Signal-flow graphs and Mason's gain formula
Draw signals as nodes and blocks as branches, then read the transfer function of a multi-loop system directly from its forward paths and loops.
THE SHORT ANSWER
A signal-flow graph draws a linear system's equations as nodes for signals and directed branches whose gains are transfer functions; each node equals the sum of its incoming branch signals. Mason's gain formula T = (1/Δ)ΣPₖΔₖ reads the transfer function from the forward-path gains Pₖ and the graph determinant Δ, built from the loop gains and the products of non-touching loops. Each cofactor Δₖ keeps only the loops that do not touch path k.
Convert a block diagram into a signal-flow graph, list its forward paths, loops and non-touching loops, compute a transfer function with Mason's gain formula, and confirm it by block reduction or direct algebra.
The block-algebra lecture combined blocks in series, in parallel and in a feedback loop, one loop at a time. Real diagrams often have several loops, and a feedback path may leave from inside one loop and return inside another. This lecture redraws the same equations as a graph and reads the transfer function from its paths and loops.
Review the previous lesson: Block diagrams: follow the signal, then reduce the blocksFirst, picture it
Block-diagram reduction works one loop at a time: combine blocks in series, close an inner loop, and repeat. That is easy when the loops are nested like the layers of an onion. When loops interlock, so that a feedback path leaves from inside one loop and returns inside another, no loop can be closed first. You have to move a summing junction or a pickoff point across a block, add a compensating block and redraw, and every redrawing is a chance to lose a sign. A signal-flow graph carries the same equations with less ink: every signal is a dot and every block is an arrow. Mason's gain formula then reads the transfer function off the graph in one pass, from two lists: the forward paths from input to output, and the loops. Nothing is redrawn.
Build the idea step by step
A signal-flow graph has two kinds of object. A node is one signal, such as the error E or the output Y. A branch is a directed line from one node to another, and its gain is the transfer function that multiplies the signal it carries. The value of a node is the sum of the signals on all branches that enter it; branches that leave a node do not change its value. An input node, or source, has only outgoing branches; the output node is the node whose signal you want. A path is a chain of branches followed in their direction that passes no node twice. A forward path runs from the input node to the output node, and its gain Pk is the product of its branch gains. A loop is a closed path that returns to its starting node without passing any other node twice; its loop gain Li is the product of its branch gains, signs included. Two loops touch when they share at least one node, and loops that share no node are non-touching. A loop touches a forward path in the same sense: it shares a node with the path.
Read the equations
Node rule: the signal at node j is the sum of the signals arriving on its incoming branches, each multiplied by its branch gain. Outgoing branches do not appear in a node's own equation.
Converting a summing junction: each incoming signal becomes a branch into the node E, and a minus sign moves into the branch gain. Here one feedback signal comes from the internal signal X₂ and one from the output Y.
A forward-path gain and a loop gain are both plain products of branch gains. Signs are part of the gains, so a negative-feedback loop has a negative loop gain.
The graph determinant. The first sum runs over all loops, the second over every pair of non-touching loops, the third over every triple of mutually non-touching loops, and so on, with alternating signs.
Mason's gain formula. Δk is Δ recomputed with only the loops that do not touch forward path k; if every loop touches path k, Δk = 1.
With one forward path and one loop touching it, Mason's formula reduces to the familiar negative-feedback formula from the block-algebra lecture.
When the only two loops do not touch, Δ factors: each loop can be closed on its own. The product term L₁L₂ is what makes the factorisation work.
The same graph as linear equations: Δ is the determinant of the node equations. Solving these equations directly is the independent check used in this lecture.
What the symbols mean
- node signals (Laplace transforms)
- gain of the branch from node i to node j
- gain of the k-th forward path from the input node to the output node
- loop gain of the i-th loop, sign included
- graph determinant, built from all loops
- Δ computed from only the loops that do not touch forward path k
- transfer function from the input node to the output node
- input and output signals
- internal signals of the worked example
- forward-path and feedback-path block transfer functions
- matrix of branch gains between nodes, and gains from the input to each node
- Laplace variable, in inverse seconds
The formula holds for linear graphs with zero initial conditions. T is the transfer function from an input node, which has only outgoing branches, to whichever node you choose as the output. Every loop gain carries its own sign, and each sum in Δ runs over sets of loops in which every pair is non-touching, each set counted once.
Each signal of the block diagram becomes a node, each block a branch, and each minus sign a negative branch gain. Orange marks G₄, which creates the second forward path R → E → X₂ → Y; green marks the feedback branches. The loops −G₁H₁ and −G₃H₂ share no node, and neither do −G₃H₂ and −G₁G₄H₃.
Two forward paths, four loops, two non-touching pairs
Write one equation per summing junction or block of the diagram in the figure. The first summing junction subtracts an internal signal X₂ through H₁ and the output Y through H₃; a second one subtracts X₄ through H₂; the output adds a shortcut G₄ from X₂.
Turn each signal into a node and each term into a branch. The minus signs move into the gains −H₁, −H₂ and −H₃, and the signal passes from R to E and from X₄ to Y through branches of gain 1.
List the forward paths from R to Y, following branches only in their direction and never passing a node twice. There are two: the main chain and the shortcut through G₄.
List the loops with their node sets. Start a loop at any of its nodes and go around once; the same loop found again from another starting node is not a new loop. There are four.
Compare node sets pair by pair. L₂ shares no node with L₁ and none with L₄; every other pair shares E or X₂. No three loops are mutually non-touching, because L₁ and L₄ touch each other.
Assemble Δ. Each loop gain is negative, so every term ends up positive.
Find the cofactors. P₁ passes through all six nodes, so every loop touches it and Δ₁ = 1. P₂ misses X₃ and X₄, so L₂ does not touch it and stays in Δ₂.
Now put in numbers: G₁ = 1/(s + 1), H₁ = 2, G₂ = 5, G₃ = 1/s, H₂ = 4, G₄ = 3 and H₃ = 1. The loop gains are L₁ = −2/(s + 1), L₂ = −4/s, L₃ = −5/(s(s + 1)) and L₄ = −3/(s + 1). Multiply the numerator and Δ by s(s + 1).
Cross-check by block reduction. These loops happen to be nested, so they can be closed from the inside out. The loop around G₁ closes to 1/(s + 3) and the loop around G₃ to 1/(s + 4). The middle branch 5/(s + 4) is in parallel with G₄ = 3, the result is in series with 1/(s + 3), and closing the outer loop through H₃ = 1 gives the same T(s). The two methods agree: the poles are −5 ± j2 and the DC gain is 17/29 ≈ 0.586.
Reference table
| Block-diagram element | Signal-flow graph | Remark |
|---|---|---|
| Signal (input, block output, summing-junction output) | Node | its value is the sum of the signals on its incoming branches |
| Block G | Branch of gain G from the block's input node to its output node | gains multiply along a path |
| Summing junction | The node of its output signal, with one incoming branch per entering signal | a node can only add |
| Minus sign at a summing junction | Negative branch gain, such as −H | the sign becomes part of the loop gain |
| Pickoff (take-off) point | One more outgoing branch from the same node | not a new node |
| Signal passed on unchanged | Branch of gain 1 | — |
| Two summing junctions with nothing between them | One node | addition does not depend on order |
A node's outgoing branches do not change its value. Converting back to a block diagram applies the same rules in reverse and keeps every transfer function.
More detail and model assumptions
From a block diagram to a graph, and back
Give every signal a node: the input, the output of each summing junction, the output of each block, and the output of the system. A pickoff (take-off) point is not a new node; it is the node itself sending out one more branch. Each block becomes a branch from its input node to its output node, carrying the block's transfer function. A summing junction becomes the node of its output signal, and every signal entering it becomes a branch into that node. A signal that enters with a minus sign gets a branch gain with a minus sign: −H instead of H. Where a signal passes on unchanged, for example from R into the first summing junction, draw a branch of gain 1.
Two details matter. A node adds up everything that comes in but cannot subtract, so the minus sign must live in a branch gain. And a signal that feeds both a block and a summing junction, as X₂ does in the worked example, is simply one node with several outgoing branches. Two summing junctions in a row with nothing between them can be merged into one node, because addition does not care about order.
The way back reads the same rules in reverse. Every node with more than one incoming branch becomes a summing junction, every branch whose gain is not 1 becomes a block, and a node with several outgoing branches becomes a pickoff point. A branch gain −H can be drawn as a block H that enters its summing junction with a minus sign. The diagram you get back may look different from the one you started with, but it encodes the same equations and therefore has the same transfer functions.
Why non-touching loops multiply, and what Δ really is
Take two first-order sections in cascade, each with its own local feedback and nothing else. Their loops share no node, so each section can be closed on its own and the transfer function is the path gain divided by (1 − L₁)(1 − L₂). Expand that denominator: (1 − L₁)(1 − L₂) = 1 − L₁ − L₂ + L₁L₂. The product term is exactly the second line of Mason's Δ. Loops that touch cannot be separated this way, which is why their product never appears.
In general Δ is a determinant. Write the node equations as x = Ax + bR, where the entry of A in row i and column j is the gain of the branch from node j to node i. Then (I − A)x = bR and Δ = det(I − A). Expanding this determinant produces one term for every set of mutually non-touching loops, with sign (−1)ᵐ for a set of m loops: that is Mason's Δ. Cramer's rule then gives the output as a ratio of two determinants, and the numerator determinant is what Σ PkΔk adds up. Mason's formula is Cramer's rule organised by paths and loops, which is why it always agrees with solving the node equations directly.
Because Δ depends only on the loops, every transfer function taken from the same graph has the same denominator Δ before any cancellation, whichever input and output you choose. That is why a disturbance input sees the same closed-loop poles as the reference: the poles belong to the loops, not to the signal you apply.
Typical mistakes, and three checks
Sign of feedback: a loop gain is the product of its branch gains, and the minus sign of a negative-feedback summing junction is one of them. In the worked example L₃ = −G₁G₂G₃H₃; with the sign lost, Δ would contain −G₁G₂G₃H₃ and describe positive feedback. Counting a loop twice: E → X₂ → E and X₂ → E → X₂ are the same loop, so list each loop once, for example by its set of nodes. Forgetting what Δk keeps: writing Δk = 1 out of habit is right only when every loop touches path k. In the worked example it would drop G₁G₄G₃H₂ from the numerator.
Two more rules are easy to break. A forward path or a loop must not pass the same node twice. And a product term in Δ needs loops that are pairwise non-touching: in a triple, every one of the three pairs must be separate.
Three quick checks catch most errors. Set one branch gain to zero and confirm that the result reduces to the simpler graph: with G₄ = 0 the worked example has a single path and T = 5/(s² + 7s + 17). Compare the denominator with the closed-loop poles you expect; every transfer function of the graph must share it. And for a numeric example, solve the node equations directly at one value of s and compare.
When to use which method
Mason's formula is not always the faster route. When loops are nested, closing the inner loops first is quick and shows useful intermediate transfer functions, as the worked example's cross-check does. The formula pays off when loops interlock, when there are several forward paths, and when you need several transfer functions from one diagram, because Δ is computed only once. For large systems, software solves the node equations or a state-space model instead; in hand calculation the formula's real value is a checklist that does not let a term go missing.
Connect the example to the lab
The block-algebra bench starts at G₁ = 2, G₂ = 3 and H = 1. Before switching its connection, draw each one as a graph. Series is one forward path of gain 6 and no loop, so T = 6. Parallel is two forward paths and no loop, so T = G₁ + G₂ = 5. Negative feedback is one forward path G₁G₂ and one loop of gain −G₁G₂H = −6, so T = 6/7 ≈ 0.857. Then predict with Mason's formula what H = 2 gives (L = −12, T = 6/13 ≈ 0.462) before you read it on the bench.
Check one-loop and two-path graphs on the block-algebra benchMake a prediction before moving a control.
- PREDICTOn the block-algebra bench choose negative feedback with G₁ = 2, G₂ = 3 and H = 1. Draw its graph: one forward path G₁G₂ and one loop −G₁G₂H. Predict y/r with Mason's formula, and predict it again for H = 0.5.
- CHANGE ONE THINGChange only H, from 1 to 0.5.
- OBSERVERead y/r: about 0.857 (6/7) at H = 1 and 1.5 at H = 0.5. The loop gain went from −6 to −3; the forward path did not change.
- EXPLAINUse Δ = 1 − L to explain why halving H raises the output, and why P₁ = G₁G₂ is unaffected. Then switch to the parallel connection and explain T = G₁ + G₂ as two forward paths with Δ = 1.
Three mistakes cause most wrong answers. Losing the minus sign of a negative-feedback summing junction turns a loop gain −GH into +GH. Listing the same loop twice, once from each starting node, counts it twice in Δ. And Δk is not always 1: it keeps every loop that does not touch path k and drops only the loops that do.
Check your understanding+
A graph has exactly two loops, with gains −2 and −3, and they share no node. Its only forward path has gain 12 and touches both loops. What is the transfer function, and what would you get if you forgot the product term in Δ?
Δ = 1 + 2 + 3 + (−2)(−3) = 12, which equals (1 + 2)(1 + 3), so T = 12/12 = 1. Without the product term Δ would be 6 and T = 2, twice the true value.Solve these before opening the solutions
Work each problem on paper first. Open the answer to check yours, then the full solution if you need the steps, and mark how it went.
The graph below has input R, output x₅ and constant branch gains. (a) List the forward paths and the loops with their gains and node sets. (b) Which pairs of loops touch? (c) Find x₅/R with Mason's formula and check it by solving the node equations.
Show answer
AnswerOne forward path, P₁ = 24. Loops L₁ = −2 {x₁, x₂}, L₂ = −1.5 {x₂, x₃, x₄}, L₃ = −1 {x₄, x₅}; only L₁ and L₃ are non-touching. Δ = 7.5 and x₅/R = 3.2.
Show full solution
Solution
There is only one forward path, and it passes through every node.
Three loops, each with its node set.
(b) L₁ and L₂ share x₂, and L₂ and L₃ share x₄; L₁ and L₃ share nothing. L₁ and L₂ overlap without either containing the other, the situation in which block reduction needs a moved junction. Mason's formula only needs the node sets.
(c) Only L₁L₃ enters the second sum of Δ. The single path touches every loop, so Δ₁ = 1.
Check with the node equations. From x₅ = 4x₄ and x₄ = x₃ − 0.25x₅ we get x₄ = x₃/2. With x₃ = 3x₂ and x₂ = 2x₁ − 0.5x₄ we get 1.75x₂ = 2x₁, and x₁ = R − x₂ then gives x₂ = 2R/3.75. So x₅ = 6x₂ = 3.2R.
A position servo: the error E = R − Y drives an amplifier of gain 8. A tachometer measures the motor speed W and its signal 4W is subtracted at a second summing junction in front of the motor, whose speed follows W = U/(s + 2); the position is Y = W/s. (a) Draw the signal-flow graph. (b) Find Y/R with Mason's formula. (c) Do the two loops touch, and what does that mean for Δ?
Show answer
AnswerThe loops L₁ = −4/(s + 2) and L₂ = −8/(s(s + 2)) share U and W, so Δ has no product term. Y/R = 8/(s² + 6s + 8) = 8/((s + 2)(s + 4)), with poles at −2 and −4.
Show full solution
Solution
(a) Nodes R, E, U, W and Y. The tachometer branch leaves W and enters U with a negative gain; the outer feedback leaves Y and enters E with gain −1.
(b) There is one forward path, and it passes through every node.
(c) Two loops, and both pass through U and W, so they touch: Δ contains no L₁L₂ term.
Divide the path term by Δ.
Check by block reduction: the tachometer loop closes to 1/(s + 6), the forward path becomes 8/(s(s + 6)), and the unity outer loop gives the same result. Without the tachometer the denominator would be s² + 2s + 8: the inner loop adds damping.
A plant G(s) = 1/((s + 1)(s + 4)) is under proportional feedback C = 6 with a unity sensor. A static feedforward F = 4 adds the reference directly to the plant input, and a disturbance D also enters at the plant input. (a) Draw one graph with both inputs R and D. (b) Find Y/R and Y/D with Mason's formula. (c) Why do the two share a denominator, and what does F do to the steady output for a unit step reference?
Show answer
AnswerY/R = 10/(s² + 5s + 10) and Y/D = 1/(s² + 5s + 10). Both have Δ = 1 + CG because the graph has a single loop. F makes the steady output 1, whereas proportional control alone gives 0.6.
Show full solution
Solution
(a) Nodes R, D, E, U and Y. Three branches enter U: from E through C, from R through F, and from D with gain 1.
There is one loop, through E, U and Y.
(b) From R there are two forward paths, through the controller and through the feedforward. Both pass through U and Y, so both touch the loop and Δ₁ = Δ₂ = 1.
From D there is one path, D → U → Y, and it also touches the loop.
(c) Δ depends only on the loops, so every input of this graph sees the same closed-loop poles, −2.5 ± j1.94. At s = 0, G(0) = 1/4: proportional control alone gives 6G(0)/(1 + 6G(0)) = 0.6, and the feedforward F = 1/G(0) = 4 raises the steady output to 1. F changes only the numerator of Y/R; the disturbance response is unchanged.
A student analysed the graph below, with G₁ = 2/(s + 1), H₁ = 1, G₂ = 4/s and G₃ = 0.5, and wrote: “Loops −G₁H₁, −H₁G₁ and +G₁G₂, so Δ = 1 + 2G₁H₁ − G₁G₂. Forward paths G₁G₂ and G₃, so T = (G₁G₂ + G₃)/Δ.” Find the three mistakes and give the correct T(s).
Show answer
AnswerThe loop −G₁H₁ was counted twice, the outer loop gain is −G₁G₂ because its branch −1 carries the sign, and Δ₂ = 1 + G₁H₁ because the path R → Y does not touch the inner loop. T(s) = (0.5s² + 1.5s + 8)/(s² + 3s + 8) = 0.5 + 4/(s² + 3s + 8).
Show full solution
Solution
Loops. E → X → E and X → E → X are the same loop, so it is counted once. The outer loop E → X → Y → E includes the branch of gain −1, so its gain is −G₁G₂, not +G₁G₂.
Cofactors. P₁ = G₁G₂ passes through E, X and Y and touches both loops. P₂ = G₃ visits only R and Y: it touches L₂ at Y but not L₁, so L₁ stays in Δ₂.
The correct transfer function.
Put in the numbers and multiply the numerator and the denominator by s(s + 1).
Sanity check at s = 0. The integrator in G₂ can hold a constant output only when its input X is zero, and then E = R − Y must be zero as well, so T(0) must be 1. The corrected result gives 8/8 = 1. The student's expression gives −1 at s = 0, a clear sign that something was wrong.
Exam-level. Find Y/R = x₆/R for the graph below in terms of G₁ … G₆ and H₁ … H₄. Write Δ both expanded and in a compact factored form.
Show answer
AnswerT = [G₁G₂G₃G₄G₅ + G₁G₅G₆(1 + G₃H₂)]/Δ with Δ = (1 + G₁H₁)(1 + G₃H₂)(1 + G₅H₃) + G₁G₅H₄[G₂G₃G₄ + G₆(1 + G₃H₂)].
Show full solution
Solution
Forward paths: the main chain, and the shortcut through G₆ that skips x₃ and x₄.
Loops: three local loops and two loops that close through H₄, one along each forward path.
Non-touching sets. L₁, L₂ and L₃ share no node with one another, which gives three pairs and one triple. L₅ misses x₃ and x₄, so it also pairs with L₂. L₄ passes through every node and touches every loop.
Assemble Δ. The triple enters as −L₁L₂L₃, which is +G₁H₁G₃H₂G₅H₃ because each Lᵢ is negative.
Cofactors: P₁ touches every loop; P₂ misses x₃ and x₄, so only L₂ survives in Δ₂.
Group the eleven terms of Δ: eight of them form (1 + G₁H₁)(1 + G₃H₂)(1 + G₅H₃), and the other three share the factor G₁G₅H₄. Block reduction gives the same compact form: close the three local loops, combine the middle section in parallel with G₆, then close the outer loop through H₄.
Mark each exercise after you check your answer. Marks stay in this browser.
Terms this lecture teaches
All terms in the glossary →Frequently asked questions
What is the difference between a block diagram and a signal-flow graph?
Both encode the same linear equations. A block diagram draws blocks, summing junctions and pickoff points; a signal-flow graph draws each signal as a node and each block as a branch, with every minus sign moved into a branch gain. Converting in either direction keeps all the transfer functions.
What are non-touching loops in Mason's gain formula?
Two loops are non-touching when they share no node. Only sets of mutually non-touching loops contribute product terms to Δ: pairs with a plus sign, triples with a minus sign, and so on. Loops that share even one node never appear multiplied together.
When is the path cofactor Δₖ not equal to 1?
Δₖ keeps every loop that does not touch forward path k, so it equals 1 only when every loop touches that path. A short path that skips part of the graph often leaves a loop untouched; dropping that loop from Δₖ out of habit removes a term from the numerator.
Why do all transfer functions of one graph share the same denominator?
Δ is built from the loops alone and does not depend on which input or output you pick, so before any cancellation every transfer function of the graph has Δ as its denominator. That is why a disturbance sees the same closed-loop poles as the reference.
Should I use Mason's formula or block-diagram reduction?
When loops are nested, closing the inner loops first is quick and shows useful intermediate results. Mason's formula pays off when loops interlock, when there are several forward paths, or when you need several transfer functions from one diagram, because Δ is computed once. Either way, check the answer by solving the node equations at one value of s.
Why the next lesson follows
With the closed-loop transfer function in hand, whether from reduction or from Mason's formula, we can ask how quickly its output rises, overshoots and settles.
Further reading
Finished the reading and the exercises? Mark the lecture as done to tick it on the syllabus. Progress stays in this browser.
EIGENROOM · https://eigenroom.com/theory/signal-flow-graphs · Updated 2026-11-15