920 PROCEEDINGS OF THE IRE July

CONCLUSION terms together with a few rational operations. These

charts, after they are prepared, will be very helpful for
The elliptic function transformation (1). is used here design purposes. For instance, if a required attenuation
for the purpose of locating zeros and poles of a low-pass beyond twice the cut-off frequency must be greater than
filter network function. Charts of the type shown in 13 db, Fig. 10 indicates that a filter function with the
Figs. 7 to 12 may be prepared for any range of applica- charge arrangement of Fig. 3(b) and values of a and c
tion whenever desired. The compactness of the expres- of 0.810 and 0.673 respectively will satisfy the require-
sions that give the tolerance and other characteristic ment. The locations of all poles and zeros of this filter
quantities makes the preparation of these charts which are determined in the z-plane. The locations of zeros and
represent a whole group of network functions with poles in the s-plane may be found by applying the in-
many singularities a matter of evaluating only a few verse transformation.

Feedback Theory -Further Properties of Signal

Flow Graphs*
SAMUEL J. MASONt

Summary-A way to enhance a linear system analysis problem should be interpretable

Writing gain at a glance. as a search for all possible combinations of something
Dr. Tustin extended
Proof appended. or other, and that the solution should take the form of
Examples illustrative a sum of products of the somethings, whatever they are,
Pray not frustrative. divided by another such sum of products. Hence, in-
stead of undertaking a sequence of operations, we can
BACKGROUND find the solution by looking for certain combinations
T HERE ARE many different paths to the solution of things. The method will be especially useful if these
of a set of linear equations. The formal method combinations have a simple interpretation in the con-
involves inversion of a matrix. We know, how- text of the problem.
ever, that there are many different ways of inverting a
matrix: determinantal expansion in minors, systematic
reduction of a matrix to diagonal form, partitioning
into submatrices, and so forth, each of which has its aAb
particular interpretation as a sequence of algebraic --2
manipulations within the original equations. A deter- 3tc
minantal expansion of special interest is
D = >, alia2ja3k an ... (1)
Fig. 1-An electrical network graph.
where amp is the element in the mth row and pth
column of a) determinant having n rows, and the sum-
mation is taken over all possible permutations of the As a concrete illustration of the idea, consider the
column subscripts. (The sign of each term is positive electrical network graph shown in Fig. 1. For simplicity,
or negative in accord with an even or odd number of let the branch admittances be denoted by letters a, b,
successive adjacent column-subscript interchanges re- c, d, and e. This particular graph has three independent
quired to produce a given permutation.) Since the solu- node-pairs. First locate all possible sets of three branches
tion of a set of linear equations involves ratios of deter- which do not contain closed loops and write the sum of
minantal quantities, (1) suggests the general idea that their branch admittance products as the denominator
of (2).
*
Original manuscript received by the IRE, August 16, 1955;
revised manuscript received, February 27, 1956. This work was sup- ab + ac + bc + bd
ported in part by the Army (Signal Corps), the Air Force (Office of
Scientific Research, Air Research and Development Command), abd + abe + acd + ace + ade + bcd + bce + bde
and the Navy (Office of Naval Research).
t Dept. of Elect. Engrg. and Res. Lab. of Electronics, M.I.T.,
Cambridge, Mass. Now locate all sets of two branches which do not form
1956 Mason: Further Properties of Signal Flow Graphs 921
closed loops and which also do not contain any paths from The purposes of this paper are: to extend the method
node 1 to ground or from node 2 to ground. Write the sum to a general form applicable to any flow graph; to
of their branch admittance products as the numerator of present a proof of the general result; and to illustrate
expression (2). The result is the transfer impedance the usefulness of such flow graph techniques by applica-
between nodes 1 and 2, that is, the voltage at node 2 tion to practical linear analysis problems. The proof
when a unit current is injected at node 1. Any imped- will be given last. It is tempting to add, at this point,
ance of a branch network can be found by this process.' that a better understanding of linear analysis is a great
So much for electrical network graphs. Our main aid in problems of nonlinear analysis and linear or non-
concern in this paper is with signal flow graphs,2 whose linear design.
branches are directed. Tustin3 has suggested that the
feedback factor for a flow graph of the form shown in A BRIEF STATEMENT OF SOME ELEMENTARY
Fig. 2 can be formulated by combining the feedback loop PROPERTIES OF LINEAR SIGNAL
FLOW GRAPHS

K)j A signal flow graph is a network of directed branches

a b c d e f g
* 9- t a, t 0 t PI t PI t t PI e

which connect at nodes. Branch jk originates at node j

h and terminates upon node k, the direction from j to k
Fig. 2-The flow graph of an automatic control system. being indicated by an arrowhead on the branch. Each
branch jk has associated with it a quantity called the
gains in a certain way. The three loop gains are branch gain gjk and each node j has an associated quan-
tity called the node signal xi. The various node signals
T,= bch (3a) are related by the associated equations
T2= cdi (3b) A X jg1k = Xk, k = 1, 2, 3, (7)
T3 = (3c) i

and the forward path gain is The graph shown in Fig. 3, for example, has equations
Go= abcdefg. (4) ax, + dx3 = X2 (8a)
The gain of the complete system is found to be bx2 + fx4 = X3 (8b)
eX2 + Cx3 = X4 (8c)
[1- (Ti + T2)](1 T3) - (5) gX3 + hx4 = X5. (8d)
and expansion of the denominator yields We shall need certain definitions. A source is a node
having only outgoing branches (node 1 in Fig. 3). A
Go (6) sink is a node having only incoming branches. A path
.~~~~
1-(Ti + T2+ T3)+ (T1T3+ T2T3) is any continuous succession of branches traversed in
the indicated branch directions. A forward path is a
Tustin recognized the denominator as unity plus the path from source to sink along which no node is encoun-
sum of all possible products of loop gains taken one at tered more than once (abch, aeh, aefg, abg, in Fig. 3).
a time (T1+T2+Ts), two at a time, (T1T3+T2T3),
three at a time, and so forth, excluding products of
loops that touch or partially coincide. The products d f
T1T2 and T1T2T3 are properly and accordingly missing
in this particular example. The algebraic sign alter-
nates, as shown, with each succeeding group of products. e 9
Tustin did not take up the general case but gave a Fig. 3-A simple signal flow graph.
hint that a graph having several different forward paths
could be handled by considering each path separately
and superposing the effects. Detailed examination of the A feedback loop is a path that forms a closed cycle along
general problem shows, in fact, that the form of (6) which each node is encountered once per cycle (bd, cf,
must be modified to include possible feedback factors in def, but not bcfd, in Fig. 3). A path gain is the product of
the numerator. Otherwise (6) applies only to those the branch gains along that path. The loop gain of a
graphs in which each loop touches all forward paths. feedback loop is the product of the gains of the branches
forming that loop. The gain of a flow graph is the signal
1 Y. H. Ku, "Resume of Maxewll's and Kirchoff's rules for net- appearing at the sink per unit signal applied at the
work analysis," J. Frank. Inst., vol. 253, pp. 211-224; March, source. Only one source and one sink need be considered,
1952.
2 S. J. Mason, "Feedback theory-some properties of signal flow since sources are superposable and sinks are independ-
graphs," PROC. IRE, vol. 41, pp. 1144-1156; September, 1953.
3 A. Tustin, "Direct Current Machines for Control Systems," ent of each other.
The Macmillan Company, New York, pp. 45-46, 1952. Additional terminology will be introduced as needed.
922 PROCEEDINGS OF TIHE IRE Jlty
GENERAL FORMULATION OF FLOW GRAPH GAIN f
To begin with an example, consider the graph shown
(a) .~I -G a-
-
abc+d(I
- -
- bf)
I ce hf cg -dgf e + aecg
in Fig. 4. This graph exhibits three feedback loops,
whose gains are e
T1-=h (9a) ( b) -- - - -- -*----a T = ae

T2= fg (9b)
T3 de (9c)
(c) 4-
- -- --A---T2=bf
,b
b

and two forward paths, whose gains are

G1 = ab (lOa) g
G2= ceb. (lOb) (d) -k--C- -
* T,= c g A =I-T1-T T2 -T 3 -T4 +rTI 3

e f g
'h (e) - _ T4=dgfe
d
e 9
(f ) *--
-10- -- - rT1 T3 = aecg
, a

(g k o G,=abc,A =I

a b c',

b
Fig. 4-A flow graph with three feedback loops. -f
(h ) 0
I
.1

Ns b
. .- -oG =d
,A =I bf

To find the graph gain, first locate all possible sets of d

nontouching loops and write the algebraic sum of their Fig. 5-Identification of paths and loop sets.
gain products as the denominator of (11).
G1(1 - T1 - T2) +G2( 1- T1) GkAk

G= - (11) k
(12a)
1 - T -T2- T3+ T,T3
Each term of the denominator is the gain product of wherein
a set of nontouching loops. The algebraic sign of the
term is plus (or minus) for an even (or odd) number of Gk = gain of the kth forward path (12b)
loops in the set. The graph of Fig. 4 has no sets of three A = 1 - Pml + L Pm2 - L Pm3 + (12c)
or more nontouching loops. Taking the loops two at a m m m

time we find only one permissible set, T1T3. When the PmrT gain product of the mth possible combina-
=

loops are taken one at a time the question of touching tion of r nontouching loops (12d)
does not arise, so that each loop in the graph is itself Ak= the value of A for that part of the graph
an admissible "set." For completeness of form we may (12e)
not touching the kth forward path.
also consider the set of loops taken "none at a time" and,
by analogy with the zeroth power of a number, interpret The form of (12a) suggests that we call A the determi-
its gain product as the unity term in the denominator nant of the graph, and call Ak the cofactor of forward
of (11). The numerator contains the sum of all forward path k.
path gains, each multiplied by a factor. The factor for A subsidiary result of some interest has to do with
a given forward path is made up of all possible sets of graphs whose feedback loops form nontouching sub-
loops which do not touch each other and which also do graphs. To find the loop subgraph of any flow graph,
not touch that forward path. The first forward path simply remove all of those branches not lying in feed-
(G1 = ab) touches the third loop, and T3 is therefore back loops, leaving all of the feedback loops, and noth-
absent from the first numerator factor. Since the second ing but the feedback loops. In general, the loop sub-
path (G2=ceb) touches both T2 and T3, only T1 enters graph may have a number of nontouching parts.
the second factor. The useful fact is that the determinant of a complete
The general expression for graph gain may be writ- flow graph is equal to the product of the determinants of
ten as each of the nontouching parts in its loop subgraph.
1956 Mason: Further Properties of Signal Flow Graphs 923
e
d(l-be) +abc flow graph are in cause-and-effect form, each variable
(a) expressed explicitly in terms of others, and since physi-
D

y -I-b
be
d cal problems are often very conveniently formulated in
just this form, the study of flow graphs assumes practi-
cal slgnificance.
a + be
(b) *
I-ad-be-cf- bcd-afe
d

f
ilI Z5
(c) c +ab
I-od- be -oabf -cd -cf
c
(a)

(d). i(I-g-h-cd+gh)+ae(l-h)+ bf (i-g)+adf +bce eI i2 e2 i3 e3

I -g-h-cd+gh I 2 4 : I * e3
-zI -Y2 -Z 3 -Y4

( b)
I a b g ( -hi -jc-hbcd +hijc) +a ie (1-jc)+ a bcde
(e ) f 1- fg -hi-jc -faie - hbcd -fabcde i e
2 e2 3 e3
+ fghi +fg jc+hijc +foiejc+fghbcd
I e d - f ghijc

Fig. 6-Sample flow graphs and their gain expressions.

ILLUSTRATIVE EXAMPLES OF GAIN EVALUATION Fig. 7-The transfer impedance of a ladder.

BY INSPECTION OF PATHS AND LooP SETS Consider the ladder network shown in Fig. 7(a). The
Eq. (12) is formidable at first sight but the idea is problem is to find the transfer impedance e3/il. One pos-
simple. More examples will help illustrate its simplicity. sible formulation of the problem is indicated by the flow
Fig. 5 (on the previous page) shows the first of these dis- graph Fig. 7(b). The associated equations state that
played in minute detail: (a) the graph to be solved; (b)- ei =zl(il -i2), i2 =y2(el-e2), and so forth. By inspection
(f) the loop sets contributing to A; (g) and (h) the for- of the graph,

e3 Zly2Z3y4Zs
(13a)
ii 1 + Z1y2 + Y2Z3 + Z3Y4 + Y4Z5 + Zly2Z3y4 + Zly2y4Z6 + y2Z3y4Z6
or, with numerator and denominator multiplied by
YlYsY5 = 1 /z1z3z6,
e3 Y2Y4
(13b)
ii Yly3y5 + Y2Y3YS + YlY2YS + yly4y5 + Y1Y3Y4 + y2y4y5 + Y2Y3y4 + yly2y4

ward paths and their cofactors. Fig. 6 gives several ad- This result can be checked by the branch-combination
ditional examples on which you may wish to practice method mentioned at the beginning of this paper.
evaluating gains by inspection. A different formulation of the problem is indicated by
the graph of Fig. 7(c), whose equations state that
ILLUSTRATIVE APPLICATIONS OF FLOW GRAPH i3-= y5e3, e2= e3+z4i3, i2=i3+y3e2, and so forth. In the
TECHNIQUES TO PRACTICAL physical problem i1 is the primary cause and e3 the
ANALYSIS PROBLEMS final effect. We may, however, choose a value of e3 and
The study of flow graphs is a fascinating topological then calculate the value of i1 required to produce that
game and therefore, from one viewpoint, worthwhile in e3. The resulting equations will, from the analysis view-
its own right. Since the associated equations of a linear point, treat e3 as a primary cause (source) and il as
924 PROCEEDINGS OF THE IRE July
the final effect (sink) produced by the chain of calcu-
lations. This does not in any way alter the physical role
of ii. The new graph (c) may appear simpler to solve
than that of (b). Since graph (c) contains no feedback
loops, the determinant and path cofactors are all equal
to unity. There are many forward paths, however, and
careful inspection is required to identify the sum of
their gains as
- = ylz2y3z4y5 + y1z2y3 + y1z2y6 + y1z4y6 (a) (b)
e3

+ yl + y3 + y5
+ y3Z4y5 (13c)
which proves to be, as it should, the reciprocal of (13b).
Incidentally, graph (c) is obtainable directly from graph
(b), as are all other possible cause-and-effect formula-
tions involving the same variables, by the process of
path inversion discussed in a previous paper.2 This
example points out the two very important facts: 1) (C)
the primary physical source does not necessarily appear Fig. 8-Voltage gain of a feedback amplifier. (a) A feedback ampli-
as a source node in the graph, and 2) of two possible fier; (b) The midband linear incremental circuit model; (c) A
flow graph formulations of a problem, the one having possible flow graph.
fewer feedback loops is not necessarily simpler to solve
by inspection, since it may also have a much more com-
plicated set of forward paths.
Fig. 8(a) offers another sample analysis problem,
determination of the voltage gain of a feedback ampli-
fier. One possible chain of cause-and-effect reasoning,
which leads from the circuit model, Fig. 8(b), to the
flow graph formulation, Fig. 8(c), is the following. First
notice that eg, is the difference of e1 and ek. Next express
ii as an effect due to causes eg1 and ek, using superposi- Rk
tion to write the gains of the two branches entering
node i1. The dependency of eg2 upon il follows directly. Fig. 9-Elimination of superfluous nodes e,j and e,2.
Now, e2 would be easy to evaluate in terms of either
eg2 or if if the other were zero, so superpose the two possible formulations. The better our perception of
effects as indicated by the two branches entering node the workings of a circuit, the fewer variables will we
e2. At this point in the formulation ek and if are as yet need to introduce at the outset and the simpler will be
not explicitly specified in terms of other variables. It is the resulting flow graph structure.
a simple matter, however, to visualize ek as the super- In discussing the feedback amplifier of Fig. 8(a) it
position of the voltages in Rk caused by i1 and if, and to is common practice to neglect the loading effect of the
identify if as the superposition of two currents in Rf feedback resistor Rf in parallel with Rk, the loading
caused by ek and e2. This completes the graph. effect of Rf in parallel with R2, and the leakage trans-
The path from ek to eg0 to i1 may be lumped in mission from ek to e2 through Rf. Such an approximation
parallel with the branch entering i1 from ek. This simpli- is equivalent to the removal of the branches from ek
fication, convenient but not necessary, yields the graph to if and if to e2 in Fig. 9. It is sometimes dangerous to
shown in Fig. 9. We could, of course, have expressed il make early approximations, however, and in this case nio
in terms of e1 and ek at the outset and arrived at Fig. 9 appreciable labor is saved, since we can write the exact
directly. All simplifications of a graph are themselves answer by inspection of Fig. 9:

lAlA2RlR2 RA 1ARkr2R2
R-
e2 (r1 + R1)(r2 + R2) L Rf I (r1 + R1)(Rf)(r2 + R2) (14)
el 1 + (/l1 + 1)Rk RA, + r2R2 +(l + 1)/2R,RkR2 (g1 + 1)Rkr2R2
r+ + 1 R+ + ++
Rf Rf1(r2 R2) Rf1(ri + RI) (r2 + R2) Rf,(r, +
± Ri) (r2 ± R2)
1956 Mason: Further Properties of Signal Flow Graphs 925

The two forward paths are elile2 and elilekife2, the first A In
having a cofactor due to loop ekif. The principal feed- )
B
back loop is ile2ifek and its gain is the fifth term of the
denominator. Physical interpretations of the various r I

paths and loops could be discussed but our main pur- ( I) (2) (3) (4)
pose, to illustrate the formulation of a graph and the (a)
evaluation of its gain by inspection, has been covered.
As a final example, consider the calculation of micro- I -r, d12 +r2 d23 1+ r3 d34
wave reflection from a triple-layered dielectric sand- AP (-r -2 P.r 3} X3 4
wich. Fig. 10(a) shows the incident wave A, the re-
flection B, and the four interfaces between adjacent I-r1 d12 I-r2 d23 1-r3 d34
regions of different material. The first and fourth inter- (b)
faces, of course, are those between air and solid. Let Fig. 10-A wave reflection problem. (a) Reflection of waves
ri be the reflection coefficient of the first interface, from a triple-layer; (b) A possible flow graph.
relating the incident and reflected components of tan-
gential electric field. It follows from the continuity of
tangential E that the interface transmission coefficient k-a k

is 1 +r1, and from symmetry that the reflection coeffi-

cient from the opposite side of the interface is the nega- (a) (b)
tive of ri. A suitable flow graph is sketched in Fig.
10(b). Node signals along the upper row are right-
going waves just to the left or right of each interface, ,* o~ ~ ~ ~ ~ ~I
those on the lower row are left-going waves, and quanti-
-

ties d are exponential phase shift factors accounting for

the delay in traversing each layer. (c) (d)

Apart from the first branch r1, the graph has the Fig. 1I-Two touching paths.
same structure as that of Fig. 6(e). Hence the reflectiv-
ity of the triple layer will be
feedback loops. This can be done by contradiction. Con-
B sider two branches which either enter the same node
--= ri + (1 + ri) (1 - rl)G (15) or leave the same node, as shown in Fig. 11(a) and (c).
A
Imagine these branches imbedded in a larger graph, the
where G is in the same form as the gain of Fig. 6(e). We remainder of which is not shown. Call the branch gains
shall not expand it in detail. The point is that the ka and kb. Now consider the equivalent replacements
answer can be written by inspection of the paths and (b) and (d). The new node may be numbered zero,
loops in the graph. whence To'=O, the other T' quantities in (16) are un-
changed, and A is therefore unaltered. If both branches
PROOF OF THE GENERAL GAIN EXPRESSION ka and kb appear in a term of the A of graph (a) then
In an earlier paper2 a quantity A was definied as the square of k must appear in a term of the A of graph
(b). This is impossible since A must be a linear function
A = (1 - Ti')(I - T2') ... (1 - Tn') (16) of branch gain k. Hence no term of A can contain the
for a graph having n nodes, where gains of two touching paths.
Now suppose that of the several nontouching paths
Tk' =loop gain of the kth node as computed with all appearing in a given term of A, some are feedback loops
higher-numbered nodes split. and some are open paths. Destruction of all other
Splitting a node divides that node into a new source and branches eliminates some terms from A but leaves the
a new sink, all branches entering that node going with given term unchanged. It follows from (16) and the
the new sink and all branches leaving that node going definitions of Tk', however, that the A for the subgraph
with the new source. The loop gain of a node was de- containing only these nontouching paths is just
fined as the gain from the new source to the new sink, A = (1 - T1)(1 - T2) . . . (1 - Tm)
when that node is split. It was also shown that A, as com- (17)
puted according to (16), is independent of the order in where Tk is the gain of the kth feedback loop in the sub-
which the nodes are numbered, and that consequently A is graph. Hence the open path gains cannot appear in the
a linear function of each branch gain in the graph. It fol- given term and it follows that each term of A is the
lows that A is equal to unity plus the algebraic sum oJ product of gains of nontouching feedback loops. More-
various branch-gain products. over, it is clear from the structure of A that a term in
We shall first show that each term of A, other than any subgraph A must also appear as a term in the A of
the unity term, is a product of the gains of nontouching the complete graph, and conversely, every term of A is a
926 PROCEEDINGS OF THE IRE _J u- y

term of some subgraph A. Hence, to identify all possible With the above notation, we have from (16) that
terms in A we must look for all possible subgraphs com-
prising sets of nontouching loops. Eq. (17) also shows (18)
that the algebraic sign of a term is plus or minus in A
accord with an even or odd number of loops in that Remembering that any A is the algebraic sum of gain
term. This verifies the form of A as given in (12c) and products of nontouching loops, we find it possible to write
(1 2d).
We shall next establish the general expression for A' = A - E TkAk. (19)
graph gain (12a). The following notation will prove con- k

venient. Consider the graph shown schematically in Eq. (19) represents the count of all possible nontouch-
Fig. 12, with node n+1 given special attention. Let ing loop sets in A'. The addition of node n +1 creates
A'= the A for the complete graph of n +1 nodes. new loops Tk but the only new loop sets of A' not al-
ready in A are the nontouching sets TkAIc. The negative
A = the value of A with node n + 1 split or removed. sign in (19) suffices to preserve the sign rule, since the
T=the loop gain of node n+1. product of Tk and a positive term of Ak will contain an
odd number of loops.
Substitution of (19) into (18) yields the general
result:
n+I FIRST n NODES
E TkAk
k
T = (20)
With node n+1 permanently split, T is just the source-
Fig. 12 A flow graph with one node placed strongly in evidence.
to-sink gain of the graph and Tk is the kth forward path.
There will in general be several different feedback loops This verifies (12a).
containing node n + 1. Let
ACKNOWLEDGMENT
Tk= gain of the kth feedback loop containing node The writer is indebted to Prof. K. Wildes and also
n+1, to J. Cruz, L. Boffi, G. Amster, and other students in
Ak = the value of A for that part of the graph not his 1955 summer-term class in subject 6.633, Electronic
touching loop Tk. Circuit Theory, M.I.T.; for helpful suggestions.

CORRECTION
James R. Wait, author of the Correspondence item calculate the radiation pattern of a dipole or a slot
"The Radiation Pattern of an Antenna Mounted on a on a conducting sphere of large radius."
Surface of Large Radius of Curvature," which appeared The last sentence of the article should read:
on page 694 of the MAy, 1956 issue of PROCEEDINGS, has "It is interesting to compare this value with the
requested that the following text, modified in editing, 6 db field strength reduction in the tangent plane
be reinstated in its original form. from a slot on a flat ground plane which is abruptly
The last sentence in the first paragraph should read: truncated."
"It is the purpose of the present note to extend Mr. Wait has also informed the editors that in
and apply the Van der Pol-Bremmer theory to (2), the second bracketed term should be zhJ(')(z).