leaving Variable Selection (Feasibility Condition / Minimum Ratio Test): To determine which basic variable leaves the basis, calculate the ratio of the current RHS values ($b_i$) to the positive coefficients in the pivot column ($a_{ik}$):

   $$\theta = \min_{i \in \{1, \dots, m\} \mid a_{ik} > 0} \left\{ \frac{b_i}{a_{ik}} \right\}$$

   The basic variable corresponding to the row with the minimum strictly positive ratio becomes the leaving variable (pivot row $r$).

   • Why strictly positive? If $a_{ik} \le 0$, increasing the entering variable does not decrease that basic variable, so it never forces it to hit zero.
   • What if all $a_{ik} \le 0$? The problem is unbounded ($Z \to \infty$).

3. Pivot Operation (Gauss-Jordan Elimination): Perform elementary row operations to transform the pivot element $a_{rk}$ into $1$ and all other entries in the pivot column into $0$:

   $$\text{New Pivot Row: } R_r' = \frac{R_r}{a_{rk}}$$ $$\text{Other Rows: } R_i' = R_i - a_{ik} R_r', \quad \forall i \neq r$$

4. Degeneracy and Cycling: If a tie occurs in the minimum ratio test, one or more basic variables will equal zero in the next iteration. This condition is called degeneracy. In degenerate tableaus, a pivot may produce a new basis without improving the objective value ($Z$ remains constant). In rare instances, the algorithm can cycle endlessly between identical bases. Cycling is prevented in practice using Bland's Rule (always choose the candidate variable with the lowest index).

---

6. Step-by-Step Worked Engineering Examples

Worked Example 8.2.1: 2D Graphical Optimization with Alternate Optima Detection

Problem: A precision machining facility manufactures two shafts, S1 ($x_1$) and S2 ($x_2$). The optimization model is:

$$\begin{aligned} \text{Maximize } & Z = 40 x_1 + 60 x_2 \\ \text{Subject to: } & 2 x_1 + 3 x_2 \le 24 \quad \text{(Lathe Capacity, hours)} \\ & 2 x_1 + 1 x_2 \le 16 \quad \text{(Grinder Capacity, hours)} \\ & x_1, x_2 \ge 0 \end{aligned}$$

1. Identify all extreme points of the feasible region.
2. Determine the optimal objective value and characterize the nature of the optimal solution.

Solution:

1. Plot Constraint Boundaries and Identify Intercepts:

   • Constraint 1 (Lathe): $2 x_1 + 3 x_2 = 24$
     • If $x_1 = 0 \implies x_2 = 8 \implies (0, 8)$
     • If $x_2 = 0 \implies x_1 = 12 \implies (12, 0)$
   • Constraint 2 (Grinder): $2 x_1 + x_2 = 16$
     • If $x_1 = 0 \implies x_2 = 16 \implies (0, 16)$
     • If $x_2 = 0 \implies x_1 = 8 \implies (8, 0)$

2. Find Constraint Intersection Point: Solve the system of linear equations:

   $$\begin{cases} 2 x_1 + 3 x_2 = 24 \\ 2 x_1 + x_2 = 16 \end{cases}$$

   Subtracting the second equation from the first gives:

   $$2 x_2 = 8 \implies x_2 = 4$$

   Substitute $x_2 = 4$ into the second equation:

   $$2 x_1 + 4 = 16 \implies 2 x_1 = 12 \implies x_1 = 6$$

   Intersection point: $(6, 4)$.

3. Identify Extreme Points of the Feasible Region: The feasible region is bounded by the axes and the two constraints:

   • Vertex A: Origin $(0, 0)$
   • Vertex B: $(8, 0)$ (limited by Grinder $x_1$-intercept)
   • Vertex C: $(6, 4)$ (intersection of Lathe and Grinder)
   • Vertex D: $(0, 8)$ (limited by Lathe $x_2$-intercept)

4. Evaluate Objective Function $Z = 40 x_1 + 60 x_2$ at Each Vertex:

   • At Vertex A $(0, 0)$: $Z = 40(0) + 60(0) = 0$
   • At Vertex B $(8, 0)$: $Z = 40(8) + 60(0) = 320$
   • At Vertex C $(6, 4)$: $Z = 40(6) + 60(4) = 240 + 240 = 480$
   • At Vertex D $(0, 8)$: $Z = 40(0) + 60(8) = 480$

5. Engineering Interpretation (Alternate Optima): Notice that vertices C and D both yield the identical maximum value of $Z = 480$!

   • Slope of objective function: $m_{\text{obj}} = -\frac{c_1}{c_2} = -\frac{40}{60} = -\frac{2}{3}$
   • Slope of Lathe constraint: $m_{\text{lathe}} = -\frac{a_{11}}{a_{12}} = -\frac{2}{3}$ Because the objective function contour is parallel to the binding Lathe constraint, multiple (alternate) optimal solutions exist. Every point along the line segment connecting $(0, 8)$ and $(6, 4)$ is optimal and yields $Z^* = 480$.

Worked Example 8.2.2: Simplex Tableau Construction and Pivot Execution

Problem: Consider the following initial Simplex tableau for a profit maximization problem with decision variables $x_1, x_2$ and slack variables $s_1, s_2, s_3$:

(Tableau convention: objective row expresses $Z - 5 x_1 - 4 x_2 = 0$, where negative entries indicate improvement opportunities). Determine the entering variable, leaving variable, pivot element, and update the tableau for one complete iteration.

Solution:

1. Identify Entering Variable: The most negative coefficient in the objective row is $-5$ under column $x_1$. Therefore, $x_1$ is the entering variable (pivot column is column 1).

2. Execute Minimum Ratio Test: Calculate $\theta_i = b_i / a_{i1}$ for entries where $a_{i1} > 0$:

   • Row 1 ($s_1$): $16 / 1 = 16$
   • Row 2 ($s_2$): $18 / 3 = 6$
   • Row 3 ($s_3$): $a_{31} = 0$ (division by zero is strictly excluded) The minimum ratio is $\min(16, 6) = 6$ in Row 2. Therefore, $s_2$ is the leaving variable, and the pivot element is $a_{21} = 3$.

3. Perform Gauss-Jordan Row Operations:

   • Normalize Pivot Row 2: Divide Row 2 by $3$: $$R_2' = \left[ 1, \; \frac{1}{3}, \; 0, \; \frac{1}{3}, \; 0, \; 6 \right]$$
   • Eliminate $x_1$ from Row 1: $R_1' = R_1 - 1 \cdot R_2'$: $$R_1' = [1, 2, 1, 0, 0, 16] - [1, 1/3, 0, 1/3, 0, 6] = [0, 5/3, 1, -1/3, 0, 10]$$
   • Row 3 remains unchanged since $a_{31} = 0$: $$R_3' = [0, 1, 0, 0, 1, 6]$$
   • Eliminate $x_1$ from Objective Row: $R_Z' = R_Z - (-5) \cdot R_2' = R_Z + 5 \cdot R_2'$: $$R_Z' = [-5, -4, 0, 0, 0, 0] + [5, 5/3, 0, 5/3, 0, 30] = [0, -7/3, 0, 5/3, 0, 30]$$

4. Updated Simplex Tableau:

Current BFS: $x_1 = 6, s_1 = 10, s_3 = 6$, with non-basic $x_2 = 0, s_2 = 0$, producing $Z = 30$. Because $x_2$ still has a negative entry ($-7/3$), further improvement is possible in the next pivot.

---

7. NCEES Reference Handbook Tips & Realistic Exam Traps

• Slope Comparison for Alternate Optima: When an exam problem asks whether alternate optima exist, calculate the slope of the objective line $-c_1/c_2$ and the slopes of all binding constraint boundaries $-a_{i1}/a_{i2}$. If the slopes match and the constraint is active at the optimal vertex, alternate optima exist.
• Minimum Ratio Divisor Rule: Never divide by zero or a negative coefficient in the minimum ratio test! Ratios with $a_{ik} \le 0$ must be completely ignored. Dividing by a negative number violates the feasibility boundary.
• Negative RHS Values in Standard Form: If an initial constraint is written as $-2 x_1 + 5 x_2 \le -10$, you must multiply by $-1$ to make the RHS positive: $2 x_1 - 5 x_2 \ge 10$. Notice that this reverses the inequality to $\ge$, which requires a surplus variable and an artificial variable ($2 x_1 - 5 x_2 - e_1 + R_1 = 10$).
• Distinguishing Slack from Artificial Variables: Slack variables represent real physical unused capacity and have an objective coefficient of $0$. Artificial variables have no physical existence and exist solely to provide an initial identity matrix in the Simplex tableau.